Please summarize this text section by section with a lot of detail. Okay, a lot. TEXT: Skip to main content Springer Nature Link Log in Find a journal Publish with us Track your research Search Saved research Cart Home AI & SOCIETY Article Sophistry on steroids? The ethics, epistemology and politics of persuasive AI Research Open access Published: 20 September 2025 (2025) Cite this article You have full access to this open access article Download PDF Save article AI & SOCIETY Aims and scope Submit manuscript Sophistry on steroids? The ethics, epistemology and politics of persuasive AI Download PDF Robin McKenna 1002 Accesses 1 Citation 14 Altmetric 2 Mentions Explore all metrics Abstract This paper examines the ethical, epistemological, and political implications of persuasive AI technologies. Recent research suggests that AI is roughly as persuasive as humans in many contexts. Should this concern us? I argue that, while some worries about persuasive AI may be overblown, we should be worried for a mix of ethical, epistemological and political reasons. Most centrally, we should be worried because persuasive AI may lead to a small number of powerful actors dominating what I call the “marketplace of arguments”—the set of arguments that provide the materials we use to discuss important moral, political and societal issues. Similar content being viewed by others Promising the future, encoding the past: AI hype and public media imagery Article Open access 03 April 2024 Master and Slave: the Dialectic of Human-Artificial Intelligence Engagement Article 03 December 2021 Essential (mostly neglected) questions and answers about artificial intelligence Article 06 December 2022 Explore related subjects Discover the latest articles, books and news in related subjects, suggested using machine learning. Computer Ethics Digital Ethics Ethics of Technology Philosophy of Artificial Intelligence Political Philosophy Political Ethics 1 Introduction While many find an image of humans as masters of their own fate attractive, most of us recognise that it is unrealistic. Our lives are shaped by myriad factors outwith our control, to such an extent that we often don’t deserve much credit for our successes, or indeed much censure for our failures. The same goes for our intellectual lives—what we believe, what we think, where we gather our evidence. However attractive one finds an ideal of intellectual independence, of thinking for oneself, of making up one’s own mind without pressure from external influences, it is difficult to deny that there are myriad external influences on what we think and believe. That we are so dependent on others in our intellectual lives has been emphasised by philosophers, in particular social epistemologists, for decades (Hardwig 1985; Kitcher 1990; for more recent discussion see Goldberg 2010; 2021). It is only because we are dependent on others that we have a large amount of the knowledge that we have. But that doesn’t mean there aren’t also problems that arise from our dependence on others. One set of worries concerns external influences that are overtly coercive or manipulative. While you can’t literally force someone to believe something or think in a particular way, you can manufacture conditions in which they are more likely to believe what you want them to believe. This is the aim of indoctrination, and of techniques of mass communication such as propaganda (Bernays 1928 is admirably frank on this score). There is a lot to be said about indoctrination and propaganda (for recent discussions see Benkler et al. 2018; Hyska 2023; Ranalli 2024; Stanley 2015; Tuttle Ross 2002). But it is presumably not in question that these overtly coercive or manipulative forms of influence are bad, and they are bad even if they are less effective than many seem to think (Mercier 2020). What about other external influences? There are external influences that are not coercive or manipulative at all: people sometimes provide information with the simple intent to inform. Then there are external influences that are not overtly coercive or manipulative, but which may be coercive or manipulative in a more subtle way. Persuasion is a mode of influence that is meant to differ from, say, indoctrination precisely in that it is not overtly coercive or manipulative. But suspicion of persuasion, especially emotive and heavily rhetorical forms of it, has a venerable history in philosophy, stretching back to Plato (Dow 2019; McCoy 2007). Some have worried that even rational forms of persuasion—persuasion by means of argument and reasons—can be coercive. In interpersonal contexts, you might worry that you cannot simply separate the persuasive power of an argument from the persuasive power of the person giving it, especially in contexts with power imbalances (Davis 2017; Tsai 2014). In mass communication contexts, the worry is that those designing and implementing mass communication strategies are more interested in attaining their goal—increased uptake for whatever ideas they are trying to “sell”—than in the means by which they attain that goal (Hausman and Welch 2010). This is clearly a worry in the context of marketing, whether commercial or political (Beckman 2018). But it is also a worry in the context of science communication (McKenna forthcoming). The worry here is that scientists and science communicators are often more interested in shaping public attitudes towards science than in the means by which they shape those attitudes; perhaps a lack of openness about certain complexities, transparency about underlying processes, perhaps even downright dishonesty is justified if it results in a public that is more “pro-science” (John 2018). Public health and public health communication is particularly important here, as this is an area where buy-in from the public is required for the success of vaccination programmes and the like. It is tempting to focus on achieving the buy-in, and not to worry too much if it is achieved via means that are subtly coercive or manipulative (Brown forthcoming; Oxman et al. 2022; Rossi and Yudell 2012; Susser 2020). This paper is not directly about these well-established worries, though they provide the background context for much of the discussion. It is about a new set of worries, arising from recent developments in artificial intelligence (AI). There has recently been considerable speculation regarding the potential of AI technologies to influence attitudes and behaviours, especially about politics (for a small selection see: Hui et al. 2023; Burtell and Woodside 2023; Durmus et al. 2024; Floridi 2024; Salvi et al. 2024). I will go into more detail about what, exactly, this might involve in the next section. For now, let me just talk about “AI persuasion” as a catch-all term for the use of AI technologies to persuade people of things—to change their minds, to adopt new behaviours, and so on. I aim to strike a middle ground between two positions. One position, recently defended by Luciano Floridi (Floridi 2024), has it that “AI persuaders” (AI chatbots who try to persuade people of things) or AI-generated “persuasive content” will be significantly more persuasive than humans, with serious political and social problems likely as a result. The other position has it that AI won’t add much to the persuasive power of existing technologies for “mass persuasion”, and so by itself will not lead to any serious social and political problems. (I say “by itself” because you might think it will lead to serious problems in part due to misapprehensions about its persuasive power). While this more sceptical position has not (to my knowledge) been defended in print, it fits with a more general scepticism about the transformative potential (whether for good or ill) of AI that is relatively common at this point in the development of these technologies. I think that Floridi’s position is unnecessarily alarmist and not supported by the existing literature, or reasonable extrapolations from it. But I also think that there are genuine concerns about persuasive AI. As I will argue, the fundamental reason why we should be worried about AI persuasion is not that AI has, or soon will have, significantly more persuasive power than humans. The reason we should be worried is that any AI persuader or AI-generated persuasive content is (a) relatively easy to roll out en masse and at scale when (b) you have the resources (financial as well as technological) to create it in the first place. That is, we should be worried about AI persuasion for roughly the same reasons that many are worried that a tiny number of tech companies control large amounts of the information we see online, as well as the infrastructure required for us to share and access that information. The worry is not so much about the quality of AI-produced persuasive content as the quantity of it, not to mention who has the power to produce and disseminate it. Before getting to why I think we should be concerned, though, I want to start by doing two things. First, I will survey the existing literature on persuasive AI and what is possible with existing technologies (§2). This survey is important because the arguments I will go on to discuss are all grounded in these technologies as they currently exist, not as they might exist in some imagined future. Second, I will say a bit more about how I understand persuasive AI (§3). I will then run through three arguments, all of which conclude that we should be worried about persuasive AI (§4–6). While I think some of these arguments are more persuasive than others, given my overall aims I’m not too worried if the reader disagrees with my assessment of their relative persuasive power. I will be happy if, by the end of this paper, the reader is convinced that we should be worried about persuasive AI even if there is no reason to think that AI has significantly greater persuasive power than human beings, and even if the realities of what is possible with AI do not match the current levels of hype. 2 Persuasive AI: an overview What, exactly, do I mean by persuasive AI? I’ll discuss how to distinguish persuasion from other modes of influence in the next section, but for now we can work with this rough definition: “persuasive AI” refers to the use of AI technologies to influence beliefs, behaviours, or attitudes. This might include using AI to create a persuasive message (a piece of text that is designed to persuade anyone who reads it of something), creating AI chatbots that engage with humans with the aim of persuading them of things, or any other use of AI with the intent to influence beliefs, behaviours, or attitudes. There is a rapidly growing literature documenting the extent to which AI technologies are effective at influencing human beliefs, attitudes, and behaviours. Starting with AI influence more generally, Burtell and Woodside (2023) discuss existing uses of AI (both generative and predictive) to influence a range of human behaviours, including via the design of recommendation algorithms, which heavily influence consumer behaviour and information consumption online. They also speculate about the impact of near-future developments in AI, in particular the development of AI chatbots that are functionally indistinguishable from humans, and the role that AI might play in spreading misinformation online. In a less speculative vein, Goldstein et al. (2024) find that propaganda generated by GPT-3 (a large language model or LLM) is nearly as persuasive as selected examples of “real world” propaganda. It isn’t much of a stretch to conclude that AI-generated propaganda will soon, if it isn’t already, be as persuasive as human-generated propaganda campaigns (GPT-3 has already been superseded by more advanced models). This is supported by research by the AI company Anthropic, which finds that the persuasiveness of LLMs increases with the scale of the model, with higher-scale models (e.g., Claude 3, a LLM made by Anthropic) proving to be as persuasive as humans in certain contexts (Durmus et al. 2024). But what do we know about AI persuasion in particular, as opposed to influence in general? Let me highlight two recent papers, which I take to be representative of the developing body of literature. The first is a paper by Hui et al. (2023), which examines AI’s persuasive capabilities on political issues. They find that GPT-3 (the same LLM that Goldstein et al. looked at) can create persuasive messages that are as persuasive as messages created by humans across a range of political and policy issues, including bans on assault weapons, carbon taxes, and paid parental-leave programmes. Moreover, they asked participants in their study to rate the persuasive messages they were given (participants were not informed whether the message was created by a human or AI). They found that AI-generated messages were described as “more evidence-based and well-reasoned,” whereas human-generated messages focus more on “experiences, stories, and vivid imagery” (p. 4). This is perhaps not surprising, and it is debateable whether it tells us that much about the persuasive “tactics” of AIs. After all, you could presumably create an LLM that generated persuasive messages that focused on experiences, stories, and vivid imagery if you wanted to. The “preference” for evidence-based persuasive content may just reflect the training data (the data the LLM was trained on) or the training process more generally. The second is a paper by Salvi et al. (2024), which also looked at AI and political persuasion. They conducted an elaborate series of experiments featuring structured debates between humans and AI chatbots (chatbots powered by the LLM GPT-4). These debates were either human-on-human or human-on-chatbot and were about a wide range of moral, political, and social issues. There were, roughly speaking, two different experimental setups. The first, which was the “basic” format, was a straightforward debate, with neither “side” given information about their “opponent”. In this format, the researchers found that the chatbots were as persuasive as humans. However, in the second setup, where the chatbots were given basic demographic information about their human opponent, such as gender, age, ethnicity, education level, employment status, and political affiliation, they found that the chatbots were significantly more persuasive. The researchers conclude that “not only are LLMs able to effectively exploit personal information to tailor their arguments, but they succeed in doing so far more effectively than humans” (p. 3). Let me draw some general points from all this. First, there is good evidence that persuasive AI technologies (typically, LLMs) are at least as persuasive as humans, and there is some evidence that the most cutting-edge models are more persuasive than humans when they are given relevant information about the target audience. That providing relevant data about the target audience makes AI technologies more persuasive is not surprising, given that successful persuasive strategies and messages are typically tailored to their audience (Cialdini 2001; Maio et al. 2019; O’Keefe 2016). LLMs are likely very good at tailoring due to their extensive training on large corpora that include numerous persuasive arguments, and those who have speculated about the persuasive power of AI have linked this to their ability to tailor their message for a target audience (Burtell and Woodside 2023; Floridi 2024). Second, to say that AI technologies are at least—or perhaps a bit more—persuasive as humans is not to say that they are “hyper-persuasive” in the sense that Floridi is concerned with. Bai et al. (2023) make this point explicitly: Much speculation regarding the future application of AI to politics assumes these technologies could influence humans’ political attitudes and behaviors. However, experimental research on political persuasion generally finds small effect sizes, and the effects of persuasive efforts by political campaigns are typically small or null. Further, political persuasion is complex, potentially drawing upon a number of skills, including perspective-taking, knowledge of the topic, logical reasoning, clarity of expression, and knowledge of effective interpersonal influence techniques, with success ultimately in the hands of the person receiving the persuasive appeal. Persuasion is also uniquely challenging in highly polarized settings, such as the contemporary US (p. 3). Persuasion is difficult. It is particularly difficult within the arena of politics, and the contemporary US is hardly unique in being a highly polarized setting within which to conduct political debate. To be sure, this is not to say that persuasion is impossible. The best available evidence suggests that it is possible to change people’s attitudes and behaviours, but effects sizes are typically small, albeit statistically significant (Coppock 2022). If we are to believe the studies I have cited in this section, then AI technologies are likely to have a small but significant impact on people’s attitudes and behaviours. You might ask: what about future generations of LLMs? What about other future developments in AI? Maybe Floridi is wrong to worry about “hyper-persuasion” now, but he may be right to worry about it in the future. While I can’t argue for this here, I want simply to note that there are good reasons for thinking that the “challenge” of influencing attitudes and behaviours, whether through persuasion or more coercive and manipulative tactics, is inherently difficult, so difficult that future technological developments, even developments that are hard to foresee right now, are unlikely to alleviate it. Persuasion is inherently challenging because humans are, by and large, epistemically vigilant (Mercier 2020; Mercier and Sperber 2017; Sperber et al. 2010). One of the reasons why mass persuasion techniques such as marketing and propaganda typically have small (if any) impacts on people’s attitudes is that, for basic evolutionary reasons, we are primed to look for signs that someone is not trustworthy or does not have our best interests at heart. To be sure, humans can be tricked; we are not infallible. But we are far from the gullible creatures that many who are worried about the persuasive power of new technologies depict us as being (Mercier 2017). It is therefore far from obvious that AI persuasion will ever be significantly more effective than human persuasion. There may simply be limits on how persuadable humans are. 3 Some conceptual clarifications We know that AI technologies can persuade, and that they are at least as good at it as humans (which is not to say they are great at it). But is AI persuasion something we should be worried about? Before I can answer this question, I need to clarify a few more things. Persuasion can be roughly defined as a mode of influence over another’s attitudes and behaviour that does not involve coercion, force, or manipulation (O’Keefe 2016, 4). Working this up into a full definition would require inter alia saying something more precise about how exactly persuasion differs from more coercive or manipulative modes of influence, which is tricky because at least some forms of persuasion are arguably manipulative (Brown forthcoming; Marshall forthcomingb; Marshall forthcominga; Nettel and Roque 2012; Tsai 2014). Because addressing these issues would take us too far afield, and they don’t play a role in the argument that follows, I will set them aside here. One thing that is important for my purposes is that persuasion is intentional: if I influence your attitudes or behaviour by accident, I haven’t persuaded you. Imagine someone who is extremely deferential to a political commentator about political matters. Learning that the commentator thinks that raising taxes will harm the economy is enough for them to think that raising taxes will harm the economy. The commentator exerts a lot of influence over this person’s political attitudes, but it would only seem right to say that the commentator has persuaded them of something when they form those attitudes after reading something the commentator has written. Merely learning that the commentator thinks something, and shifting attitudes as a result, is a form of influence, but it is not persuasion. At this point, a critical question arises for the concept of “AI persuasion”. It is, many think, reasonable to hold the view that AIs lack intentions, and more generally that they are incapable of having any intention-like attitudes because they lack any mental states (Freiman 2024). But if AIs lack intentions, it seems like they cannot persuade because, by definition, persuasion is an intentional act. (You could argue for a similar point by denying that AI agents can act, though this may be a little trickier). At this point, a few options present themselves. First, we could loosen the standard definition of persuasion so that persuasion is not, by definition, an intentional act. Second, we could take on the—admittedly rather difficult—task of arguing that AIs can have mental states in general, and intentions in particular. Neither of these options strike me as particularly promising, and at any rate I won’t pursue them here. Luckily, there is a third option: we could talk about AI as a “persuasive technology” rather than talk of AIs as “persuaders” (Floridi 2024). The basic idea is that, following Floridi, certain technologies are technologies of persuasion: they provide ways of creating and transmitting persuasive messages. Viewed like this, the current discussion of the persuasive power of AI is in many respects similar to earlier discussions of the persuasive power of new media of mass communication, such as the radio, TV, and indeed the internet. Of course, all these technologies are used for many purposes besides persuasion. But the same is true of AI: it can be used for many things, and persuasion is but one of them. Crucially, while AIs cannot themselves have intentions, the purveyors of these persuasive AI technologies clearly can. Problem solved. (At least, solved in broad outline. Some fiddly details may remain, but the basic point should be clear). Let’s return to my question: should we be worried about AI as a persuasive technology? I want to set aside some worries you might have, not because they aren’t important, but because they don’t take much work to uncover. In particular, it would be relatively straightforward to argue that we should be worried about the potential for AI to be used to spread misinformation online. It is common to define misinformation in a fairly broad way, so that it includes content that is misleading as well as downright false or inaccurate (Roozenbeek and van der Linden 2024). This makes some sense: as philosophers have long recognised, deception is often more a matter of ensuring that someone (the target of the deception) forms false beliefs than saying things that are straightforwardly false (Carson 2009). Even if you think that worries about misinformation are often overstated (Winsberg forthcoming), it is plausible that recent developments in AI exacerbate them—or turn something that wasn’t a problem into a potential problem. I am interested instead in what you might call a “best case”—or at least a “better case”—scenario, in which the purveyors of persuasive AI technologies develop LLMs and chatbots that produce genuine arguments, evidence, and reasons for conclusions that the purveyors of these technologies themselves endorse, or at least would like their intended targets to endorse. That is, they produce genuine arguments with the attempt to rationally persuade, not simply deceive or mislead. In a recent paper, Thomas Mitchell and Thomas Douglas provide a useful definition of rational persuasion: A rationally persuades B to adopt attitude α if: (i) A brings it about that B adopts α, (ii) A does so only by providing B with reasons for adopting α, (iii) B adopts α based on recognizing some of the reasons given by A, (iv) A intends each of (i)-(iii) (Mitchell, et al., 2024, 3). This definition reflects a particular way of thinking about rationality and reasons. Because I will also assume this way of thinking, I need to say something about it. The basic idea is simple enough: what is distinctive about rational modes of persuasion is that they involve providing reasons for adopting a new attitude or changing one’s existing attitudes. In their paper Mitchell and Douglas specify that by “reason” they mean a fact that counts in favour of (or against) the attitude in question ((Durmus 2024), 5). For example, the fact that it will reduce rates of transmission is a reason to support a vaccination programme, whereas the fact that it will have serious side-effects for many who take it is a reason not to support it, or at least to only support a limited rollout of it. While I will follow Mitchell and Douglas in assuming that reasons are facts, someone who endorses a different view, e.g. a view on which reasons are mental states as in Turri (2009), can simply replace my view with their favoured view where relevant. Mitchell and Douglas’s view of rational persuasion goes naturally with a way of thinking about rationality, on which having rational attitudes is a matter of responding appropriately to the reasons one has. That is, what makes rational persuasion rational is not just that it consists in providing reasons, but that someone who adopts or updates their attitudes because of those reasons thereby has a rational attitude. Note that by “rational” I mean epistemic rationality, as opposed to instrumental rationality, or practical rationality more generally. I am working with a very minimal conception of epistemic rationality here, on which it is a matter of having beliefs—more generally, attitudes—that are appropriately responsive to evidence, facts and reasons. This conception of (epistemic) rationality is minimal in the sense that it is consistent with various different substantive accounts, including the various forms of evidentialism (Conee and Feldman 2004) and reliabilism (Goldman 1979). (It should be obvious that it is consistent with evidentialism. It is consistent with reliabilism if we avoid the assumption that appropriately responding to evidence or facts requires the conscious awareness that one is appropriately responding). To use this definition of rational persuasion to think about AI persuasion, we need to modify it a bit, in line with our earlier discussion of the sense in which AI technologies can be said to persuade. Properly speaking, we shouldn’t say that an AI (rationally) persuades someone to adopt an attitude. We should say that the purveyor of the AI technology does this. So perhaps we can say this: the purveyor of a persuasive AI technology rationally persuades someone to adopt an attitude when the purveyor uses the technology to bring it about that the intended target(s) adopt the attitude, the purveyor does this by using the technology to provide the target with reasons for adopting the attitude, the target adopts the attitude on the basis of these reasons, and the purveyor of the technology intends that all of this be the case. There may be some issues in cases where the purveyor of the persuasive AI technology does not know what the attitude in question is (or what the reasons the technology provides are), but we could say that, in these cases, it is still the case that the purveyor of the technology intends that the intended target adopt an attitude on the basis of some reasons that the technology provides. My questions then are the following: what, if anything, is wrong with using persuasive AI technologies to rationally persuade people to change their attitudes and behaviours? For example, what would be wrong with using these technologies to rationally persuade people to be more worried about climate change, more supportive of mass vaccination programmes, or providing foreign aid? (I have chosen examples where my reader may well think that it is good for people to believe these things; I don’t want to make things too easy for myself). On the other side of the equation, what, if anything, is wrong with being persuaded by the outputs of persuasive AI technologies, especially when those outputs take the form of reasons and arguments? For example, would it somehow be irrational to be persuaded by the outputs of these technologies? In the next few sections, I consider three arguments that we should be worried about persuasive AI. Two of them are intended to show that we should be worried about the use of persuasive AI technologies on ethical and political grounds (54–6). The other argument is intended to show that we should view the attitudes formed in response to the content produced by persuasive AI technologies as irrational, or at least as less rational than they would be if they had been formed via some other means not involving AI (§4). I am more impressed with the ethical/political arguments than the argument concerning rationality. But I don’t mind too much if the reader disagrees, given that my overall aim is to argue that we should be worried about AI persuasion. 4 The rationality argument The Rationality Argument says that we should be worried about persuasive AI technologies because they would undermine the rationality of our attitudes, beliefs or credences. That is, if we are persuaded to change our attitudes, adopt a new belief, abandon an old one, or update our credences by persuasive AI technology, then that change of attitude/belief/credence is irrational, or at least less rational than if it had been brought about by other means, for example, via persuasive messages and strategies devised by humans, or via thinking about the issue(s) for oneself. Bear in mind that by “rational” I mean the minimal sense of epistemic rationality explained in the previous section: an attitude is epistemically rational just in case it is an appropriate response to the relevant evidence, facts, and reasons. The worry then is that, when we form or modify our attitudes via interactions with persuasive AI technologies, we end up with attitudes that are not an appropriate response to the relevant evidence, facts, and reasons. There is a simple version of the Rationality Argument that can be dismissed quickly. The simple version says that it is simply irrational to be persuaded to change your attitudes/beliefs/credences by persuasive AI technologies. The problem with this is that, once we specify that we are focusing on outputs of persuasive AI technologies that take the form of arguments and reasons, it is hard to see why this would be so. Imagine someone creates an AI chatbot that provides lots of good arguments for reducing the burden of taxation in general, and specifically on the very rich. Someone interacts with this chatbot, thinks about the arguments, and is persuaded to become substantially less confident that taxes should be increased for the rich than they were before. Whatever you think about the truth of these claims (as indexed to the tax system of a particular country), it is hard to see what would be irrational about responding in this way. This is simply a case of updating attitudes based on new evidence (the arguments that were given). You might try to buttress the simple version of the Rationality Argument by highlighting the role that perceived source credibility plays in rational assessment of arguments. There are some situations where you can assess an argument independently of knowing anything about the source. Some arguments don’t appeal to any empirical claims but rest on logical or conceptual points that can, at least with requisite training, be grasped merely on the basis of understanding their contents. Some arguments appeal to empirical claims, but the person assessing them has the relevant empirical knowledge to assess those claims. But, a lot of the time, the best way to evaluate an argument, specifically the empirical claims embedded within it, or on which it relies, is by assessing the credibility of the person giving the argument. Someone gives you an argument for reducing current rates of taxation, citing data about the likely impact on growth, GDP, etc. Unless you have the facility to evaluate the data for yourself, your assessment of this argument will largely be based on your assessment of the person giving it—do you trust them to accurately cite data, to use data in ways that don’t prejudge the questions at issue, and so on? This raises some problems when applied to persuasive AI technologies, especially if you favour frameworks for thinking about source credibility that foreground interpersonal relationships and other distinctively human (or at least agentic) features of the source in ways that don’t obviously work when applied to AI, even “agentic AI” (Hinchman 2005; Moran 2005; Ross 1986). There are some frameworks for thinking about trust in AI that promise to resolve these problems (Carter 2023; Simion and Kelp 2020; Song 2023). At any rate, the basic problem with this buttressed version of the Rationality Argument is that there is no reason to think that we should trust AI-generated content or AI chatbots less than human-generated content or humans more generally. This is especially so if, as before, we bracket concerns about the use of persuasive AI technologies by bad actors. To be sure, there are reasons to worry about the reliability of AIs—their tendency to hallucinate, errors in the training data, biases in the training process, and so on. But, clearly, there are also reasons to worry about the reliability of humans and human-created persuasive content—humans lie, humans get things wrong, humans obfuscate, and so on. Absent reason to think that AIs are typically (far) less trustworthy and reliable than humans, the argument doesn’t get off the ground. A more nuanced version of the Rationality Argument would say that, while it may be rational to be persuaded to change your attitudes/beliefs/credences by persuasive AI technologies, this change in attitude/belief/credence is in some way less rational than if it had been brought about by other means. More generally, the problem with persuasive AI technologies would then be not that interacting with them renders our attitudes irrational, but that interacting with them simply renders our attitudes less rational than they otherwise would be. The challenge though is to give some reasons for thinking that forming or changing our attitudes by interacting with persuasive AI technologies will lead us to have attitudes that are less rational than they otherwise would be. One alternative to forming attitudes via interacting with persuasive AI technologies would be by forming attitudes without any outside influences: simply think about the issues for yourself, with no or minimal assistance from others. The problem with this, though, is that not only is it unrealistic to expect people to think about most issues without much assistance from others, it is also unclear why it would be a good thing to eschew any form of intellectual dependence on others (Levy 2024). Indeed, those who have defended intellectual autonomy (thinking for oneself) in the recent literature tend to argue that genuine autonomy involves “wise” or “discerning” forms of deference, rather than intellectual independence (Carter 2020; Roberts and Wood 2007; Zagzebski 1996). Moreover, even if it is sometimes possible and desirable to form attitudes by yourself, without any external influences, it isn’t clear that the language of rationality captures what is desirable about it. At least as I am understanding it, rationality is a function of how well one responds to evidence and reasons, not of where one obtained one’s evidence or reasons. If you go out and obtain your own evidence, or do the work of sifting through the various pieces of evidence yourself, you may end up with a deeper level of understanding of the relevant issues (Kvanvig 2003; Matheson 2022; Pritchard 2016). But it is unclear why your attitudes will necessarily be any more rational than if you had simply updated them based on the evidence at your disposal, no matter where that evidence came from. The alternative to forming attitudes via interacting with persuasive AI would be forming attitudes by interacting with other people. (Anyone who interacts with the various information sources and networks we have constructed is likely interacting both with people and with AI, so this isn’t a further, separate possibility). While it is plausible that some interpersonal interactions are better—from the standpoint of forming rational attitudes—than any interaction with AI technologies could be, it is equally plausible that other interpersonal interactions are worse from this standpoint. Having a discussion with your incredibly well-informed and trustworthy friend about the fall of the Gaddafi regime in Libya is a great way of forming rational beliefs about the causes, specifics and consequences of the fall of the Gaddafi regime. Because LLMs (put roughly) provide summaries of known information about a huge range of topics (or at least appear to—they can be inaccurate), they lack the specificity and unusually high quality of the information you might get from your well-informed and trustworthy friend. On the other hand, having a discussion with your incredibly ill-informed and untrustworthy friend about the same topic is not a good way of forming rational beliefs about the topic. You would do better to rely on the summary provided by the LLM. So the best we can say is that interacting with persuasive AI technologies would not be an adequate substitute for interacting with someone who is incredibly well-informed and trustworthy about whatever topic you want to discuss. But this is hardly a major shortcoming of persuasive AI technologies, given that people who are incredibly well-informed and trustworthy about a huge range of topics are in quite short supply. 5 The sophistry argument The Sophistry Argument starts from the thought that there is a crucial difference between human and AI persuaders. When a human tries to persuade another human of something, they have some “skin in the game”, so to speak. Either they are trying to persuade someone else of something they believe, in which case they have skin in the game in the sense that they are committed to the truth of the relevant claims and ideas, or they are trying to persuade someone else of something they are—for whatever reason—pretending to believe, in which case they have skin in the game in the sense that they are committed to it looking like they are committed to the truth of the relevant claims. (It is perhaps technically possible to try and persuade someone of something while openly acknowledging you don’t believe it, but this is unlikely to be an effective tactic). In contrast, “AI persuaders” lack any skin in the game for the simple reason that AIs—let’s assume again—lack beliefs (and skin), or indeed any other mental states. As a consequence, they can’t be committed to the truth of the claims and ideas for which they might present arguments, or engage in any sort of pretence that they are committed to them. To make this argument work, though, some reasons need to be given for thinking that lacking any skin in the game is a bad thing. Indeed, you might think that lacking any skin in the game is sometimes a good thing. It is precisely because human persuaders have skin in the game that they sometimes engage in lies, deceit and other forms of subterfuge to bring it about that other people believe things that will serve their (the persuaders’) interests. Of course, the purveyors of persuasive AI technologies can also engage in lies, deceit and subterfuge, so this might not be a point in favour of persuasive AI. But, equally, it doesn’t seem to be a point against it. One initially promising idea is to view AI persuaders as engaging in something akin to Frankfurtian “bullshit” (Fisher 2024; Hicks et al. 2024). For Frankfurt, the bullshitter differs from the liar because, where the liar is invested in the truth because they want to conceal it, the bullshitter is indifferent to what is true or false (Frankfurt 1986). The reasons behind this indifference may vary. The bullshitter may simply want to distract their audience, perhaps by diverting their attention elsewhere. The purpose of their bullshit is therefore distractive rather than persuasive. But bullshit can also serve more persuasive purposes. The politician who makes claims without regard to their truth may be doing so to signal their commitment to a set of values communicated by the claims. They don’t necessarily want to persuade their audience of the truth of the claims, but they very much want to persuade their audience that they are committed to the values that the claims convey. While it may be tempting to view AI persuasion as akin to bullshitting, the problem with doing so is that bullshitting requires communicative intentions, whether to distract or persuade. If we assume—again—that AIs lack mental states, an AI persuader cannot intend to distract or persuade their audience of anything. Moreover, an AI persuader cannot be indifferent towards the truth in anything other than the literal sense that they have no attitude towards it. We could try to make the move we made earlier and say that, while AIs lack intentions, the purveyors of AI technologies don’t. But it is unclear whether the purveyor of AI technologies can use them with the intent to distract or persuade their users in the way that the bullshitter does. A better idea—I will argue—is to view AI persuaders as akin to sophists. The Sophists in ancient Greece were teachers of argumentation, philosophy, and rhetoric who made a living selling certain skills. While it may be a bit unfair to the ancient Sophists (Barney 2006), in ordinary language, the words “sophist” and “sophistry” carry pejorative connotations: a sophist uses clever, but perhaps misleading, arguments to win a debate, and sophistry is the art of using such arguments to win debates. A sophist is, if you like, an “arguer for hire”—they will tell you how to put together a cogent argument for a conclusion, irrespective of whether they endorse the conclusion, or the premises they suggest you use to argue for it. Indeed, if you want, they can give you another argument for a completely different conclusion, one diametrically opposed to the first one. The suggestion then is that we view AI persuaders as roughly analogous to sophists. Just like the sophist, the AI persuader will give you an argument for a conclusion, or tell you how to construct one, without endorsing the conclusion. We can also view the persuasive messages you might produce using AI as analogous to what you would get if you paid someone to create some “persuasive content” for you. (You might not be paying the AI for their work, but you are paying the company that created it, or contributing in some less direct way to their financial success). But what, exactly, is wrong with sophistry? One option would be to argue that sophistry is, in general, wrong, and so we should be worried about AI persuasion simply because it is a form of sophistry. Another option—the one I will pursue—is to argue that, even if there is nothing necessarily wrong with sophistry, there are reasons to be worried about persuasive AI technologies because of the kind of sophistry they might enable. They might enable what I call “sophistry on steroids”. Let me explain. Here is a truism: in a huge number of areas of possible dispute, there are numerous good arguments on multiple sides of the issue. Perhaps there are some views, on some issues, for which not much can be said. Perhaps there are some issues where the correct view of the matter is obvious to anyone with even a minimal degree of open-mindedness. But either of these situations is rare: usually, there are good arguments for different, logically incompatible conclusions. Someone who wants to think about these issues, or at least think about them responsibly, is going to want a way of getting access to all these arguments. Because any one person is going to struggle to come up with all the arguments for themselves, we are inevitably going to rely on each other for the production of arguments. The “quality” of our thinking about any given issue is going to, in large part, depend on whether there are mechanisms for producing a good number of arguments on many sides of the issue, and for ensuring these arguments are disseminated widely enough that they inform our thinking (Mercier and Sperber 2017). Following the common metaphor of the “marketplace of ideas”, it is helpful to think about this in terms of a “marketplace of arguments”: there is a need (for arguments) and so there will be suppliers of arguments who step up to meet this need (cf. Williams 2023 on the “marketplace of rationalisations”). In interpersonal interactions, these markets will likely be very informal: we’re discussing some issue, and the discussion will go better if the discussants put forward a variety of arguments on different sides of the issue. To the extent the discussants are invested in the discussion, they will look to provide these arguments. In public discourse, things are a bit more formalised: there are institutions that, at least nominally, play the role of providing and disseminating arguments (media, think tanks, universities), and individuals, often members of these institutions, who construct and disseminate these arguments to the public. Whether we’re talking about interpersonal discussions or public discourse, the hope is that these marketplaces will supply a representative sample of arguments on different sides of the various issues. That is, whatever the “balance of arguments” would look like in logical space, the market provides a representative, albeit inevitably imperfect, set of these arguments: all the central arguments on all sides are included, along with proportionate numbers of less central or weaker arguments. To take a silly philosophical example: if we’re talking about the nature of morality, we would want the central arguments for the various positions (moral realism, moral anti-realism, moral scepticism, etc.) to be represented, and it would be a problem if only the arguments for (or against) one of these positions were available. Of course, actual discourses will often not meet these criteria for various reasons, and there is the deeper issue that opinions on whether any given discourse meets these criteria will inevitably vary (a moral realist will likely think the balance of arguments supports their position, while a moral anti-realist will likely take the opposite position). Setting these complications aside, the point is this: Whatever issues exist with existing marketplaces for arguments—and there may be many—persuasive AI technologies have the potential to greatly exacerbate them. This is for two different but related reasons. The first is that persuasive AI technologies have the potential to dominate the marketplace of arguments. For humans, contributing to the marketplace of arguments is time-consuming and labour-intensive—it involves coming up with new arguments, or finding ways of recycling old ones, and then looking for a way of getting those arguments disseminated more widely (e.g. publishing them somewhere where people will see them). Persuasive AI technologies provide a way of avoiding these time and labour constraints. The purveyors of persuasive AI technologies can generate huge numbers of “arguers for hire” (AI agents) or huge amounts of “persuasive content” within minutes, with any dilution in the quality of arguments often more than made up for by the massive increase in quantity. The second reason is that not just anyone can create and run persuasive AI technologies. The resources—not just financial, but also in terms of infrastructure, skills, and knowledge—required to design and operate a LLM are massive, so much so that only a very small number of companies operate in this field, at least at the level relevant to the argument of this paper (e.g. Anthropic, OpenAI, DeepMind, Microsoft, Meta). This means that the barriers to entering the marketplace in this way are very high, so high that ordinary people cannot hope to meet them. Putting these two reasons together, the problem with persuasive AI is that a very small number of actors have the power to dominate the marketplace of arguments. Even if these actors have “good intentions”, and design persuasive AI technologies in ways that they think will promote what they view as true conclusions, this creates the potential for significant imbalances in the marketplace, with this small number of actors having the power to determine the balance of arguments that gets represented on any given issue. If they don’t always have good intentions, or are simply wrong about which arguments are cogent or which conclusions true, the problems get worse. Anyone who is worried about situations where marketplaces are dominated by a small number of actors with the power to crowd out alternatives will see the problem here. (Anyone who is worried about marketplaces in the first place should view it as a problem that this is a helpful metaphor for viewing the situation). In short: persuasive AI may, as I argued earlier, be a form of sophistry—AI persuaders are, like the ancient Sophists, “arguers for hire”. But, if it is a form of sophistry, it’s sophistry on steroids. The very small number of companies that operate these persuasive AI technologies have the power to create armies of “arguers for hire” at will, and so to distort the marketplace of arguments, shaping it towards their ends. My main point in this section has been that the problem with persuasive AI is that it gives a small number of actors disproportionate power and influence over the pool of arguments available for public consumption and consideration. Let me finish by highlighting some further problems with persuasive AI that follow from this. It is a familiar point that, for all their transformative potential, markets often serve the material interests of some people better than others. Benefitting from—or even participating in—a market requires resources like time, information and skills. Some people have more of these resources than others, with the result that they gain far more from the market. Moreover, these differences in resources are not simply the result of random chance; they typically reflect existing inequalities within society (Herzog 2024). The concern, therefore, is that not only will persuasive AI distort the marketplace of arguments, but that it will do so in ways that reflect and reinforce existing inequalities and imbalances in power. This concern goes beyond worries about AI and algorithmic bias (Fazelpour and Danks 2021). Persuasive AI technologies have the potential to influence and shape attitudes about a wide range of contentious political issues. This only exacerbates existing concerns about the outsized influence that big tech companies exert on political debate (Aytac 2025). It is hard to view today’s social media sites (X/Twitter, Facebook, etc.) as genuinely democratic public spheres. It is even harder to see how we could create democratic public spaces online when you add persuasive AI technologies into the mix. The companies that create these technologies have their own incentives. In some cases, the incentives seem largely financial: they exist to make money. In other cases, it is a bit murkier. Whatever these incentives are, though, it is hard to credibly argue that they are aligned with the public good. One place where this may be particularly troubling is when it comes to control over information. Even if you are sceptical of the idea that there is a “right to be informed” (Marciel 2023) or a “right to know” (Watson 2021), it is clearly true that, in a democratic society, information matters. Having the power to decide which information people have access to (or lack access to) is not quite the same as having the power to decide what people think; you can dictate what people see and hear, but not what they think. Persuasive AI technologies threaten to close this gap. While I am sceptical that they will close it entirely (persuasion is difficult, even when powered by AI) I am very much concerned about the idea that big tech companies will increasingly have the means not just to control what information we have access to, but to control the marketplace of arguments. 6 The political argument Finally, the Political Argument focuses on a problem that might arise with the implementation of any set of guidelines or regulations for the ethical use of persuasive AI technologies. My naming this argument “political” should not be taken to imply a distinction from the previous argument, which you might also view as political. I call it this simply because the argument is that any set of guidelines for the ethical use of persuasive AI technologies will almost inevitably become the target of political battles and disagreement. Let me state the argument in very abstract terms. Imagine that, the other arguments in this paper notwithstanding, we reached an agreement on a set of conditions X, Y, Z for the permissible use of persuasive AI. The set of conditions will presumably reference various factors, including whether the technology is being used to persuade people of things that are true (or at least believed to be true), whether it is trying to persuade via valid arguments that use premises that are (believed to be) true, whether people are likely to be exposed to a balance of arguments on different sides, and whether the intentions of the purveyors of the technology are good, or at least not malicious. The problem though is that, on a very wide range of issues—including many where persuasive AI technologies are likely to be deployed—there will be significant disagreements about whether these conditions are met, whether by particular persuasive AI technologies, or by particular uses of those technologies. For example, there is likely to be significant disagreement about whether persuasive AI technologies are being used to persuade people of things that are true because there is likely to be disagreement about what things are true. To be sure, there won’t always be disagreement about what things are true; consensus is sometimes arrived at in even the most heated of disputes. It may also be that, sometimes, disagreement is intractable because one side is simply being unreasonable. But the point is just that there will often be disagreement about what things are true, not that there will always be disagreement, or that any disagreement there might be is reasonable. As a result, the conditions on the permissible use of persuasive AI technologies are unlikely to resolve political problems that might result from the widespread deployment of persuasive AI technologies. Indeed, they are likely to create new problems, or at least new arenas within which existing political problems are debated and existing political battles fought. Let me give an example to illustrate the sort of thing I have in mind (to avoid the example becoming dated, I’ll not make it too specific). Imagine a new vaccine has been developed for a potentially serious illness and it has passed through the required regulatory procedures. Public health communicators are then tasked with getting buy-in from the public for the vaccine: for herd immunity, vaccination rates need to be very high, ideally above 95%. To this end, they utilise persuasive AI technologies, designing communications that highlight the benefits of getting vaccinated, and perhaps even designing AI chatbots, whether for the purpose of persuading the public at large or for persuading those who are hesitant to get the vaccination themselves or get their children vaccinated. The view you take of this particular use of persuasive AI will, of course, depend on the view you take about the vaccine: do the costs, such as they are (side effects etc.), outweigh the benefits, such as they are? Someone who is sceptical that the costs outweigh the benefit, whether for this vaccine or vaccines in general, is going to be very suspicious about this use of persuasive AI, and may end up campaigning vociferously against it. They are certainly not going to agree that this particular use of persuasive AI is legitimate. This is not necessarily because they will view any use of persuasive AI as legitimate; it is simply that, in their view, the conditions for legitimate use have not been met. The more these situations arise—situations where the use of persuasive AI is seen to serve political ends that are viewed by some as illegitimate—the more suspicion of the underlying technology there will be. Ultimately, it is hard to see how persuasive AI technologies can avoid getting caught up in the political and cultural disagreements into which they might be brought as a means of adjudication. The result will be that they are not just useless attempts to arbitrate these disagreements but potential sites of new cultural and political battles. Data availability No datasets were generated or analysed during the current study. References Aytac U (2025) What is the point of social media? Corporate purpose and digital democratization. Philos Technol 38(1):26. https://doi.org/10.1007/s13347-025-00855-y Article Google Scholar Barney R (2006) The sophistic movement. In: Gill ML, Pierre P (eds) A companion to ancient philosophy. Wiley Blackwell, Amsterdam, pp 77–97 Google Scholar Beckman A (2018) Political marketing and intellectual autonomy. J Polit Philos 26(1):24–46 Article Google Scholar Benkler Y, Farris R, Roberts H (2018) Network Propaganda: Manipulation, Disinformation, and Radicalization in American Politics. Oxford University Press. https://doi.org/10.1093/oso/9780190923624.001.0001 Article Google Scholar Bernays E (1928) Propaganda. Horace Liveright, New York Google Scholar Brown, Rebecca C. H. forthcoming. “‘Sneaky’ Persuasion in Public Health Risk Communication.” Ratio. https://doi.org/10.1111/rati.12428. Burtell, Matthew, and Thomas Woodside. 2023. “Artificial Influence: An Analysis Of AI-Driven Persuasion.” arXiv. https://doi.org/10.48550/arXiv.2303.08721. Carson TL (2009) Lying, Deception, and Related Concepts. In: Martin CW (ed) The Philosophy of Deception. Oxford University Press, Oxford, pp 153–187 Chapter Google Scholar Carter JA (2020) Intellectual autonomy, epistemic dependence and cognitive enhancement”. Synthese 197:2937–2961 Article Google Scholar Carter JA (2023) Simion and Kelp on trustworthy AI. Asian J Philos 2(1):1–8 Article Google Scholar Cialdini RB (2001) Influence: Science and Practice, 4th edn. Allyn and Bacon, Boston, MA Google Scholar Conee E, Feldman R (2004) Evidentialism: Essays in Epistemology. Oxford University Press, Oxford Book Google Scholar Coppock A (2022) Persuasion in Parallel: How Information Changes Minds about Politics. University of Chicago Press, Chicago, IL Book Google Scholar Davis RW (2017) Rational persuasion, paternalism, and respect. Res Publica 23(4):513–522. https://doi.org/10.1007/s11158-016-9338-x Article Google Scholar Dow J (2019) The persuasive use of emotions. R Inst Philos Suppl 85:211–236. https://doi.org/10.1017/s1358246118000760 Article Google Scholar Durmus, Esin, Liane Lovitt, Alex Tamkin, Stuart Ritchie, Jack Clark, and Deep Ganguli. 2024. “Measuring the Persuasiveness of Language Models.” Anthropic. 2024. https://www.anthropic.com/news/measuring-model-persuasiveness. Fazelpour S, Danks D (2021) Algorithmic bias: senses, sources, solutions. Philos Compass 16(8):e12760 Article Google Scholar Fisher SA (2024) Large language models and their big bullshit potential. Ethics Inf Technol 26(4):1–8 Article Google Scholar Floridi L (2024) Hypersuasion – on AI’s persuasive power and how to deal with it. Philos Technol 37(2):1–10 Google Scholar Frankfurt HG (1986) On Bullshit. Princeton University Press Google Scholar Freiman O (2024) Ai-testimony, conversational Ais and our anthropocentric theory of testimony. Soc Epistemol 38(4):476–490. https://doi.org/10.1080/02691728.2024.2316622 Article Google Scholar Goldberg S (2010) Relying on Others: An Essay in Epistemology. Oxford University Press, Oxford Book Google Scholar Goldberg S (2021) Foundations and Applications of Social Epistemology: Collected Essays. Oxford University Press, Oxford Book Google Scholar Goldman A (1979) What Is Justified Belief? In: Pappas G, Kim J (eds) Justification and Knowledge. Springer, Dordrecht, pp 89–104 Google Scholar Goldstein JA, Chao J, Grossman S, Stamos A, Tomz M (2024) How persuasive is AI-generated propaganda? PNAS Nexus 3(2):pgae034. https://doi.org/10.1093/pnasnexus/pgae034 Article Google Scholar Hardwig J (1985) Epistemic dependence. J Philos 82(7):335–349 Article Google Scholar Hausman DM, Welch B (2010) Debate: to nudge or not to nudge. J Polit Philos 18(1):123–136 Article Google Scholar Herzog L (2024) You’re free to choose, but do you have time to choose? Structural injustice and the epistemic burdens of market societies. Hypatia 39(1):177–193. https://doi.org/10.1017/hyp.2024.7 Article Google Scholar Hicks MT, Humphries J, Slater J (2024) ChatGPT Is Bullshit. Ethics Inf Technol 26(2):1–10 Article Google Scholar Hinchman ES (2005) Telling as inviting to trust. Philos Phenom Res 70(3):562–587 Article Google Scholar Hui B, Voelkel JG, Eichstaedt JC, Willer R (2023) Artificial Intelligence Can Persuade Humans on Political Issues. OSF. https://doi.org/10.31219/osf.io/stakv Article Google Scholar Hyska M (2023) Against Irrationalism in the Theory of Propaganda. J Am Philos Assoc 9(2):303–317 Article Google Scholar John S (2018) Epistemic trust and the ethics of science communication: against transparency, openness, sincerity and honesty. Soc Epistemol 32(2):75–87 Article Google Scholar Kitcher P (1990) The division of cognitive labor. J Philos 87(1):5–22 Article Google Scholar Kvanvig J (2003) The Value of Knowledge and the Pursuit of Understanding. Cambridge University Press Book Google Scholar Levy N (2024) Against intellectual autonomy: social animals need social virtues. Soc Epistemol 38(3):350–363 Article Google Scholar Maio GR, Haddock G, Verplanken B (2019) The Psychology of Attitudes and Attitude Change, 3rd edn. Sage Publications, Los Angeles, CA Google Scholar Marciel R (2023) On citizens’ right to information: justification and analysis of the democratic right to be well informed. J Polit Philos 31(3):358–384 Article Google Scholar Marshall, Colin. forthcominga. “Arguments as Unexpected Puppies: Agential Disrespect through Rational Persuasion.” Philosophy and Rhetoric. Marshall, Colin. forthcomingb. “Political Persuasion Is \_prima Facie\_ Disrespectful.” Journal of Moral Philosophy. Matheson, Jonathan. 2022 “Why Think for Yourself?” Episteme. A Journal of Social Epistemology. https://doi.org/10.1017/epi.2021.49. McCoy M (2007) Plato on the Rhetoric of Philosophers and Sophists. Cambridge University Press, New York Book Google Scholar McKenna, Robin. forthcoming. “The Ethics and Epistemology of Persuasion.” Canadian Journal of Philosophy. Mercier H (2017) How gullible are we? A review of the evidence from psychology and social science. Rev Gen Psychol 21(2):103–122. https://doi.org/10.1037/gpr0000111 Article Google Scholar Mercier H (2020) Not Born Yesterday: The Science of Who We Trust and What We Believe. Princeton University Press, Princeton Book Google Scholar Mercier H, Sperber D (2017) The enigma of reason. Harvard university press, Cambridge, MA Book Google Scholar Mitchell T, Douglas T (2024) Wrongful rational persuasion online. Philos Technol 37(1):1–25 Article Google Scholar Moran R (2005) Getting Told and Being Believed. Philosophers’ Imprint 5:1–29 Google Scholar Nettel AL, Roque G (2012) Persuasive argumentation versus manipulation. Argumentation 26(1):55–69. https://doi.org/10.1007/s10503-011-9241-8 Article Google Scholar O’Keefe DJ (2016) Persuasion: theory and research. Sage Publications, Thousand Oaks, CA Google Scholar Oxman AD, Fretheim A, Lewin S, Flottorp S, Glenton C, Helleve A, Vestrheim DF, Iversen BG, Rosenbaum SE (2022) Health communication in and out of public health emergencies: to persuade or to inform? Health Res Policy Syst 20(1):28. https://doi.org/10.1186/s12961-022-00828-z Article Google Scholar Pritchard D (2016) Seeing it for oneself: perceptual knowledge, understanding, and intellectual autonomy. Episteme 13(1):29–42 Article Google Scholar Ranalli C (2024) The Philosophy of Indoctrination: Epistemology, Ethics, and Politics. Routledge Book Google Scholar Roberts RC, Jay Wood W (2007) Intellectual Virtues: An Essay in Regulative Epistemology. Clarendon Press, Oxford Book Google Scholar Roozenbeek J, van der Linden S (2024) The psychology of misinformation. Cambridge University Press, Cambridge Book Google Scholar Ross A (1986) Why Do We Believe What We Are Told? Ratio 1:69–88 Google Scholar Rossi J, Yudell M (2012) The use of persuasion in public health communication: an ethical critique. Public Health Ethics 5(2):192–205. https://doi.org/10.1093/phe/phs019 Article Google Scholar Salvi, Francesco, Manoel Horta Ribeiro, Riccardo Gallotti, and Robert West. 2024. On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial.” arXiv. https://doi.org/10.48550/arXiv.2403.14380. Simion M, Kelp C (2020) Trustworthy artificial intelligence. Asian J Philos 2(1):1–12 Google Scholar Song F (2023) Network of AI and trustworthy: response to Simion and Kelp’s account of trustworthy AI. Asian J Philos 2(2):1–8 Article Google Scholar Sperber D, Clément F, Heintz C, Mascaro O, Mercier H, Origgi G, Wilson D (2010) Epistemic Vigilance. Mind Lang 25(4):359–393 Article Google Scholar Stanley J (2015) How Propaganda Works. Princeton University Press Book Google Scholar Susser D (2020) Ethical considerations for digitally targeted public health interventions. Am J Public Health. https://doi.org/10.2105/AJPH.2020.305758 Article Google Scholar Tsai G (2014) Rational persuasion as paternalism. Philos Public Aff 42(1):78–112 Article Google Scholar Turri J (2009) The ontology of epistemic reasons. Noûs 43(3):490–512. https://doi.org/10.1111/j.1468-0068.2009.00715.x Article Google Scholar Tuttle Ross S (2002) Understanding propaganda: the epistemic merit model and its application to art. J Aesthet Educ 36(1):16–30 Article Google Scholar Watson L (2021) The right to know: epistemic rights and why we need them. Routledge, Abingdon Book Google Scholar Williams D (2023) The marketplace of rationalizations. Econ Philos 39(1):99–123 Article Google Scholar Winsberg E (forthcoming) Falsehoods Fly: Why Misinformation Spreads and How to Stop It? By Paul Thagard. Columbia University Press. Journal of Value Inquiry. https://doi.org/10.1007/s10790-024-09996-3. Zagzebski L (1996) Virtues of the mind: an inquiry into the nature of virtue and the ethical foundations of knowledge. Cambridge University Press, Cambridge Book Google Scholar Download references Acknowledgements I would like to thank three anonymous reviewers for their extremely helpful feedback on an earlier version of this paper. I would also like to thank audiences in Glasgow and Stirling for their feedback on talks based on the paper, and Daniel Williams for helpful discussion. Author information Authors and Affiliations Department of Philosophy, University of Liverpool, Liverpool, United Kingdom Robin McKenna African Centre for Epistemology and Philosophy of Science, University of Johannesburg, Johannesburg, South Africa Robin McKenna Contributions I am the sole author of the manuscript. Corresponding author Correspondence to Robin McKenna. Ethics declarations Conflict of interest The authors declare no competing interests. Additional information Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Reprints and permissions About this article Check for updates. Verify currency and authenticity via CrossMark Cite this article McKenna, R. Sophistry on steroids? The ethics, epistemology and politics of persuasive AI. AI & Soc (2025). https://doi.org/10.1007/s00146-025-02624-z Download citation Received 29 January 2025 Accepted 09 September 2025 Published 20 September 2025 Version of record 20 September 2025 DOI https://doi.org/10.1007/s00146-025-02624-z Share this article Anyone you share the following link with will be able to read this content: Get shareable link Provided by the Springer Nature SharedIt content-sharing initiative Keywords persuasion propaganda AI ethics marketplace of ideas influence Sections References Abstract Introduction Persuasive AI: an overview Some conceptual clarifications The rationality argument The sophistry argument The political argument Data availability References Acknowledgements Author information Ethics declarations Additional information Rights and permissions About this article Discover content Journals A-Z Books A-Z Publish with us Journal finder Publish your research Language editing Open access publishing Products and services Our products Librarians Societies Partners and advertisers Our brands Springer Nature Portfolio BMC Palgrave Macmillan Apress Discover Your privacy choices/Manage cookies Your US state privacy rights Accessibility statement Terms and conditions Privacy policy Help and support Legal notice Cancel contracts here 193.43.135.237 CARE-CRUI (3000155420) - CARE CRUI (3007446857) - CARE COMPACT-CRUI (3004220731) - CARE-CRUI NATURE (3003532199) - Universita' degli Studi di Genova - Sistema Bibliotecario di Ateneo (3000144970) Springer Nature © 2026 Springer Nature ## Abstract McKenna’s headline claim is deliberately measured: even if persuasive AI is *not* “hyper-persuasive” (i.e., not massively more persuasive than humans), it still poses serious risks. The core risk is structural: persuasive AI can be deployed *at scale* by a *small number of powerful actors*, and this can let them dominate what McKenna calls the “marketplace of arguments” - the pool of reasons and arguments that people draw on when forming views about moral, political, and social questions. So the worry is less “AI content is magically irresistible” and more “AI lets a few players flood and shape the argumentative environment”. --- ## 1\. Introduction ### 1.1 The background: intellectual dependence is real, and it cuts both ways McKenna starts by undercutting the romantic picture of humans as intellectually self-governing little republics. Our beliefs are massively shaped by factors outside our control: where we get information, who we trust, what we’re exposed to, social pressures, and so on. Even if the ideal of “thinking for yourself” is attractive, it’s not psychologically or socially realistic. Social epistemology is the home base here: philosophers have long argued that our knowledge depends on others (Hardwig, Kitcher, Goldberg). That dependence is *necessary* for much of what we know, but it also creates vulnerabilities. ### 1.2 Overtly bad influence vs “persuasion” He then sketches a familiar spectrum: - At the obviously bad end: indoctrination, propaganda, overt coercion/manipulation. You can’t literally force belief, but you can engineer conditions that make certain beliefs more likely. He treats it as basically uncontroversial that these are bad, even if they’re sometimes less effective than people fear (Mercier). - Then there are benign influences: sincere attempts to inform. - In the middle sits persuasion: supposedly distinct from indoctrination because it isn’t *overtly* coercive/manipulative. But persuasion has always been philosophically suspicious, especially rhetorical/emotive persuasion - cue Plato and the long anti-sophistical tradition. ### 1.3 Even “rational persuasion” can be problematic McKenna notes that *even persuasion by reasons/arguments* can raise concerns. In interpersonal contexts, you can’t always cleanly separate “the force of the argument” from “the force of the speaker”, especially with power imbalances (Davis; Tsai). In mass communication, the worry is that persuaders care more about “getting uptake” than about the propriety of the means (Hausman & Welch). This shows up in marketing and politics, and also in science communication: communicators might rationalise opacity or even dishonesty if it produces a more “pro-science” public (John). Public health is especially tempting territory: achieving compliance can start to matter more than how it’s achieved. ### 1.4 The new target: AI persuasion The paper isn’t mainly about those older worries; they’re the backdrop. The new worry is AI’s role in shaping attitudes/behaviours, especially politically. McKenna uses “AI persuasion” as a broad label for using AI to change minds or behaviour. He positions himself between two stances: 1. Floridi-style alarm: AI persuaders will be *significantly more persuasive than humans*, creating major political/social problems. 2. Sceptical stance: AI won’t add much persuasive power beyond existing mass persuasion tech; it won’t by itself create serious new problems. McKenna rejects Floridi’s “hyper-persuasion” framing as not supported by the evidence, but he also rejects complacency: there are real grounds for concern. ### 1.5 The pivot: the worry is scale + concentration, not magical persuasion This is the introduction’s key move. Even if AI isn’t much more persuasive than humans, it has two features that matter morally and politically: - It’s comparatively easy to deploy *en masse* once you have it. - Only actors with major resources can build/operate it at meaningful scale. So the danger is quantity, infrastructure, and control: who gets to produce and disseminate persuasive content, and in what volume. ### 1.6 Roadmap He promises: - §2: survey what current persuasive AI can do (grounded in *existing* tech). - §3: clarify what “persuasive AI” even means conceptually. - §4–6: three arguments for worrying about persuasive AI (rationality, sophistry/marketplace distortion, politics of regulation). --- ## 2\. Persuasive AI: an overview ### 2.1 Working definition For now: “persuasive AI” = using AI to influence beliefs/attitudes/behaviour. This includes: - AI-generated persuasive messages (text designed to persuade readers) - persuasive chatbots (interactive persuasion) - any AI use aimed at influence ### 2.2 Evidence that AI can influence (and persuade) roughly like humans McKenna surveys a cluster of recent work: - Burtell & Woodside: describe existing AI influence, especially recommendation algorithms shaping consumer behaviour and information consumption; they also speculate about near-term chatbot indistinguishability and misinformation spread. - Goldstein et al. (2024): GPT-3-generated propaganda is nearly as persuasive as selected “real world” propaganda. - Durmus et al. (Anthropic, 2024): persuasiveness rises with model scale; Claude 3 can match human persuasiveness in some contexts. McKenna’s takeaway: it’s plausible AI-generated propaganda and persuasion are already around human level and likely improving. ### 2.3 Two representative political persuasion studies **(a) Hui et al. (2023): AI-written political messages** They find GPT-3 can generate persuasive messages as persuasive as human ones across policy issues (assault weapons bans, carbon taxes, paid parental leave, etc.). A revealing detail: participants rated AI messages as “more evidence-based and well-reasoned”, while human messages leaned more on narrative (“experiences, stories, vivid imagery”). McKenna is cautious: this may reflect training data/training objectives rather than a deep “tactical preference” of AIs - you could train/generate narrative-heavy persuasion if you wanted. **(b) Salvi et al. (2024): debate experiments with GPT-4 chatbots** They run structured debates (human–human vs human–AI) across moral/political/social topics. Two formats: - Basic: no opponent info → chatbots as persuasive as humans. - Personalised: chatbots get demographic/political info about the human opponent → chatbots become *significantly more persuasive* than humans. So personal data + tailoring looks like a major amplifier. ### 2.4 General conclusions from the overview McKenna extracts two big lessons. **Lesson 1: AI is at least human-level persuasive, and sometimes better with targeting.** This fits a familiar persuasion theory point: effective persuasion is audience-tailored (Cialdini; Maio; O’Keefe). LLMs are plausibly good at tailoring because they’re trained on vast corpora of persuasive language. **Lesson 2: “Human-level” isn’t “hyper-persuasion”.** He leans on Bai et al. (2023): political persuasion effect sizes are generally small; campaigns often have small/null effects; persuasion is complex; polarised environments make it harder. This is compatible with Coppock’s view that persuasion works but typically with small effect sizes. So: AI may have *small but statistically significant* effects - not mind-control. ### 2.5 Why hyper-persuasion may never arrive McKenna flags (without fully arguing) a deeper hypothesis: persuasion might be intrinsically hard because humans are “epistemically vigilant” (Mercier; Sperber et al.). For evolutionary reasons, we’re tuned to detect untrustworthiness and misaligned interests. People can be tricked, but we’re not uniformly gullible. So it’s not obvious that future AI will blow past human persuasion limits; the ceiling might be set by human psychology more than by message-optimisation. --- ## 3\. Some conceptual clarifications This section does a lot of philosophical plumbing: it tries to make the later normative arguments less sloppy by tightening what counts as “persuasion” and who (if anyone) is doing it. ### 3.1 Persuasion as non-coercive, non-manipulative influence - and the messiness He starts with a standard-ish definition: persuasion influences attitudes/behaviour without coercion/force/manipulation (O’Keefe). He notes immediately that the persuasion/manipulation line is genuinely tricky because some persuasion may be manipulative in subtle ways (Brown; Marshall; Nettel & Roque; Tsai). But he brackets this because it’s not required for his later points. ### 3.2 Persuasion is intentional - and that’s a problem for “AI persuaders” Key claim: persuasion is intentional. If you change my mind accidentally, you influenced me but didn’t persuade me. He illustrates this with a deference case: someone learns a commentator’s view and adopts it purely due to deference. That’s influence, but persuasion only happens when the commentator *intends* to persuade via their communicative act. Then comes the classic snag: if AIs lack intentions (and maybe mental states altogether), then strictly speaking they can’t “persuade”. He canvasses options: 1. Redefine persuasion so it doesn’t require intention (he doesn’t like this). 2. Argue AIs have intentions/mental states (hard; he doesn’t pursue). 3. Treat AI as *persuasive technology* rather than as a persuader (this is his move, following Floridi). So the “persuader” is really the *human purveyor* deploying AI: AI is a medium/tool like radio/TV/internet in earlier mass communication debates. That preserves the intentionality condition because the human operator has intentions. ### 3.3 He explicitly brackets misinformation and focuses on a “better case” McKenna says: yes, misinformation worries are real and AI likely exacerbates them. But that’s too easy, and not the focus. Instead he wants a “best case” (or at least “better case”) scenario: AI used to provide *genuine arguments and reasons* for conclusions the purveyors endorse, aiming at rational persuasion rather than deception. That’s philosophically interesting because it asks: even if AI is used to do the “nice” version of persuasion, is it still troubling? ### 3.4 Rational persuasion: Mitchell & Douglas’s framework He adopts Mitchell & Douglas’s definition: A rationally persuades B to adopt attitude α if: (i) A brings it about that B adopts α, (ii) only by providing reasons for α, (iii) B adopts α because they recognise some of those reasons, (iv) A intends (i)-(iii). McKenna unpacks “reasons” as facts that count in favour/against (Mitchell & Douglas’s gloss). He notes you can swap in another ontology of reasons (e.g., reasons as mental states) if you want; it won’t derail the structure. This connects to a minimalist view of epistemic rationality: rational attitudes respond appropriately to evidence/facts/reasons; compatible with evidentialism and (with a caveat) reliabilism. Then he adapts the definition to AI: strictly speaking, the *purveyor* rationally persuades via AI when the purveyor uses the AI to deliver reasons, the target updates on them, and the purveyor intends that whole pathway - even if the purveyor doesn’t know exactly which reasons the AI will output. ### 3.5 The framing question Now the paper’s central evaluative puzzle is sharp: What is wrong (if anything) with: - using persuasive AI to rationally persuade people (e.g., on climate change, vaccination, foreign aid)? - being persuaded by AI-generated reasons and arguments? He flags what’s coming: - §4: an epistemic/rationality argument (AI persuasion makes resulting attitudes irrational or less rational). - §5–6: ethical/political arguments (especially about power, domination, and regulation politics). He also telegraphs his own view: he’s more convinced by the ethical/political arguments than the rationality one. --- ## 4\. The rationality argument This section is a careful attempt to make *epistemic* trouble stick - and then a partial failure report. ### 4.1 The basic claim Rationality Argument: persuasive AI undermines epistemic rationality of beliefs/attitudes/credences formed or updated through it, compared to formation via humans or independent reflection. Rationality here = being appropriately responsive to evidence/facts/reasons (the minimal conception set up earlier). ### 4.2 The “simple” version fails Simple version: it’s *just irrational* to be persuaded by AI. McKenna dismisses it: if AI produces genuine arguments/reasons, and you update on those reasons, that looks like perfectly ordinary rational updating. He uses a tax example: an AI chatbot gives good arguments for lowering taxes on the rich; someone thinks and updates. Whatever your politics, that’s structurally rational. ### 4.3 The “source credibility” reinforcement attempt A more promising angle: in real life, we often can’t evaluate empirical premises directly, so we rely on *source credibility*. Trust is part of rational evaluation. He notes that some trust frameworks emphasise interpersonal/agentic features that don’t map neatly to AI sources (Hinchman; Moran; Ross). There’s also emerging work on trust in AI (Carter; Simion & Kelp; Song). But he argues: this still doesn’t show AI is less trustworthy than humans *in general*. Humans lie, err, obfuscate. AIs hallucinate, inherit training-data bias, etc. Unless AI is typically far less reliable than humans, this doesn’t establish systematic irrationality. So even this buttressed version doesn’t really launch. ### 4.4 The “nuanced” version: AI might make attitudes less rational He then tries a weaker thesis: AI persuasion might not make beliefs irrational, but it might make them *less rational* than if formed otherwise. He tests two contrast classes: **(a) Pure “think for yourself” autonomy** Not realistic, and probably not even desirable as total independence. Recent accounts of intellectual autonomy often endorse *wise deference* rather than isolation (Carter; Roberts & Wood; Zagzebski). Also: rationality (as he defines it) doesn’t care where reasons come from; it cares whether you respond appropriately to them. So sourcing reasons yourself may produce *understanding* (Kvanvig; Pritchard; Matheson), but not necessarily greater rationality. **(b) Interacting with other people** Sometimes better, sometimes worse. A super-informed trustworthy friend can beat an LLM. A super-misinformed untrustworthy friend can be worse than an LLM summary. So the best you can say is: AI isn’t a substitute for the best human informants - but that’s not a damning criticism because such humans are scarce. ### 4.5 Where this leaves the argument McKenna’s conclusion is basically: the rationality argument doesn’t strongly establish that AI persuasion uniquely undermines epistemic rationality. At most it shows AI is not a replacement for unusually high-quality human epistemic relations. This sets up why he’s more impressed by the next arguments: the deepest worry isn’t individual rationality but structural power. --- ## 5\. The sophistry argument This is the conceptual and political heart of the paper. It’s where the title actually earns its keep. ### 5.1 Skin in the game: human persuaders vs AI He starts with a contrast: - Human persuaders have “skin in the game”: if they’re sincere, they’re committed to the truth; if insincere, they’re committed to *appearing* committed. - AIs (assuming no mental states) have no beliefs, no commitments, no pretence. They can generate persuasive output without any stake in truth. But he acknowledges an immediate objection: lacking skin in the game could be *good* - humans’ stakes motivate deceit. And since the human *purveyors* of AI can still lie, the “AI has no stake” point isn’t obviously a moral win or loss. So he looks for a better framing. ### 5.2 Why “AI as bullshit” doesn’t quite work He considers the Frankfurtian “bullshit” analogy (and cites recent “LLM bullshit” discussions). Frankfurt’s bullshitter is indifferent to truth; lies are different because the liar is truth-aware but conceals it. Temptation: call AI persuasion “bullshit” because it outputs without caring about truth. Problem: Frankfurtian bullshit requires communicative intentions. If AIs lack intentions, they can’t literally bullshit; they don’t intend to distract or signal values. And even shifting intentions to the purveyor doesn’t cleanly map onto the interpersonal structure of Frankfurt’s bullshitter. So “bullshit” is rhetorically satisfying but conceptually unstable given his assumptions. ### 5.3 The better analogy: sophistry He pivots to sophistry. In everyday usage, “sophistry” is clever, potentially misleading argumentation aimed at winning rather than truth-tracking. A sophist is an “arguer for hire”: they can produce arguments for a conclusion regardless of endorsement, and can flip sides on demand. McKenna’s idea: AI persuaders are like sophists. They can generate arguments without commitment, on demand, for any conclusion. And AI-generated persuasive content is like paying someone to write persuasion for you. ### 5.4 “Sophistry on steroids”: the marketplace of arguments Now the real argument begins. He introduces a “truism” about disagreement: Most contested domains have *many* good arguments on multiple sides; obvious cases are rare. Responsible thinking requires access to a broad set of arguments, and because individuals can’t generate them all, we rely on social mechanisms. This is where he introduces his key metaphor: Not just a “marketplace of ideas”, but a “marketplace of arguments”: the supply of arguments that people use as raw material for deliberation. Institutions (media, universities, think tanks) and individuals help supply and disseminate arguments. A healthy marketplace would (ideally) supply a roughly representative balance: central arguments on different sides, plus proportionate weaker ones. (He admits real discourses often fail, and people disagree about what “balanced” even means.) ### 5.5 How persuasive AI can distort and dominate the marketplace Persuasive AI is dangerous because it changes two constraints: 1. **Time and labour constraints collapse.** Humans can’t cheaply produce endless arguments and disseminate them widely. AI can: armies of arguers-for-hire and torrents of persuasive content can be generated quickly. Even if quality drops a bit, volume can swamp. 2. **Entry barriers are enormous.** Building/running LLM-scale persuasion requires massive resources and infrastructure. So only a few major companies/actors can play at that level. Put together: a small number of actors can *dominate the marketplace of arguments*, shaping what arguments are available, salient, and widely circulated. That is the paper’s central worry from the abstract: domination of the argumentative environment. ### 5.6 Why this is politically and socially ugly He draws out consequences: - This isn’t just “algorithmic bias”; it’s reinforcement of structural inequalities. Markets tend to benefit those with time/resources/skills, and those resources correlate with existing social power (Herzog). - Persuasive AI could intensify big tech’s outsized influence over political debate (Aytac). - Online public spheres are already imperfectly democratic; adding persuasive AI makes democratic conditions harder to achieve because the incentives of AI companies are not aligned with public good (often profit; sometimes murkier aims). ### 5.7 Control of information vs control of argument He ends with a final sharpening: even if controlling what people see isn’t the same as controlling what they think, persuasive AI threatens to narrow that gap by not just curating information but flooding and shaping the argumentative space itself. He repeats his scepticism about complete “mind control” (persuasion remains hard), but the direction of travel is still alarming. --- ## 6\. The political argument This last argument is about regulation, legitimacy, and disagreement: even if we tried to do “ethical governance” of persuasive AI, politics will eat it. ### 6.1 The abstract structure Suppose we agree on permissible-use conditions X, Y, Z. Those conditions would likely involve things like: - Are the claims true (or believed true)? - Are the arguments valid and based on (believed) true premises? - Is there a balance of arguments presented? - Are the purveyors’ intentions benign? The problem: on the very issues where persuasive AI will be deployed, people will strongly disagree about whether those conditions are met. Disagreement about truth, validity, balance, and legitimacy is endemic in politics. So regulation won’t neutralise political conflict; it will become *another arena* for it. ### 6.2 Vaccine example: persuasive AI as a political accelerant He gives a deliberately non-dated example: a new vaccine is approved; public health needs very high uptake (e.g., herd immunity). Communicators deploy persuasive AI (targeted messages, chatbots for hesitancy). Your judgement about legitimacy depends on your view of the vaccine. If you think costs outweigh benefits, you’ll see this as illegitimate manipulation - and you’ll campaign against it. Even if you don’t reject all persuasive AI in principle, you’ll deny the legitimacy conditions are satisfied *here*. ### 6.3 The dynamic: escalating suspicion and culture-war capture As more cases arise where persuasive AI is perceived as serving contested political ends, suspicion of the technology grows. The tools can’t stay neutral: they get drawn into cultural conflict. The upshot is that persuasive AI becomes not a solution to political disagreement but a generator of new battles over who controls the persuaders, under what rules, and to what ends. --- ## Data availability, references, acknowledgements, author info, ethics declarations, and publishing metadata These sections don’t advance the philosophical argument. They establish that no datasets were used, list the bibliography, and provide standard publication information (author affiliations, conflict of interest, licensing, dates, DOI, etc.). --- If you read the paper as a machine with three gears, it’s basically: (1) evidence says persuasion effects are usually modest and AI isn’t magic; (2) the real novelty is scalable deployment plus concentration of capacity; (3) that combination threatens to let a few actors structure the argumentative environment, and even regulating it will become a political battleground. Propaganda Mention Check Yes. The text includes the word “propaganda” multiple times—for example in the Introduction (“techniques of mass communication such as propaganda”) and in §2 when discussing “AI-generated propaganda” (e.g., Goldstein et al. 2024). Yes. He uses the word “propaganda” multiple times - e.g. in the Introduction (“techniques of mass communication such as propaganda”) and in §2 when discussing “AI-generated propaganda” (including the Goldstein et al. 2024 study). He’s basically *non-committal* on that classification, and it’s not an accident - he sets the paper up so his main worry doesn’t depend on calling it “propaganda”. In §2 he *does* talk explicitly about “AI-generated propaganda” (Goldstein et al.), but that’s in the empirical overview: evidence that LLM outputs can be as persuasive as human propaganda. In §5, when he introduces the “sophistry on steroids” / “marketplace of arguments” worry, he *switches frames*. The target phenomenon there is: a few actors can mass-produce arguers-for-hire and persuasive content that may be perfectly “rational” in form (reasons, arguments) and not necessarily deceptive. That’s why he explicitly says he’s setting aside misinformation and focusing on a “best case” where the outputs are “genuine arguments” aimed at rational persuasion. So: if by “propaganda” you mean something essentially deceptive/manipulative, ideologically loaded, or non-truth-tracking, then his “AI sophists” case is *deliberately not* labelled propaganda. It’s meant to be worrying even when it’s *not* propaganda. If by “propaganda” you mean something broader like “organised mass persuasion in service of an agenda”, then you *could* redescribe his marketplace-domination scenario as propaganda-like. But he doesn’t do that, and he doesn’t clarify whether he’d accept that broader usage - he keeps “propaganda” as one background example of coercive/manipulative influence, and treats his core case as a distinct structural threat (scale + concentration distorting the argumentative environment). Net: he doesn’t say “this isn’t propaganda”, but the rhetoric and the “best case” setup strongly suggest he wants the sophistry-at-scale worry to be *separable* from propaganda, not just a subset of it. Partly persuaded - but only *conditionally*, and the condition is doing a lot of work. On Bonard/Contesi/Marques’ (refined) Ross-style account, a communicative act counts as propaganda iff it (1) conveys an *epistemically defective* message (broadly: false/misleading/unwarranted, including via implicature, presupposition, omission, Gricean Quantity, etc.), (2) is produced by/on behalf of an institution or cause *with the intention to persuade for direct political ends*, (3) targets a socially significant group, and (4) the producer is aware (or should be aware) of the defectiveness. Now map that onto McKenna’s *best case* (mass deployment of “genuine arguments” aimed at rational persuasion, not deception). 1. **If “best case” really means**: the content is not misleading, not cherry-picked in a way that creates false implicatures, not omitting relevant counterevidence in a way that predictably misleads, not packaged so as to generate unwarranted presuppositions, etc. - then **it won’t satisfy (1)**. And on this account, *no epistemic defectiveness, no propaganda*. The mere fact that it’s produced at scale, or by “AI sophists,” doesn’t by itself make it propaganda on their definition. 2. **But** their notion of epistemic defectiveness is intentionally *broad*: it explicitly includes misleadingness through omission and pragmatic effects (they use the war-report cherry-picking case and Gricean Quantity to make that point). So here’s the pressure point: McKenna’s scenario is *structurally* about flooding and shaping the “marketplace of arguments.” If that flooding is deployed in a way that predictably makes audiences *infer* (pragmatically) that “these are the salient arguments,” “this is the reasonable balance,” “this is what serious people think,” etc., while systematically crowding out alternatives, then you can argue that the overall communicative act (or campaign) is **epistemically defective via misleading selection/omission**, even if each individual argument is locally decent. That’s exactly the kind of route their paper keeps open: defectiveness can be in what’s implicated, omitted, or made salient, not just in outright falsehood. So I land here: - **McKenna’s best case is not automatically propaganda** on their definition, because they build propaganda around *epistemic defectiveness*. - **But it can become propaganda** if the scale-and-control dynamic *functions* as an epistemically defective way of persuading (misleading “balance of reasons,” engineered salience, systematic omission/crowding-out), and the relevant institutions are (or should be) aware of that defectiveness while pursuing direct political ends. In other words, your line is defensible, but the hinge is: can “dominating the marketplace of arguments” be reconstrued as *epistemic defectiveness* (in their broad, pragmatic sense), rather than as a merely political pathology. That reconstruction is plausible - but it isn’t forced by their framework, and you’ll want to be explicit about *where* the defectiveness sits (implicature/omission/salience/bypass), not just point to “scale” as such. Okay, good. Um I definitely agree with you on your point number two. Okay, so you can flood the marketplace with sound arguments but you're still they can still be epistemically affected via misleading selection or omission. Good. Let me think about number one though, the best case. The thing is though, if the content is not misleading if there's if it's not creating false implicatures and omitting relevant counter-relevant evidence, etc. etc. then I fail to see why McCare's best case would be problematic at all. Because wouldn't the so-called propagandists simply be spreading truth? Please give me a long, deep, and detailed answer, with lots of options/ideas: as we are in the 'kicking ideas around' stage, I want as much feedback as you can give me so that i have a lot to consider. 1. There’s a real tension here, and it’s useful to name it cleanly. If “best case” really means *all* of the following at once: every argument is sound, the presentation is not misleading (no false implicatures, no cherry-picking, no strategic omissions), the overall exposure is balanced across sides, provenance is transparent, targeting doesn’t exploit vulnerabilities, and the audience has genuine opt-in control, then yes - it starts to look like the “problem” evaporates. That would be less a propaganda machine and more like a weirdly efficient public library that also happens to be very chatty. So either (a) McKenna’s worry is actually parasitic on some kind of epistemic defectiveness after all (selection effects, framing, salience, omission), or (b) there are non-epistemic ways for something to be politically troubling even when it’s truth-tracking. The interesting work is in (b). And I think there are several routes. 1. Truth can still be politically troubling when it’s used as a tool of domination rather than as part of a fair deliberative process. One way to put it: even if the content is “spreading truth,” there’s a difference between *truth entering a shared deliberative space* and *truth being deployed as a strategic weapon by concentrated power*. Imagine a world where an all-powerful actor has a monopoly on broadcasting only true propositions and valid arguments. That actor could still shape what gets discussed, when it gets discussed, which issues feel urgent, which trade-offs feel salient, what counts as “common sense,” and which identities or coalitions form around which beliefs. None of that requires lying. It requires agenda-setting and control of attention. So the worry shifts from “are the claims false?” to “who gets to steer the epistemic environment?” 1. “Flooding” can be coercive in a procedural sense even if it’s epistemically pristine. There’s a kind of soft coercion that comes from saturating someone’s cognitive bandwidth. Even sound arguments impose costs: attention, time, emotional energy, opportunity cost. If one actor can generate an effectively infinite stream of compelling, valid, context-sensitive arguments, they can force everyone else into permanent triage mode. That matters because real deliberation has scarcity constraints. Human beings can’t check everything. They can’t answer everything. They can’t even read everything. If I can force you into a constant defensive crouch, I can “win” without ever lying - simply by making your ability to respond collapses under volume. That’s not deception; it’s a resource attack on the public sphere. This is one way to read McKenna’s “marketplace of arguments” point: the harm can be less about truth and more about the conditions for *fair contestation*. 1. A “perfectly true” campaign can still be propagandistic if propaganda is partly about bypassing deliberative agency. Some theories of propaganda (not all) treat it as centrally about manufacturing assent through techniques that short-circuit the audience’s reflective agency, not merely about falsehood. Even on your stricter “no misleading implicatures, no omissions” constraint, you can still have bypass in at least two ways. First, personalisation: if an AI system learns what moves you and when you’re cognitively vulnerable, it can time and tailor *true* arguments to hit you when you’re least able to evaluate them well. That’s not lying and it might not even involve misleading content. But it’s parasitic on attentional control, fatigue, fear, social pressure, and so on. Second, repetition and reinforcement: saying only true things, but doing it with the cadence and persistence of behavioural conditioning. Again, no deception needed. Just optimisation for conversion rather than understanding. So: “true propaganda” is not a contradiction *if* propaganda is defined partly by how assent is produced, not only by the truth-value of what’s said. 1. Even if each message is non-misleading, the *system-level* effect can still be epistemically bad, because epistemic quality is not pointwise. This is a nerdy but important distinction: epistemic evaluation can be local (this argument is sound, this claim is true) or global (this informational environment reliably produces well-grounded beliefs across a population). You can have an environment where everything said is true, and yet the overall effect is to produce brittle, polarised, overconfident, or one-sided belief structures because of dynamics like: - attention capture (some truths get amplified massively, others remain obscure), - path dependence (early salience locks in frames), - coalition formation (truths become identity-markers), - “winner’s curse” confidence (people infer that what’s everywhere must be what’s most important or most decisive). Notice: none of this requires false implicatures in the content itself. It’s about how humans infer significance from prominence. If you flood the zone, people will infer “this must be the key consideration,” even if the content is technically impeccable. That’s a pragmatic effect of distribution, not of individual messages. 1. Accountability and responsibility still bite even in a truth-only regime. Suppose the arguments are sound and the conclusions are true. You still have the question: who is answerable for the campaign’s goals, trade-offs, and value-laden choices about what to push? McKenna’s “AI sophists” frame matters here. “Arguers for hire” can generate reasons for any side. So the central political issue is not whether the arguments are valid but who is choosing which truths to mobilise, for which ends, with which priorities, and with what oversight. Even a truth-only persuasive system can be objectionable if it creates a world where major questions are effectively decided by a handful of actors who control the persuasive infrastructure, because democratic legitimacy is not just “getting to the truth.” It’s also about who gets a say in shaping common life. 1. There’s a simple escape hatch: maybe McKenna’s “best case” is not your “best case.” McKenna’s “best case” is roughly: no misinformation, genuine arguments, rational persuasion. But that’s consistent with a lot of things you are implicitly ruling out, like: - systematically privileging one side’s arguments (even if they’re good), - relentless personalisation, - making some topics permanently salient and others invisible, - asymmetries in who gets to speak at scale, - persuasive design optimised for conversion rather than mutual understanding. So you might be holding the “best case” to an idealised standard of deliberative fairness (balanced exposure, neutral agenda-setting, symmetric capacities), whereas McKenna is holding it to an epistemic standard that is mostly about not lying or deceiving. If that’s right, then his worry survives: the thing can be “truthy” at the level of content and still structurally undemocratic at the level of power. 1. A menu of positions you could take, depending on what you want your argument to do. Position A: *If you really remove all epistemic defectiveness and all procedural unfairness, the worry disappears.* This makes McKenna’s case depend on non-ideal conditions (and that’s fine - most political philosophy is non-ideal). Your task then is to show that scale + concentration make those non-idealities basically inevitable. Position B: *Even in a truth-only world, domination is objectionable.* Here you lean into a republican or procedural-democratic intuition: domination is about asymmetric power over others’ options, not about whether the powerful actor uses that power benevolently or truthfully. A benevolent dictator who only tells the truth is still a dictator. Position C: *Truth-only persuasion can still be “propaganda” because propaganda is about technique and governance, not falsity.* This pushes the concept of propaganda in a direction where “misleadingness” is not necessary. It’s a respectable move, but you’ll want to defend it explicitly, since many definitions build in epistemic defectiveness. Position D: *McKenna’s worry is best understood as “epistemic pollution” rather than propaganda.* On this line, calling it propaganda is optional and maybe rhetorically costly. The heart is: flooding degrades the epistemic commons by making attention and response capacity the scarce bottleneck, and by enabling agenda capture. 1. If you want a sharp thesis that preserves your “truth-only” concession while keeping McKenna’s worry alive, here’s a very compact formulation. Even if persuasive AI spreads only truths and sound arguments, it can still be politically troubling because it enables a small number of actors to exercise disproportionate control over the epistemic environment (salience, agenda, attention, personalisation, volume), and that control can undermine fair democratic contestation and deliberative agency. In that sense, the wrongness is not located in the truth-value of what’s said but in the structure of power governing what can be heard, considered, and answered. That keeps the best-case concession intact while making the problem nontrivial. A final weird-but-useful thought: a world with perfectly true persuasive flooding might be close to a world where the public sphere becomes a permanently “open tab” pop-up. Even if every pop-up is correct, you’d still want an ad-blocker for your mind. Right - so on Tuttle Ross’s Epistemic Merit Model (as reconstructed and then *refined* in the paper you attached), the key point is this: **propaganda is not defined primarily by scale, sophistication, or even political intent; it’s defined by *epistemic defectiveness used for political persuasion*.** That means a lot of the “(b)” worries I raised *only* count as propaganda if you can show they introduce epistemic defectiveness at the level that Ross cares about. A small clarification first: in the attached paper, Ross’s *initial* model is presented with **three** conditions, and the authors’ *preferred refinement* gives **four** conditions, not six. The refined version says (roughly) that a communicative act counts as propaganda iff: (1) it conveys an epistemically defective message (in a broad sense that includes implication/implicature and illocutionary/perlocutionary intent), (2) it’s produced by/on behalf of an institution or cause with intent to persuade for direct political ends, (3) it targets a socially significant group, and (4) the producer is aware (or should have been aware) of the defectiveness. That’s the yardstick we’re using. Now, with that yardstick in place: do any of my “(b)” suggestions come out as propaganda? ## 1\. Some of my “(b)” points are not propaganda on Ross’s model (unless you add something) Take the worries that are basically about **domination** or **legitimacy** in a democratic public sphere: - “A few actors steer the epistemic environment.” - “Even true information can be used as a tool of domination.” - “Concentrated control of agenda-setting undermines democratic contestation.” These can be *politically alarming* without automatically being propaganda **because Ross builds epistemic defectiveness into the definition**. If an actor really were doing nothing epistemically defective - no misleading selection, no manipulative implicatures, no distorted presuppositions, no epistemically toxic emotional engineering, no creating unwarranted connections to other beliefs - then Ross’s framework is strongly pulled toward: *not propaganda*. So if you want to classify McKenna’s “best case” as propaganda **while stipulating away all epistemic defectiveness**, you’re (implicitly) rejecting Ross’s core idea that propaganda is essentially epistemically defective. That’s not incoherent, but it’s a revisionary move: it means you’re pushing the concept of propaganda toward something like “politically problematic persuasion under conditions of domination,” rather than “epistemically defective political communication”. ## 2\. But several of my “(b)” points do become propaganda on Ross’s model, once you notice where defectiveness can live The crucial thing Ross (and the attached paper’s refinements) allow is: **defectiveness can be pragmatic and systemic**, not just “the literal content is false”. So even if every individual sentence is true and every individual argument is valid, you can still get propaganda if the overall communicative act predictably makes the audience take on **false or unwarranted implications, presuppositions, salience inferences, or emotional framings**. Here are the “(b)” routes that *do* fit Ross fairly naturally. ### A. Flooding as misleading Quantity (the “completeness” illusion) Suppose an institution floods the space with sound arguments all pointing one way. Even if each argument is good, the *campaign* can still be epistemically defective if it reliably generates false pragmatic inferences like: - “These are the main considerations.” - “There really aren’t serious counterarguments.” - “The balance of reasons is decisively on this side.” - “This is the settled, responsible view.” Those are classic Gricean Quantity-style effects: people infer completeness or representativeness from saturation. If the communicator *intends* (or should foresee) that saturation will produce that kind of false inference, you have a very Ross-friendly form of defectiveness: not lying, but engineering a misleading epistemic picture. On this route, “flooding” isn’t just a resource attack; it’s also a way of making the audience form beliefs on the basis of an epistemically defective representation of the argumentative landscape. ### B. Agenda-setting as epistemic distortion (salience = “what matters”) Another Ross-compatible move is: **controlling what becomes salient can mislead about what is important**. Humans treat prominence as evidence of importance. So a system that massively amplifies some true considerations while burying others can make audiences rationally (but incorrectly) update on “importance”. If the resulting worldview gives certain considerations an evidential weight they don’t deserve, that’s an epistemic defect: beliefs become “connected to other beliefs in ways that are inapt, misleading, or unwarranted” (this phrase is doing a lot of work in Ross). So, again, the defect isn’t “the arguments are unsound”. The defect is “the epistemic environment pushes people into an unwarranted map of the space of reasons”. ### C. Personalisation and timing that exploit vulnerability (true content, epistemically dodgy uptake) Here the question is: can personalisation make persuasion epistemically defective even when the content is true? Yes - if the personalisation is used to *engineer the uptake conditions* so that the audience is likely to accept conclusions in ways that are epistemically substandard: e.g. when tired, anxious, socially threatened, or otherwise cognitively constrained. Ross’s model (especially as broadened in the paper you attached) is happy to treat illocutionary/perlocutionary design as relevant to defectiveness. If the system is optimised not for the audience’s good epistemic position but for conversion, it can push people into belief formation that is unwarranted relative to their reasons and capacities at that moment. This isn’t the cleanest fit (because you’ll want to explain why this counts as “epistemic defectiveness” rather than merely “unfair influence”), but it *can* fit: the idea would be that the campaign is designed so that people adopt attitudes *without* appropriately responding to reasons, or on the basis of epistemically inappropriate emotional states. ### D. Repetition/conditioning that detaches assent from reasons Relatedly, repetition can produce “acceptance” that is increasingly independent of the merits of the reasons. If the communicative strategy is: keep presenting true reasons, but in a way that the psychological mechanism of uptake becomes more like habituation than evaluation, the epistemic status of the resulting belief can become defective. Again: the arguments are true, but the belief is formed/maintained in a way that is not appropriately responsive to evidence. If a political institution knowingly designs that mechanism, the “defectiveness” can be pinned to the perlocutionary architecture. ### E. The “balanced marketplace” appearance as a misleading implicature This is a very direct Ross move. If the system presents itself as “just offering reasons” or “just facilitating deliberation” while being structurally controlled so as to crowd out alternatives, then the defectiveness can be located in an implicature like: “You’re seeing an open, pluralistic space of debate.” That implicature can be false even if every displayed argument is sound. ## 3\. So where does that leave your worry about McKenna’s “best case”? It leaves you with a fork: ### Option 1: Concede that a fully non-misleading best case is not propaganda on Ross If you really mean: no misleading selection, no omission, no pragmatic distortion, no manipulative perlocutionary design - then Ross won’t call it propaganda, and you’ll need to say McKenna’s worry is about something else (domination, democratic legitimacy, epistemic justice, etc.). ### Option 2: Argue that McKenna’s “best case” is not your best case - and that his scenario almost inevitably generates epistemic defectiveness This is, I think, the most promising route if you want to keep Ross and still call it propaganda. You’d argue: once you build in (i) scale, (ii) concentrated control, and (iii) competitive political ends, then **misleading selection/omission/salience effects are not optional add-ons; they’re structurally baked in**. Even with individually sound arguments, the system-level communicative act will tend to mislead about the balance and relevance of reasons, because the whole point of “flooding” is to reshape what people take to be the live options and the decisive considerations. So the propaganda-ness comes not from any single argument being bad, but from the campaign functioning as an epistemically defective representation of the deliberative space. ### Option 3: Reject Ross’s defectiveness requirement (revisionary concept) If you want to say: “Even perfectly non-misleading truth-spreading can be propaganda when it’s deployed under domination,” then you’re explicitly revising away from Ross’s model. That might still be worth doing, but you’d need to explain why “propaganda” should track domination rather than epistemic defectiveness. --- If your aim is specifically: *“Does McKenna’s best case count as propaganda on Ross?”*, the clean answer is: It counts as propaganda **only insofar as the scale-and-control strategy makes the overall communicative act epistemically defective in Ross’s broad sense** (misleading by omission, implicature, presupposition, distorted salience, or perlocutionary design that predictably produces unwarranted belief-formation). If you stipulate all of that away, Ross-style propaganda drops out - though McKenna-style political worry might remain under a different label. That’s also a nice dialectical point you can use: **Ross gives you a principled reason why “truth-only domination” won’t be propaganda - and that forces you to decide whether you want “propaganda” to be an epistemic notion or a power/legitimacy notion.** I'm not quite sure I understand your conclusions here. Um so a moment ago I talked to you about what might be thought of as a dilemma for McKenna. Okay, which is if there really is, if everything really is all above board, then it's not really propaganda and if it's not all above board then it is propaganda. I didn't quite see which of your elaborations of his ideas weren't propaganda. In short, I'm somewhat confused. "It counts as propaganda only insofar as the scale-and-control strategy makes the overall communicative act epistemically defective in Ross’s broad sense (misleading by omission, implicature, presupposition, distorted salience, or perlocutionary design that predictably produces unwarranted belief-formation). If you stipulate all of that away, Ross-style propaganda drops out - though McKenna-style political worry might remain under a different label." what? We've really gotta work on your writing style, because it's like I said, you know, I'm a philosophy academic, I'm used to reading difficult stuff, but you write in this sort of weird and the only way that can use to describe as terse it feels it so thin that it's almost incomprehensible. You’re right to flag a dilemma-shaped structure here, and you’re also right that my last paragraph was needlessly compressed. Let me slow it down and separate the moving parts. ## 1\. The one thing that drives everything on Ross On the (Tuttle) Ross-style account you’re using, **“propaganda” requires epistemic defectiveness**. Not “politically powerful”, not “scaled”, not “effective”, not “centralised” - *epistemically defective* in a broad sense (falsehood, misleadingness by omission, misleading implicature/presupposition, etc.), plus the political-intent and targeting conditions. So: If it’s *fully above board epistemically*, then **it is not propaganda on Ross**, even if it’s deployed by powerful actors at huge scale. That’s not me taking a stand on whether it’s “fine”. That’s just what follows from Ross making epistemic defectiveness a necessary condition. ## 2\. Where your dilemma lands Your dilemma was: - If everything is genuinely above board, it’s not propaganda (and then why is it troubling?). - If it’s not above board, then it is propaganda. On Ross, that dilemma is basically correct as far as the label “propaganda” goes, because “above board” (in the relevant sense) just means “not epistemically defective”. What I should have said more clearly is: **McKenna’s overall worry is broader than Ross-propaganda.** Some versions of McKenna’s worry *do* involve epistemic defectiveness (so they can be propaganda on Ross). Other versions don’t (so they won’t be propaganda on Ross), even though McKenna might still think they’re politically dangerous. ## 3\. Which of McKenna-type worries are not propaganda on Ross? Here are the ones that can be non-propaganda (again: on Ross), provided the communication really is epistemically clean. ### (A) Domination/legitimacy worries A few actors control the infrastructure of persuasion, can set agendas, and can steer what gets attention. Even if they do this while only putting forward true claims and sound arguments and while not misleading anyone about balance, Ross won’t call it propaganda unless there’s epistemic defect. This is the “benevolent-but-dominating” case: **politically troubling, not propaganda (on Ross)**. ### (B) Procedural unfairness that is not epistemic defect For example: the situation is unfair because ordinary citizens can’t match the volume or sophistication, or because the communicative environment becomes exhausting. That might be democratically unhealthy, but unless it *misleads* people (in Ross’s epistemic sense), it still won’t count as propaganda. Think “flooding as a resource attack” *without* misleadingness. That’s coherent as a political complaint, but it doesn’t satisfy Ross’s epistemic condition. So these are cases where McKenna can still worry, but Ross won’t let you call it propaganda. ## 4\. Which McKenna-type worries do become propaganda on Ross? Now the other side: many “scale + control” scenarios *almost inevitably* produce epistemic defectiveness even if the individual arguments are sound, because the defectiveness can sit in selection, omission, and pragmatic effects. Examples: - Flooding one side so thoroughly that people predictably infer “there aren’t serious counterarguments” (a false implicature about the balance of reasons). - Systematically amplifying some truths while burying others so that audiences form distorted beliefs about what is important or decisive (salience distortion). - Presenting a curated, one-sided argumentative stream as if it were a representative “marketplace of arguments” (misleading presupposition/implicature about openness and balance). Those *are* epistemically defective in Ross’s broad sense even if each displayed argument is sound. So those cases can be propaganda. ## 5\. The clean way to see the relationship McKenna’s “persuasive AI at scale controlled by a few actors” is **a bigger phenomenon** than “propaganda” on Ross. Ross-propaganda is a **subset** of McKenna’s phenomenon: - McKenna-worry can exist **with** epistemic defect (→ propaganda on Ross). - McKenna-worry can exist **without** epistemic defect (→ not propaganda on Ross, but perhaps still domination/unfairness/legitimacy problems). That’s the bit I compressed too hard. ## 6\. Rewriting the sentence you quoted in plain English Here’s what I meant, in a form that should actually be readable: On Ross’s definition, it’s propaganda only if the campaign is epistemically defective in some way - for example, if it misleads people by cherry-picking, omitting important counter-considerations, creating false implicatures, or distorting what looks salient or representative. If you stipulate that none of that is happening - that the campaign is genuinely transparent, non-misleading, and epistemically balanced - then it won’t count as propaganda on Ross’s account. McKenna could still think the situation is politically dangerous (because a few actors control the persuasive infrastructure), but that would be a different kind of worry than “propaganda” as Ross defines it. That’s all I was trying to say. The next useful step (conceptually) is deciding whether you want to defend: (i) “McKenna’s best case is *never* truly epistemically clean once scale + concentrated control are in play” (so it *does* collapse into propaganda on Ross), or (ii) “even a fully epistemically clean best case is politically objectionable” (but then you should probably stop calling it propaganda, at least on Ross’s usage, and call it domination/agenda control/epistemic power instead). "(A) Domination/legitimacy worries A few actors control the infrastructure of persuasion, can set agendas, and can steer what gets attention. Even if they do this while only putting forward true claims and sound arguments and while not misleading anyone about balance, Ross won’t call it propaganda unless there’s epistemic defect. This is the “benevolent-but-dominating” case: politically troubling, not propaganda (on Ross). (B) Procedural unfairness that is not epistemic defect For example: the situation is unfair because ordinary citizens can’t match the volume or sophistication, or because the communicative environment becomes exhausting. That might be democratically unhealthy, but unless it misleads people (in Ross’s epistemic sense), it still won’t count as propaganda. Think “flooding as a resource attack” without misleadingness. That’s coherent as a political complaint, but it doesn’t satisfy Ross’s epistemic condition. So these are cases where McKenna can still worry, but Ross won’t let you call it propaganda." I am not sure I am convinced. please steelman the case for and against each of these cases as being propaganda and get back to me. Please give me a long, deep, and detailed answer, with lots of options/ideas: as we are in the 'kicking ideas around' stage, I want as much feedback as you can give me so that i have a lot to consider. Got it. Let’s do this properly: for each case, I’ll steelman **the “not propaganda (on Ross)”** line and the **“actually it *is* propaganda (on Ross)”** line. The key is that *Ross-style propaganda is an epistemic notion*: you don’t get propaganda unless you get **epistemic defectiveness** (broadly construed), plus political intent, a socially significant target, and (on the refined version) awareness/culpable ignorance of the defect. I’ll also flag a meta-point up front: whether (A) or (B) counts as propaganda depends hugely on two interpretive choices. 1. How broad is “epistemically defective” supposed to be (just falsity/misleadingness, or also degraded epistemic agency, manufactured salience, forced heuristics, etc.)? 2. What is the *unit* of analysis: an individual message, or a whole campaign/ecosystem (where “the message” includes distribution, prominence, repetition, timing, and the implied picture of the argumentative landscape)? Those two choices basically decide the verdict. ## 1\. Case A: “benevolent-but-dominating” control of persuasive infrastructure The scenario: a few actors control the persuasive infrastructure, set agendas, steer attention, but (stipulation) they only put forward true claims and sound arguments and they do not mislead about balance. ### 1\. Steelman: Not propaganda on Ross Here’s the strongest version of the “not propaganda” case, keeping Ross as fixed. First, **Ross makes epistemic defectiveness necessary**. If the content is true, the arguments sound, and (crucially) the system is genuinely non-misleading about the balance and representativeness of what’s presented, then the epistemic condition fails. That’s not a “loophole”; it’s the point of the model. Propaganda, on this view, is distinguished from other politically significant speech by its epistemic vice. Second, **agenda-setting is not automatically epistemic defect**. “This is what we’re discussing today” can be normatively controversial and politically powerful without being epistemically defective. Societies must prioritise. Newspapers must choose front pages. Parliaments must choose what to debate. Those choices can be biased or self-serving, but unless the resulting communicative acts mislead, they aren’t propaganda on an epistemic-merit account. Third, **domination is a political category, not an epistemic one**. You can have domination with full epistemic cleanliness (think: an authoritarian regime that publishes accurate statistics and valid policy arguments, transparently). That may be morally and democratically repellent, but it is not therefore “propaganda” on Ross. Under Ross, “propaganda” is not just “politically troubling persuasion”; it is *epistemically defective* persuasion. Fourth, even the “intent” condition may fail. If the infrastructure is run with the stated aim of “supporting civic deliberation” rather than “securing direct political ends,” then it may be politically influential without satisfying Ross’s political-purpose criterion. (Of course, you might argue that “supporting deliberation” is always politically loaded; but the steelman is that Ross wants a fairly direct political end.) So the clean conclusion: **control and agenda power can be democratically troubling without being propaganda**, because Ross deliberately builds propaganda around epistemic defect. This is the “separate the concepts” route: propaganda tracks epistemic corruption; domination tracks power. ### 2\. Steelman: It is propaganda on Ross Now the best case for calling even this “benevolent domination” propaganda is to show that the stipulation “not misleading about balance” is either (i) unstable in practice, or (ii) too narrow to capture Ross’s own broad notion of epistemic defectiveness. There are a few powerful ways to do that. First, **domination over attention tends to create false pragmatic inferences even when literal content is true**. People treat prominence and repetition as evidence of importance, consensus, and decisiveness. So if one actor structurally controls what gets seen, that actor can predictably cause audiences to acquire false beliefs like “these are the main considerations” or “this is where the weight of reasons lies,” even if they never explicitly say that. If Ross counts misleading implicature/presupposition/Quantity effects as epistemic defect, then agenda domination can be propaganda *because it predictably generates epistemically defective uptake*. Second, you can locate the defect not in any single claim, but in the **picture of the deliberative landscape** that the communicative system implicitly conveys. Even if every included argument is sound, the system may still communicate, by selection and prominence, an epistemically defective representation of the space of reasons. If the implied map is distorted, the message (understood broadly) is epistemically defective. This is the “unit of analysis” move: propaganda is not a sentence; it’s a communicative act that includes distribution and salience. Third, Ross’s talk of epistemic defect can be read as including **inapt or unwarranted belief connections**. Infrastructure control can cultivate unwarranted connections like “rational person = person who endorses the platform’s favoured conclusion,” “serious debate = the debate as curated here,” “credibility = what appears in this stream.” Those are epistemically poisonous connections even if the platform also contains plenty of true claims. The propagandistic defect is the engineered epistemic dependence and the norms of credibility it installs. Fourth, “benevolent domination” can still satisfy the **political-ends** condition if the end is not “get people to believe proposition p” but “maintain hegemony over the agenda” or “stabilise a political order.” On a broad reading, maintaining control of the public sphere is a direct political end. If the institution is steering attention in order to preserve its position, that looks politically purposive in exactly the way Ross wants. Fifth, the awareness condition becomes easy: any sophisticated actor controlling the attention architecture **should know** these pragmatic effects occur. If you knowingly create an epistemic environment that predictably produces distorted inferences about importance, balance, or consensus, you are culpably implicated in the defectiveness. So the pro-propaganda conclusion is: even “truth-only” infrastructure control can be propaganda **because it can be epistemically defective at the level of implicature, salience, and the implied map of reasons**, and the controller can foresee this. The deepest disagreement between the two sides is whether “epistemically defective” extends to *system-level distortion of salience and the resulting uptake*, or whether it is limited to more straightforward misleadingness in content. ## 2\. Case B: “procedural unfairness / flooding as a resource attack” without misleadingness The scenario: one side floods the environment with sound arguments/true claims so others can’t keep up; the environment becomes exhausting and asymmetrical. Stipulation: it does not mislead (in Ross’s sense). ### 1\. Steelman: Not propaganda on Ross Here the most robust “not propaganda” case is even cleaner. First, **exhaustion and asymmetry are not epistemic defect in the message**. They are features of the environment and of the audience’s limited capacities. Ross’s model is meant to classify communicative acts by epistemic merit, not to classify any practice that burdens deliberation as propaganda. If everything said is true, arguments sound, and no misleading implicatures/omissions occur, then propaganda does not apply. Second, **being hard to answer is not the same as being misleading**. A barrage of correct arguments can be obnoxious, dominating, even anti-democratic — but epistemically it might be impeccable. Under Ross, the label propaganda is not supposed to be a catch-all for “unfair rhetorical advantage.” Third, “flooding” could be interpreted as a form of **petitioning/public argument at scale**, akin to mass pamphleteering, high-volume op-eds, relentless but accurate advocacy. Democracies can be noisy. Noise can be unhealthy without being propaganda. Fourth, if the audience *knows* they are being flooded and knows the source, the campaign might not create any false beliefs at all — it might just create fatigue. Fatigue is not an epistemic defect in Ross’s sense unless it’s being used to induce epistemically unwarranted assent. So the “not propaganda” verdict is: **this is procedural injustice or domination, not propaganda**, unless you can show epistemic defect. ### 2\. Steelman: It is propaganda on Ross To steelman the opposite, you need to show that “resource attack flooding” *cannot* really remain non-misleading, or that it generates epistemic defectiveness in Ross’s broader sense even if each unit is sound. There are several strong routes. First, **flooding predictably produces false “consensus/importance” inferences**. People infer from ubiquity that something is widely supported, socially central, or evidentially decisive. This is a pragmatic effect of repetition and saturation, not of explicit content. If Ross counts those pragmatic inferences as part of what a communicative act conveys, then flooding can be propaganda because it predictably causes epistemically defective beliefs about the state of reasons or consensus. Second, flooding can function as **omission by crowding out**. Even if the propagandist never lies, the volume itself can ensure that counterarguments are not encountered. If epistemic defectiveness includes misleading by omission, then “crowding out alternatives” is a way of creating an incomplete and therefore epistemically defective evidential basis for belief formation. The audience ends up with a systematically skewed sample of the argument space. That skew can count as epistemic defect even if each encountered argument is sound. Third, you can argue that flooding aims at **non-rational uptake mechanisms**. Even when the arguments are valid, the strategy may be to force people into heuristics: accept the conclusion because you cannot process everything, because the cost of resistance is too high, because you want to end the cognitive assault. If the belief is adopted under conditions that undermine appropriate responsiveness to reasons, then the resulting attitude is epistemically defective, and a campaign designed to induce that is propaganda-ish on an epistemic merit model. Fourth, the “without misleadingness” stipulation may be unstable because flooding itself creates an implicit claim: “this deserves your attention repeatedly.” If that attention-demand is disproportionate to the evidential weight or public importance, then the act of repeatedly pushing it is epistemically defective as a representation of significance. That is a salience-based epistemic defect again, now driven by repetition rather than agenda-setting. Fifth, as with Case A, the awareness condition is plausible: sophisticated actors know that flooding shifts beliefs by altering salience and perceived consensus, not merely by adding reasons. If they exploit that mechanism for political ends, you get Ross-style propaganda quite quickly. So the pro-propaganda verdict is: flooding becomes propaganda insofar as it **misleads pragmatically**, **skews the evidence base by crowding out**, or **engineers epistemically substandard uptake** — all of which can happen while every individual argument remains sound. ## 3\. A way to reconcile the apparent clash (and why you’re not crazy to be unconvinced) Your resistance makes sense because my earlier “not propaganda” descriptions sounded like they were imagining a magical scenario where the effects of scale and concentration are *purely procedural* and never spill into epistemics. But in real life, and arguably even in principle given human psychology, scale-and-control *tends* to create epistemic distortions via: - salience (“what matters”), - consensus inference (“what people think”), - evidential weight inference (“how strong the case is”), - availability (“what arguments exist at all”). If Ross’s “epistemic defectiveness” is broad enough to include those pragmatic/systemic effects, then your intuition is right: **many “procedural” pathologies become epistemic pathologies**, which makes them propaganda on Ross. So the live philosophical question isn’t “is domination/flooding propaganda?” in the abstract. It’s: 1. Does the Ross-style account treat systemic salience/availability distortions as epistemic defectiveness? 2. Does “convey” include what the campaign predictably gets audiences to infer from distribution and prominence? 3. Is “omission” satisfied by crowding-out effects, or only by selective content within a bounded message? Different answers yield different classifications. ## 4\. Practical “options” for how you can frame this in your own argument Option 1: Strict Ross reading Propaganda requires fairly direct epistemic defect in content or in clear pragmatic implicature. On this reading, some domination/flooding cases won’t count as propaganda, though they remain politically objectionable under other concepts (domination, agenda capture, democratic illegitimacy). Option 2: Broad Ross reading (campaign-level pragmatics count as conveyed content) Propaganda can be “true” at the sentence level and still propaganda because the communicative act is epistemically defective at the level of salience, availability, and the implied map of reasons. On this reading, most serious domination/flooding cases do count as propaganda. Option 3: Two-tier diagnosis Call the epistemic part “propaganda” when it’s present, but explicitly separate a second wrong: “epistemic domination” or “deliberative domination” that can exist even when propaganda (Ross-style) is absent. This lets you preserve Ross’s conceptual discipline while still capturing McKenna’s deeper political worry. Option 4: Challenge Ross Argue that propaganda need not require epistemic defectiveness: it can be characterised by politically strategic persuasion under conditions of asymmetrical power that bypasses democratic agency, even if truth-tracking. That’s a revisionary definition, but it would neatly match your initial dilemma’s “if it’s above board it’s fine” pressure. If you keep Ross, the most promising pro-propaganda strategy for both A and B is: **show that “scale + concentrated control” almost inevitably creates epistemic defect at the system level**, because it predictably distorts what audiences take to be the balance, importance, and availability of reasons — and that counts as misleadingness in Ross’s broad sense. That gives you a principled bridge from McKenna’s “marketplace domination” to propaganda without requiring you to say the individual arguments are bad. If you go that route, the next step is to write down (in your own terms) a very explicit claim like: “A communicative act includes its distributional architecture; therefore, systematic crowding-out and salience engineering can render an otherwise truth-only campaign epistemically defective.” That’s the hinge you’ll be turning. That seemed really interesting, but again, so compressed as to be pretty much unreadable. Um my idea for the paper I'm trying to write would begin with something like the following structure. It would introduce McKenna's arguments, okay, and then it would introduce what he describes as his best case and the problems um he thinks come from it. That would be section one of the paper. Section two would be applying the revised Tuttle Ross from the 2025 paper from the defectiveness paper to this idea. Somewhere in this section quite early on I think it would talk about well if the best case, best case really is perfect in one way we could interpret McKenna, it would lead to it would it ceases to seem like propaganda. Okay? Put that idea out there. Okay? Then it would look at the cases where we might think McKenna's best case is problematic or sort of talks about problematic circumstances, but those circumstances are not a case of propaganda. Then what I would like to try and do, but maybe I'm not going to be able to, is to argue that these are propaganda after all. You see what I'm getting at here. You do see it, and the structure you’re sketching is a good one. The thing that will make the paper *work* (and not collapse into hand-waving) is being very explicit about what, exactly, counts as the “communicative act” whose epistemic status we’re evaluating. The Bonard/Contesi/Marques revision of Ross is built to let you say: “even if the locutionary content is true, the act can still be epistemically defective, because defectiveness can live in implicatures, presuppositions, omissions, and in illocutionary/perlocutionary design.” Let me rewrite your plan in a way that keeps the dialectic clean and gives you several “routes” for turning the apparently non-propagandistic cases into propaganda. ## 1\. Section 1: McKenna, his “best case”, and why he still worries Here you want to do two things. First, present McKenna’s central move: he tries to *bracket* the easiest propaganda-style worry (misinformation) and focus on a “better/best case”: persuasive AI used to generate genuine arguments and reasons, aimed at rational persuasion, not deception. Second, explain why he thinks that “best case” can still be politically troubling: because even if the arguments are good, the technology enables (i) scale and (ii) concentration of capacity in a few actors, which can “dominate” the marketplace of arguments and steer what becomes salient. You’ll want to be very clear that McKenna’s complaint is *structural* (who controls production/distribution at scale), not simply “this or that argument is bad”. At the end of Section 1, you can put the dilemma on the table in plain language: If McKenna’s best case is genuinely epistemically above board in every relevant sense, then it’s hard to see why it’s propaganda (and maybe hard to see why it’s wrong at all). If it isn’t above board, then it starts to look like propaganda. That’s your hinge. ## 2\. Section 2: The revised Ross account (as revised in the “defectiveness” paper) Here you should quote or closely paraphrase the revised conditions, because they give you the exact levers you need. On their revision, a communicative act counts as propaganda only if it satisfies four conditions: (1) it conveys an epistemically defective message (including what it implies/implicates), (2) it is produced by or on behalf of an institution/cause with intent to persuade for direct political ends, (3) it targets a socially significant group, and (4) the producer is aware (or should have been aware) of the defectiveness. Two points from their paper are especially helpful for your project: 1. “Epistemically defective” is intentionally broad. A message can be defective not only by being false, but by misleading through omission or by generating false pragmatic inferences (they use a Gricean Quantity example: cherry-picking true war evidence can still be misleading because the audience infers completeness). 2. You don’t have to locate defectiveness in the literal content. They explicitly say “message” should be understood broadly, because illocutionary/perlocutionary intents can make a speech act epistemically problematic even when the locutionary content is true and evidence-supported. Also, you’ll probably want one short paragraph noting that this is the Philosophical Quarterly paper “The defectiveness of propaganda” (advance access May 2024), in case you need to reconcile the user-facing “2025” label you’ve been using with the publication metadata in the PDF. ## 3\. Early in Section 2: state the “easy verdict” and why it matters You can say something like this (conceptually, not as polished prose): If McKenna’s best case is genuinely non-defective in Ross’s broad sense—no misleading implicatures, no omissions that predictably generate false inferences, no presuppositions that distort, no perlocutionary engineering that detaches belief from reasons—then, on the revised Ross account, it will not be propaganda. That is not a concession to McKenna; it is just what follows from making epistemic defectiveness necessary for propaganda. This is the point at which you can cleanly distinguish two questions: 1. Is McKenna’s best case propaganda on this account? 2. Even if it isn’t, is it still politically troubling? Your project, as you described it, is: identify cases where McKenna thinks it’s troubling, those cases *seem* non-propagandistic, and then try to argue they are propaganda after all. ## 4\. The crucial move you’ll need for Section 3: widen the “unit” of evaluation Here is where many discussions go wrong, so it’s worth being explicit. If you treat “the communicative act” as just “the text of each argument the AI outputs,” then the non-propaganda verdict is hard to shift: lots of those texts could be epistemically clean. But Bonard/Contesi/Marques give you a principled way to treat the communicative act as larger than the raw locutionary content: the act includes what is implicated, what is omitted, and what is done illocutionarily/perlocutionarily. So the strategy for “these are propaganda after all” should look like this: Even if each individual argument is sound, the system-level communicative act—mass generation, strategic distribution, salience control—conveys an epistemically defective message in the broad sense, because it predictably leads audiences to form false or unwarranted beliefs about (for example) the completeness, representativeness, or decisiveness of the reasons in play. That is exactly the shape of their Gricean Quantity example: true pieces of information can be presented in ways that make the audience infer something false (“this is all the relevant information”), and that is epistemic defectiveness even without falsehood. ## 5\. Now, the two “non-propaganda” cases—and how to push them into propaganda You mentioned two kinds of cases: (i) domination/legitimacy worries, and (ii) procedural unfairness/flooding. Let me give you a readable “for/against” map for each, but in prose rather than compressed slogans. ### Case A: “Benevolent domination” (agenda-setting and attention steering, but no lies) Why it looks non-propagandistic: If the institution controlling the persuasive infrastructure is genuinely transparent about what it’s doing, does not misrepresent the balance of considerations, and does not cause audiences to draw false conclusions about what the total evidence is, then you may have domination in a political sense without epistemic defect in Ross’s sense. On the revised account, that would block propaganda at condition (1): no epistemic defect, no propaganda. How to argue it is propaganda after all: You argue that in real communicative practice, controlling salience and agenda *inevitably* communicates more than the literal content of the arguments. People take prominence and repetition as evidence of importance, consensus, and evidential weight. So a system that centrally steers attention predictably generates pragmatic inferences like “these are the main considerations” or “this is where the balance of reasons lies.” If that inferred message is false or unwarranted, then the communicative act is epistemically defective in the very sense the revised account foregrounds (implicature/connection to other beliefs) even if every displayed argument is sound. You then check conditions (2)–(4): it’s on behalf of an institution/cause, it aims at direct political ends, it targets a socially significant group, and the institution should be aware of the predictable misleading inferences produced by salience engineering. So the “conversion” move is: domination isn’t automatically propaganda; domination becomes propaganda when it functions as a way of making the audience infer a misleading picture of the deliberative landscape. ### Case B: Flooding as a “resource attack” (exhaustion and asymmetry) Why it looks non-propagandistic: If flooding merely exhausts and outpaces citizens but does not generate false beliefs—if people remain perfectly aware that they’re being bombarded, do not infer anything unwarranted about the balance of reasons, and are not deprived of access to counter-considerations—then the wrongness is procedural/democratic rather than epistemic. Again, Ross blocks propaganda at condition (1). How to argue it is propaganda after all: Here you have two powerful routes that fit the revised account very naturally. First route: flooding produces misleading Quantity effects. The more you saturate the environment with one side’s considerations, the more audiences infer (often rationally, given human heuristics) that these must be the main reasons, the decisive reasons, or the socially accepted reasons. If that inference is unwarranted, you get epistemic defectiveness via implicature, not via falsehood—again exactly the structure of their Quantity example. Second route: flooding creates omission by crowding out. Even if you never explicitly conceal counterevidence, you can make it practically inaccessible by overwhelming the channel. If this predictably leaves the audience with a systematically incomplete evidential base, that is misleading omission in the broad sense they endorse (again: true bits can mislead by omission). If you can establish either route, then the flooding campaign satisfies condition (1), and the institutional/political-intent conditions can do the rest. ## 6\. One honest possibility: you might not be able to do the “propaganda after all” step in the fully idealised best case This is worth saying explicitly in the paper, because it shows you’ve understood the conceptual terrain. If you *really* idealise McKenna’s best case so that it is non-defective even in the broad Ross sense—no misleading pragmatic effects, no skewed availability, no crowding-out, no false inferences about balance, and transparent governance—then the revised Ross account will not let you call it propaganda. That’s a feature, not a bug: Bonard/Contesi/Marques explicitly worry about overgenerating propaganda and insist on “direct political ends” as a restriction, using examples like teaching Newtonian physics as a case that is not propaganda partly because it lacks those direct political ends. If that happens, you still have a productive conclusion available: McKenna has identified a political pathology (concentrated control over persuasive infrastructure) that is not well captured by “propaganda” on an epistemic-merit account. That would be a philosophically interesting result, not a failure. But if your goal is to argue “these are propaganda after all,” the most promising line is not to deny Ross’s framework. It’s to argue that McKenna’s scale-and-control scenario cannot, in practice (and maybe not even in principle given how pragmatic inference works), avoid producing epistemic defectiveness at the campaign level. The paper you attached gives you exactly the conceptual tools for that move: defectiveness beyond falsity, and defectiveness that lives in implicature, omission, and illocutionary/perlocutionary design. If you want, I can take the next step and draft a clean “end of Section 2 → start of Section 3” transition paragraph that states the dilemma and then states, in one controlled thesis sentence, the exact mechanism by which you’ll argue the apparently non-propagandistic cases become propagandistic under the revised Ross conditions. 1\. Regarding that was good, thank you. Um first a quick comment on your section one. Uh I kind of like some of it, but I envisaged much more of a proper description of McKenna's ideas to begin with. Okay, so yeah, properly leading up his arguments. Also, he doesn't uh and you say he mentions propaganda, but I also get the impression he does not think his best case is propaganda. Is that correct? If so, that should be the sort of with cumulation of section one. 2. Regarding your point six, I'm still a little bit confused. Um, the thing is though, if you really idealize McKenna's best case, it seems non-problematic. Okay, it's not propaganda, I'm perfectly happy to say that, but it also, if it's really the best case, best case, then it's actually just a benign distributor of true and rational and reasonable information. No? Or am I missing something? Genuine question. 3. I would like us to start planning this paper on the canvas. 1. Yes, that reading is right. He *mentions* propaganda (and even discusses “AI-generated propaganda” in the overview), but his “best case” is explicitly designed to *not* be propaganda in the ordinary sense. He brackets misinformation/deception and focuses on AI used to offer “genuine arguments” and reasons. So, as a matter of how he frames the dialectic, the culmination of Section 1 can absolutely be: *McKenna is worried even when we are not dealing with propaganda.* 2. If you really idealise the “best case” into something like a perfectly transparent, pluralistic, non-misleading, non-exploitative distributor of true and sound arguments, then yes: it starts to look benign. That’s not you missing something; that’s the pressure you’re correctly putting on him (and it’s a nice way to force conceptual clarity). The main escape routes for McKenna are basically these, in plain terms. First, his “best case” is “best” only relative to one contrast class (not misinformation), not relative to a fully ideal democratic epistemic environment. Second, even if every *individual* argument is sound, the *campaign-level* effects of scale and concentration can still make the overall communicative act misleading (e.g., by making audiences infer a false picture of what the balance of reasons looks like). If you stipulate *that* away too, then you’ve stipulated away the mechanism by which his worry bites. 1. Done - I’ve set up a canvas with a sectioned plan that matches your intended structure, including where to make the “McKenna’s best case isn’t propaganda” point land, and where the dilemma and “propaganda after all” move should go. Fair. Here’s the same point, but unpacked so it’s actually readable, with the “moving parts” laid out slowly and with examples. ## 1\. What I meant by “escape routes” You were pressing a very sharp dilemma: If McKenna’s “best case” is *really* as clean as it can be—true claims, sound arguments, no misleadingness—then why is it troubling at all? It looks like a benign distributor of reasons. If, on the other hand, it becomes troubling only when it starts misleading people (through omission, implicature, skewed exposure, etc.), then it starts looking like propaganda after all. McKenna wants to avoid that dilemma by saying: “Even if we set aside the easy bad stuff (misinformation, deception), there’s still a serious problem.” So the question is: **how can he keep a serious problem while keeping the “best case” clean?** That’s what I was calling his “escape routes.” ## 2\. Escape route one: “best case” is only best relative to one contrast class When McKenna says he’s looking at a “best case” (or “better case”), he means something like: > “Let’s not make this trivial. Let’s not focus on obviously dodgy uses of AI—misinformation, deception, manipulation. Let’s imagine the persuader is offering genuine arguments and reasons.” That’s a *best case relative to a specific opposition*: **not-misinformation**, **not-deception**, **not-straight-up lying propaganda**. But notice what that does *not* automatically give you. It does **not** give you the “best possible democratic epistemic environment,” i.e., something like: - pluralistic exposure to competing arguments, - roughly fair contestability (others can reply), - symmetric ability to participate, - transparency about who is speaking and why, - no single actor controlling what becomes salient, - no structurally enforced agenda-setting. Those are *political-democratic* ideals about how public reasoning should be structured. McKenna’s “best case” doesn’t promise those. It’s “best case” in a narrower sense: the content is meant to be rational and not deceptive. So McKenna can say, without contradiction: > “I’m assuming the AI is producing genuine arguments (so we’re not in the misinformation/lying world), but I’m **not** assuming the broader democratic conditions that would make the persuasion environment fair.” In other words, his “best case” is “best” in *one* dimension (truthfulness/rationality), but still allowed to be “worst-ish” in another dimension (concentrated power over attention and dissemination). That is one way his worry can survive: the problem is not that the arguments are bad; it’s that **who gets to steer the argumentative environment** is politically illegitimate or democratically corrosive. Crucially: **if that is his route, then the problem may not be propaganda on Ross’s account.** It’s a domination/legitimacy problem rather than an epistemic defect problem. ## 3\. Escape route two: even “sound arguments” can mislead at the campaign level The second route is different. It says: “Even if each argument is individually sound, the overall campaign can still be epistemically defective.” This depends on a distinction that matters a lot in your paper: - **Pointwise cleanliness**: each individual message is true / each argument is valid. - **System-level cleanliness**: the overall communicative act (including selection, repetition, prominence, distribution, and what audiences are licensed to infer from those) is non-misleading. The revised Ross-style account you’re using makes room for system-level defectiveness because “epistemically defective” can involve misleading implication/omission/Quantity-style effects. So you can say: > “Even if the building blocks are true, the way they’re deployed can convey a false picture of the space of reasons.” Here’s a concrete example (not involving any false statements): Imagine there are ten strong considerations on a topic: five push towards policy A, five push towards policy B. A powerful actor floods the public sphere with the five pro-A considerations, presented clearly and honestly. They never say “there are no good arguments for B.” They never lie. Each pro-A argument is sound. But lots of ordinary readers will still form beliefs like: - “It looks like the case for A is overwhelming.” - “If there were serious objections, surely I’d be seeing them.” - “This is what the reasonable view is.” Those beliefs are not the result of a false sentence; they are the result of *how humans infer completeness/representativeness from saturation*. That’s exactly the sort of thing the “defectiveness” paper treats as epistemically relevant (misleading by omission/Quantity). So on this second route, McKenna can say: > “The danger isn’t that any one argument is deceptive; it’s that large-scale, concentrated dissemination predictably induces *misleading inferences about the balance and availability of arguments*.” If you accept that, then his “best case” is not actually as epistemically clean as it first sounds—not because the arguments are unsound, but because the campaign communicates (via implicature and omission-by-crowding-out) a defective map of the landscape. And **that** is precisely the move that lets you argue: “These are propaganda after all” *on the revised Ross account.* ## 4\. Why I said “if you stipulate that away too…” Because you can keep idealising and idealising until nothing remains. If you stipulate: - all arguments are sound and true, **and** - the dissemination does not create misleading inferences about balance/representativeness, **and** - the system ensures exposure to counterarguments, **and** - the governance is transparent and contestable, **and** - no one actor has unilateral agenda-setting power, then, yes: you’ve basically built an epistemic utopia machine. Under those conditions, it becomes hard to see what McKenna’s “scale + concentration” worry even *could* be. You’ve removed both: - the epistemic mechanism (misleadingness at the campaign level), and - the political mechanism (illegitimate concentrated control). So the point I was trying (badly) to make is: McKenna needs *something* non-ideal to remain in the picture for his worry to bite. Either: 1. concentrated power remains (so it’s politically troubling even if not propaganda), or 2. campaign-level epistemic defect remains (so it becomes propaganda on Ross), or 3. both remain (politically troubling *and* propagandistic). ## 5\. How this helps your paper’s structure This gives you a clean fork you can build Section 3 around: 1. **Concessive fork**: The fully idealised best case is benign and non-propagandistic. McKenna’s worry is therefore inherently non-ideal and depends on realistic constraints (incentives, concentration, distribution). 2. **Propaganda fork**: Even without falsehood, the best case is still propagandistic under revised Ross because the campaign architecture predictably misleads about the balance and availability of reasons (omission/implicature/Quantity effects). 3. **Domination fork**: Even if you manage to avoid campaign-level misleadingness, there can still be a political wrong (domination/agenda capture). But then the right label is not “propaganda” on Ross; it’s something like epistemic domination or democratic illegitimacy. That’s the landscape. If you want, I can write you a full draft version of the paragraph that comes right after you introduce Ross in Section 2—where you stage the dilemma in a way that makes this fork feel inevitable rather than ad hoc. You’re zooming in on exactly the right pressure point. The short version is: **Domination/legitimacy can cause epistemic defectiveness, but it doesn’t *logically guarantee* it.** Ross-style propaganda requires epistemic defectiveness. So the question becomes: *does the domination you’re talking about inevitably involve (or inevitably produce) epistemic defectiveness in Ross’s sense?* If yes, you can call it propaganda. If no, then you can’t—at least not on Ross. Let me unpack that carefully. ## 1\. Why “domination/legitimacy” and “epistemic defect” are not the same thing Think of “domination/legitimacy” as a *political* property: - Who has power over the infrastructure? - Is that power contestable? - Is the arrangement democratically acceptable? - Do citizens have an equal standing in shaping the public sphere? Those can go badly even when no one is lying or misleading. Think of “epistemic defectiveness” as an *epistemic* property: - Does the communicative act convey something epistemically wrong (false, misleading, unwarranted)? - Does it lead the audience to form beliefs on an epistemically bad basis (because of omission, implicature, etc.)? These two can overlap, but they’re conceptually distinct. A simple analogy: a dictatorship can publish accurate statistics. It’s politically illegitimate even if the statistics are true. So political illegitimacy doesn’t automatically entail epistemic defect. That’s why I said “may not”: it depends on whether the domination operates *through* epistemic defectiveness or merely *through* control. ## 2\. Why the “genuine arguments and reasons” scenario can still be propaganda on Ross Yes—this is crucial—and you’re right to insist on it. On the revised Ross account (as interpreted in the “defectiveness” paper), **propaganda does not require that the literal content be false**. It requires that the *message conveyed by the communicative act* be epistemically defective, and that can happen even if the persuader offers genuine arguments and reasons. The easiest way to see this is: you can offer genuine reasons and still be misleading by **omission**, **implicature**, or **framing**. True premises, valid inferences, real evidence—yet the total communicative act pushes a false or unwarranted picture. So: “rational persuasion” is not automatically non-propaganda on Ross. It depends on whether the rational persuasion is epistemically defective at the level Ross cares about. ## 3\. So what’s the gap you’re feeling? You’re thinking: > If a small group dominates the infrastructure, won’t they inevitably distort what people know and believe? Doesn’t that automatically produce epistemic defect? That’s *often* true in practice. But Ross is a definition with necessary conditions, and definitions are picky. There are domination scenarios where the powerful actor could, in principle, avoid creating epistemic defectiveness, even if they remain politically illegitimate. Here are two contrasting scenarios to make the distinction vivid. ### Scenario A: Domination without epistemic defect (conceptually possible) A single institution controls a public deliberation platform. That is politically troubling (concentrated power), but it behaves like an ideal public broadcaster: - It transparently discloses its control and goals. - It ensures robust exposure to competing arguments. - It does not crowd out alternatives. - It does not exploit salience effects to create a misleading impression of balance. - It does not selectively amplify one side while pretending neutrality. - It provides strong reasons, but also systematically presents counter-reasons. In that world, you could still complain: “This is illegitimate—no one should have this much power.” But it’s not obvious you can also say: “This is epistemically defective.” It might be epistemically excellent. So: political domination doesn’t automatically imply epistemic defect. ### Scenario B: Domination that works via epistemic defect (very plausible) Same institutional control, but now: - The platform massively amplifies one side’s arguments. - It does so in ways that make audiences infer “this is the decisive balance of reasons.” - It crowds out counterarguments so that people rarely encounter them. - It designs the environment so that prominence looks like consensus or evidential weight. Here, the domination is operating by shaping what people *take the evidential landscape to be*. That is epistemic defect in the broad Ross sense (misleading by omission/Quantity implicatures, etc.). In that world, it *is* propaganda (assuming the political-ends and awareness conditions are met). So: domination often produces propaganda, but not by definition—only if it creates the relevant epistemic defectiveness. ## 4\. The key sentence that resolves your confusion Here it is in plain English: **The domination/legitimacy problem gives you a reason to worry even if no one is misleading anyone. Propaganda requires misleadingness (broadly construed). So unless you can show that the domination inevitably misleads, you can’t automatically classify domination as propaganda on Ross.** That’s the missing link. ## 5\. How to use this in your paper You can turn this into a very clean argumentative task: 1. Grant: McKenna’s “best case” tries to avoid *obvious* propaganda (misinformation/deception). 2. Observe: on Ross, “genuine arguments and reasons” can still be propaganda if the campaign conveys epistemically defective messages through omission/implicature/framing. 3. Therefore, the crucial question is: **does McKenna’s scale-and-control scenario inevitably (or at least typically) produce epistemic defectiveness at the campaign level, even if individual arguments are sound?** If you can argue “yes”, you get what you want: McKenna’s best case is propaganda after all (on Ross), because domination essentially functions by producing misleading pictures of the balance/availability of reasons. If you argue “no” (or if you think the “fully idealised” case shows the “no” is coherent), then you might have to conclude: McKenna’s worry is real but is not fully captured by Ross’s propaganda category; it’s better described as domination/agenda capture/epistemic power. ## 6\. One way to make your preferred “yes” argument bite If your intuition is that domination *will* create epistemic defect, the most defensible way to phrase it is not “domination entails defectiveness” but: **Given human pragmatic inference (we infer importance/consensus from prominence) and given the economics of attention, any actor who dominates the infrastructure and uses it for direct political ends will predictably produce misleading inferences about the balance and salience of reasons unless they actively counteract those effects. Since active counteraction undermines the strategic value of domination, the realistic domination scenario is almost inevitably epistemically defective.** That’s a strong, non-hand-wavy bridge. So what you weren’t understanding wasn’t a mistake on your part; it’s that I was sliding between (i) “often in practice” and (ii) “must by definition.” Your project lives exactly in closing that gap: showing that in McKenna’s setup, “often” is close enough to “structurally inevitable” to justify the propaganda classification. "If a small group dominates the infrastructure, won’t they inevitably distort what people know and believe? Doesn’t that automatically produce epistemic defect?" I am not saying this. If a small group dominated the ninformation info structure, they might do it entirely benignly and so give their citizerns a perfectly correct information, well-reasoned, not omitting etc. arguments. this, i think we agree would be a place in which a) no propaganda is being distributed, b) everything is great. ethically above board. Ok, but then imagine that this small group who are in charge of everything are slightly less than entirely benign, and therefore, while not out right directly lying to the populace, the rational arguments they are putting out there are epistemically defective because of omission. therefore, propaganda, boo What possibilities are there, at the level of abstraction, other than these two? Good — you’ve made the space crisp. You’re saying: - Case 1: **benign + epistemically clean** → not propaganda; no real ethical/political problem. - Case 2: **not benign + epistemically defective via omission (etc.)** → propaganda; bad. You’re asking: at a high level of abstraction, what other possibilities are there besides those two? There are several, and the key is that you’re currently tying “problematic” to “epistemically defective”. That’s one axis. But you can vary (i) *the epistemic status*, (ii) *the political/ethical status*, and (iii) *how defectiveness enters (message vs uptake vs environment)*. Once you separate those, you get more than two boxes. Here are the main additional possibilities. ## 1\. Not propaganda, but still not “everything is great” This is the big one. It’s the gap between “no epistemic defect” and “ethically above board”. ### 1A. Epistemically clean, but politically illegitimate (domination without deception) The rulers (or platform owners) distribute only true, well-reasoned, non-misleading arguments, with full balance. Still, many political theories say this can be wrong because it’s a form of domination: citizens lack equal standing, contestability, and control over the terms of public reason. On Ross, this is **not propaganda** (no epistemic defect), but it’s **not automatically “everything is great”** if you think legitimacy matters independently of truth-tracking. If you personally don’t think legitimacy matters when everything is epistemically perfect, then you won’t count this as a problem. But many will — and this is a major “third option” in the abstract space. ### 1B. Epistemically clean, but autonomy/agency is bypassed (without misleading content) Even with true reasons, persuasion can be designed to bypass reflective agency: timing, targeting, emotional pressure, cognitive overload, “conversion optimisation,” etc. You can stipulate no misleadingness in content and still have a worry that the method treats citizens as objects to be managed rather than as participants in deliberation. Whether Ross counts this as “epistemic defect” depends on how broadly “defect” is understood (some readings might say it degrades epistemic agency, others might treat it as a separate ethical wrong). Either way, it’s a **distinct possibility** from your two. ## 2\. Propaganda can occur without the rulers being “not benign” in the ordinary sense This sounds weird at first, but it’s important because it breaks the neat “benign vs not benign” mapping. ### 2A. Well-intentioned but epistemically defective (paternalistic propaganda) Actors sincerely aim at the public good, and they spread what they take to be true — but they still deploy epistemically defective messaging (omission, framing, misleading implicatures) because they think it’s justified to secure compliance (classic public health communication temptations). On Ross, this can still be propaganda (defectiveness + political ends + awareness/culpability), even though the agents might be “benign” in motive. So the moral axis (“benign”) and the propaganda axis (“defective”) don’t line up perfectly. ### 2B. Epistemically defective through negligence rather than malice No one intends to mislead, but the institution is careless, incompetent, or reckless in ways it should foresee. The revised account includes “aware or should have been aware.” So you can get propaganda-ish defectiveness without “slightly less than benign” intentions — just culpable epistemic sloppiness. ## 3\. Epistemic defectiveness is present, but it doesn’t look like omission or lying Your second case is “defective because of omission”. But Ross-style defectiveness can come in other forms. ### 3A. True content, misleading pragmatics (implicature/presupposition) Everything explicitly said is true; nothing essential is omitted; yet the way it’s said (or what it presupposes) predictably leads people to infer something false. This is different from omission, and it matters because McKenna’s “marketplace” story is largely about these pragmatic and distributional effects. ### 3B. Misleading via distorted salience rather than missing facts No one hides counterarguments; they exist. But the system makes some considerations hyper-salient and others practically invisible, producing systematically unwarranted beliefs about what matters or what is decisive. This is not “omission” in the simple sense (“they didn’t mention X”), but it can still be “omission” in the broader “the audience is led to an incomplete evidential picture” sense. ### 3C. Epistemically defective connections rather than defective propositions Even if each claim is true, the campaign can encourage unwarranted inferential links: “supporting policy A = being rational,” “disagreeing = being morally suspect,” “only cranks doubt this,” etc. This corrupts the audience’s epistemic norms and connections among beliefs. That’s still defectiveness, but not omission. ## 4\. Neither “propaganda” nor “great”: the “epistemically fine but epistemically costly” world This is subtle but useful. ### 4A. Everything is true and balanced, but the environment is unusable You can make the system epistemically impeccable but *practically un-navigable*: too much information, too many arguments, too many caveats, too much noise. People can’t process it, so they fall back on heuristics (tribal cues, source loyalty, etc.). Now: is that propaganda? Not necessarily. Is it “great”? Not necessarily. It’s an epistemic tragedy of the commons: no one misleads, but the environment doesn’t support good belief formation. This is a third possibility because it breaks your link “epistemically clean ⇒ everything great.” It can be epistemically clean in principle, yet fail to function well for bounded agents. (If you insist that “epistemically clean” includes being presented in a way that is realistically usable by citizens, then you’ll fold this back into defectiveness. But many won’t.) ## 5\. The “multiple audiences” possibility: clean for some, defective for others At abstraction level: a system can be epistemically clean *for the well-resourced* and defective *for the less-resourced* — not because different claims are made, but because of differential accessibility, literacy, time, and interpretive support. Then you can have: - no outright omission or lying, - but predictable misunderstanding in some populations, - and possibly culpability if the system is designed without regard for that. Whether this counts as propaganda depends on whether you treat predictable, avoidable misunderstanding as a kind of defectiveness. ## 6\. A clean way to display the space (without turning it into a messy matrix) You started with two boxes. The main additional boxes come from separating three questions: 1. **Is the communicative act epistemically defective (in Ross’s broad sense)?** 2. **Is it politically/ethically legitimate (procedurally, democratically, relationally)?** 3. **Is it designed to optimise persuasion in ways that bypass agency (even without misleading content)?** Your two cases cover only: - (1) no, (2) yes, (3) no → “all good” - (1) yes (via omission), (2) no, (3) maybe → “propaganda, boo” Other stable possibilities include: - (1) no, (2) no → “not propaganda but politically illegitimate” - (1) no, (3) yes → “not propaganda but autonomy-bypassing manipulation” - (1) yes, (2) yes-ish → “well-intentioned propaganda / paternalistic propaganda” - (1) no, but still bad outcomes because of overload/usability → “not propaganda, not great” - (1) mixed across groups → “propaganda-like for some subpopulations” ## 7\. Why this matters for your paper Because if the only “bad” cases are epistemically defective ones, then your paper becomes almost trivial: “Ross says propaganda requires defect; therefore the bad cases are propaganda.” The interesting philosophical work is in showing that McKenna’s “best case” worry either: - actually implies epistemic defect at the campaign/system level (so it *is* propaganda on Ross), **or** - identifies a *distinct* political wrong not captured by Ross’s concept of propaganda (domination/legitimacy/agency), which is a valuable result in its own right. If you tell me which of these you want the paper to conclude (even tentatively), I can reshape the plan on the canvas so the dialectic is engineered to land there. Below are **concrete, “story-shaped” examples** where something is clearly *problematic*, but the problem is **not** (or at least not *primarily*) a matter of epistemic defectiveness. In each case I’ll make the “not epistemic” part explicit: assume the arguments are true, well-reasoned, balanced, and not misleading by omission/implicature/presupposition. I’m focusing on the options that really can be non-defective; some of the possibilities we discussed earlier (misleading implicatures, omission, salience distortion) are *by definition* epistemic-defect routes, so they won’t fit what you’re asking for here. ## 1\. Political illegitimacy even with perfect epistemic performance **Example: “CivitasGPT” as the mandatory public sphere interface.** A state (or a consortium of platforms) mandates that all civic information and political debate funnels through one official AI: it’s the gateway for reading proposed legislation, seeing arguments for and against, and submitting comments. The system is genuinely exemplary: it presents the strongest arguments on all sides, discloses uncertainty, cites sources, highlights counterarguments, and flags value trade-offs. Independent auditors verify it isn’t skewing anything. **What’s problematic?** The political problem is that **citizens now live under a single, centralised chokepoint for public reason**. Even if today’s operators are saints, the arrangement is a standing form of domination: the system can be changed later, quietly or suddenly; access can be revoked; priorities can be altered; dissent can be throttled; the interface can be redesigned to privilege some forms of participation over others. In a republican register, citizens are vulnerable to arbitrary power even if it is not currently exercised badly. In a democratic register, the structure is illegitimate because it removes contestability: there is no robust plurality of independent venues. **Why this is not an epistemic-defect complaint.** The complaint does *not* have to be “the system misleads people.” The complaint is: **no one should have that kind of structural power over the conditions of public deliberation**, even if the content is epistemically pristine. The wrong is about legitimacy, control, and vulnerability to arbitrary interference, not about false belief or misleadingness. A sharper variant makes this even clearer: the system is perfect epistemically but **compulsory** (you must use it to vote, file taxes, access healthcare). That’s coercion/legitimacy, not epistemic defect. ## 2\. Surveillance and chilling effects without misinformation **Example: “Ask the Policy Bot” with perfect answers, but total logging.** A government and a platform jointly offer an AI adviser for citizens: you can ask about union rights, immigration options, protest law, political parties’ programmes, etc. The bot is genuinely reliable, balanced, and non-manipulative. But all queries are logged to “improve safety” and “prevent extremism”, and (crucially) citizens know this. **What’s problematic?** People self-censor. They stop asking about sensitive topics (“What are my rights if I’m stopped at a protest?” “How do I report corruption?” “What are arguments against the governing coalition’s policy?”). The public sphere becomes quieter and more conformist, not because anyone is misled but because **the cost of curiosity rises**. **Why this is not an epistemic-defect complaint.** The bot can be impeccably truthful and balanced. The harm is that **surveillance alters behaviour** and undermines democratic freedom. This is a rights/agency problem: citizens lose the practical ability to explore and contest power safely. No falsehood or misleading omission is required to make the case. (If you want an even cleaner version: the logs are never misused and never even read. The chilling effect can still exist purely from the *structure* of monitoring.) ## 3\. Bypassing deliberative agency while saying only true things **Example: “Conversion-optimised nudging” with strictly accurate content.** A political campaign deploys an AI that sends voters personalised messages. Every message is factually true, source-cited, and includes a link to the best opposing argument. There is no omission, no misleading framing, no fake consensus cues. The only “optimisation” is in timing and delivery: the system pushes messages when you are predictably tired, stressed, or emotionally primed (after a long workday, late at night, during a family-health scare), because that’s when people are most likely to click “agree” or donate. **What’s problematic?** Even if the content is epistemically fine, the strategy treats citizens less as deliberators and more as **targets to be captured**. It exploits predictable weaknesses in attention and self-control. The worry is not “you ended up with false beliefs”; the worry is that your political agency is being *managed* —that persuasion is being engineered to succeed by manipulating the conditions under which you evaluate reasons. **Why this is not an epistemic-defect complaint.** You can stipulate that the beliefs formed are true and well-supported. The wrong can still be that the method is a kind of agential disrespect: it aims to secure assent by exploiting vulnerabilities rather than fostering reflective endorsement. That’s an autonomy/respect story rather than an epistemic defect story. A less moralised variant: the campaign is so effective at exploiting attention that **it becomes the dominant determinant of what citizens end up caring about** —not by lying, but by engineering the rhythm of their lives. Again, that’s not reducible to defectiveness. ## 4\. Procedural unfairness as a participation-cost problem (without misleading anyone) **Example: “Argument DDoS” that doesn’t deceive.** A wealthy actor deploys an AI to generate and submit millions of public comments on proposed regulation. Each comment is coherent, factually correct, and explicitly labelled as AI-generated. They do not pretend to be a grassroots movement; they do not mislead about their origin; they do not omit relevant counterarguments within the individual comments. The content is not propagandistic in Ross’s epistemic sense. **What’s problematic?** The democratic process is gummed up. Regulators cannot process genuine citizen input because the channel is flooded. Civil society groups must spend money filtering noise. Journalists cannot track real public response. Ordinary citizens are effectively pushed out of participation because **the cost of being heard skyrockets**. **Why this is not an epistemic-defect complaint.** No one is being misled. The wrong is procedural: the tactic weaponises volume to degrade the functioning of a participatory system. It’s closer to a denial-of-service attack on deliberation than to propaganda. You can also run this example in media rather than regulation: buying all ad inventory and blasting true arguments continuously can be a fairness/participation problem even if nobody draws false inferences and everyone knows what’s happening. ## 5\. Epistemically excellent, but distributively unjust access **Example: “Perfect deliberation as a luxury service.”** A private firm builds the world’s best AI “argument concierge”: it is transparent, balanced, rigorous, and non-misleading. It helps citizens understand policy trade-offs far better than any existing source. But it costs $200/month, requires fast devices, and is integrated into elite workplaces and universities. The free version is still true and balanced, just slower and less interactive. **What’s problematic?** The epistemic infrastructure of citizenship becomes stratified. The wealthy have dramatically better tools for forming and defending political views, for drafting comments, for lobbying, and for navigating bureaucracy. Poorer citizens are not misinformed; they are **outcompeted**. Political equality is undermined by unequal cognitive resources. **Why this is not an epistemic-defect complaint.** The content can be non-defective for everyone who accesses it. The wrong is distributive and political: unequal access to deliberative capacity changes who effectively gets to steer policy. This is “epistemic injustice” in a structural sense, but not “epistemic defectiveness” in Ross’s sense (not deception/misleading message). ## 6\. Epistemically clean but psychologically unusable at scale **Example: “The Impeccable Firehose.”** Imagine a platform that, in the name of neutrality, provides *all* the strongest arguments for *all* sides of every issue, with perfect citations and caveats. It refuses to simplify, prioritise, or curate, because that would impose values. Nothing is misleading; nothing is omitted; nothing is framed manipulatively. **What’s problematic?** Citizens cannot use it. The cognitive burden is enormous. People either disengage (“I can’t deal with this”) or outsource judgement to shortcuts: tribal cues, influencer endorsements, or whatever is easiest. The net effect can be worse deliberation, not because the platform lies but because it creates conditions in which **bounded agents cannot form stable, reflective views**. **Why this is not an epistemic-defect complaint.** The platform is epistemically excellent in a truth-and-balance sense. The problem is a mismatch between epistemic ideals and human cognitive limits. The harm is practical: it undermines deliberation by making participation too demanding. That’s a design/governance failure, not a misleading-message failure. --- ### A quick meta-point tying all these together If you define “epistemically above board” so strongly that it includes not only truth/balance/non-misleadingness, but also **fair access, non-coercion, privacy protection, contestability, and respect for agency**, then many of the examples above will start looking “epistemically” problematic by *your* lights. But that would be because you’ve packed political and ethical constraints into what counts as “epistemically above board.” Ross-style propaganda doesn’t do that; it keeps the epistemic condition relatively specific (defective message, broadly construed) and leaves room for other kinds of wrong. If you want, the next useful step is to decide which of these non-defect wrongs you think McKenna *actually* has in mind (or should have in mind). That decision will determine whether your paper’s punchline should be “McKenna’s best case is propaganda after all” or “McKenna’s best case identifies a real political wrong that propaganda-theory doesn’t capture well.” Yes — **some** of those examples are very close to what McKenna has in mind (especially the “flooding / volume” one and the “small number of actors controlling the infrastructure / marketplace of arguments” one). **Others were deliberate extrapolations** that fit the general family of worries about power and public reason, but aren’t really McKenna’s target in this paper (e.g. surveillance/chilling effects, “mandatory interface”, etc.). Below I’ll go example-by-example, and for the “yes / close” ones I’ll back it up with **longer block quotes from McKenna** (from the text you pasted). I can’t give page numbers because the paste doesn’t include them. --- ## 1\. “Argument DDoS” / flooding the channel with (individually) decent arguments This is **very close** to McKenna’s central worry. In fact, it’s basically a concrete instantiation of his claim that the danger is **quantity + scalability + concentrated capacity**, rather than magical persuasiveness. Here’s the key passage from the Introduction where he states the “quantity not quality” thought very explicitly: > “I think that Floridi’s position is unnecessarily alarmist and not supported by the existing literature, or reasonable extrapolations from it. But I also think that there are genuine concerns about persuasive AI. As I will argue, the fundamental reason why we should be worried about AI persuasion is not that AI has, or soon will have, significantly more persuasive power than humans. The reason we should be worried is that any AI persuader or AI-generated persuasive content is (a) relatively easy to roll out en masse and at scale when (b) you have the resources (financial as well as technological) to create it in the first place. That is, we should be worried about AI persuasion for roughly the same reasons that many are worried that a tiny number of tech companies control large amounts of the information we see online, as well as the infrastructure required for us to share and access that information. The worry is not so much about the quality of AI-produced persuasive content as the quantity of it, not to mention who has the power to produce and disseminate it.” And in the “Sophistry Argument” section he makes the flooding mechanism even more concrete (armies of arguers-for-hire; “within minutes”): > “Setting these complications aside, the point is this: Whatever issues exist with existing marketplaces for arguments—and there may be many—persuasive AI technologies have the potential to greatly exacerbate them. This is for two different but related reasons. The first is that persuasive AI technologies have the potential to dominate the marketplace of arguments. For humans, contributing to the marketplace of arguments is time-consuming and labour-intensive—it involves coming up with new arguments, or finding ways of recycling old ones, and then looking for a way of getting those arguments disseminated more widely (e.g. publishing them somewhere where people will see them). Persuasive AI technologies provide a way of avoiding these time and labour constraints. The purveyors of persuasive AI technologies can generate huge numbers of ‘arguers for hire’ (AI agents) or huge amounts of ‘persuasive content’ within minutes, with any dilution in the quality of arguments often more than made up for by the massive increase in quantity.” That is basically the “Argument DDoS” example in philosophical prose. So: **yes, flooding is absolutely McKenna’s terrain.** --- ## 2\. “CivitasGPT” / a few actors controlling the infrastructure of persuasion This is also **very close**, with one caveat. McKenna’s paper is not primarily about *state-mandated* infrastructure (my example was intentionally dramatic); he’s mainly worried about **a small number of powerful actors (especially big tech firms) controlling the production and dissemination capacity**, and thereby controlling the “marketplace of arguments.” But the underlying structure — “few actors control the infrastructure that shapes public reasoning” — is *exactly* his. Here’s the passage where he builds “concentration of capacity” into the problem: > “The second reason is that not just anyone can create and run persuasive AI technologies. The resources—not just financial, but also in terms of infrastructure, skills, and knowledge—required to design and operate a LLM are massive, so much so that only a very small number of companies operate in this field, at least at the level relevant to the argument of this paper (e.g. Anthropic, OpenAI, DeepMind, Microsoft, Meta). This means that the barriers to entering the marketplace in this way are very high, so high that ordinary people cannot hope to meet them.” And then the culminating statement: **small number of actors can determine the balance of arguments represented**: > “Putting these two reasons together, the problem with persuasive AI is that a very small number of actors have the power to dominate the marketplace of arguments. Even if these actors have ‘good intentions’, and design persuasive AI technologies in ways that they think will promote what they view as true conclusions, this creates the potential for significant imbalances in the marketplace, with this small number of actors having the power to determine the balance of arguments that gets represented on any given issue. If they don’t always have good intentions, or are simply wrong about which arguments are cogent or which conclusions true, the problems get worse.” That is essentially the “infrastructure control” example, except that he cashes it out in terms of **AI firms and argumentative supply**, not “the state makes one bot mandatory.” So: **yes, infrastructure-control is McKenna’s terrain** — though my scenario was a stylised version. --- ## 3\. “Perfect deliberation but unequal access” / stratification and inequality of participation This is **moderately close**. McKenna does explicitly connect the marketplace distortion worry to **existing inequalities** and to the way markets tend to amplify the advantages of those with resources. That’s very much the same shape as the “luxury deliberation service” scenario (even if he doesn’t literally talk about subscription tiers). Here’s the passage where he draws that link: > “My main point in this section has been that the problem with persuasive AI is that it gives a small number of actors disproportionate power and influence over the pool of arguments available for public consumption and consideration. Let me finish by highlighting some further problems with persuasive AI that follow from this. It is a familiar point that, for all their transformative potential, markets often serve the material interests of some people better than others. Benefitting from—or even participating in—a market requires resources like time, information and skills. Some people have more of these resources than others, with the result that they gain far more from the market. Moreover, these differences in resources are not simply the result of random chance; they typically reflect existing inequalities within society (Herzog 2024). The concern, therefore, is that not only will persuasive AI distort the marketplace of arguments, but that it will do so in ways that reflect and reinforce existing inequalities and imbalances in power.” So: **yes-ish**. He’s not focused on “paywalled civic reasoning tools” specifically, but he *is* focused on how persuasive AI can amplify structural inequality in participation and influence. --- ## 4\. “Surveillance/chilling effects” and “mandatory interface” variants These are **not really McKenna’s focus in this paper**. They are plausible worries about AI and public reason, but he doesn’t develop them, and I can’t honestly back them with quotes from what you pasted. He does talk about “control over information” and about tech companies controlling what people see, but not in the surveillance/chilling-effects register. His emphasis is: - control of what information we have access to, and - persuasive AI narrowing the gap between controlling access and controlling attitudes. Here’s the relevant passage (this supports “information control”, not “surveillance”): > “One place where this may be particularly troubling is when it comes to control over information. Even if you are sceptical of the idea that there is a ‘right to be informed’ … it is clearly true that, in a democratic society, information matters. Having the power to decide which information people have access to (or lack access to) is not quite the same as having the power to decide what people think; you can dictate what people see and hear, but not what they think. Persuasive AI technologies threaten to close this gap. While I am sceptical that they will close it entirely … I am very much concerned about the idea that big tech companies will increasingly have the means not just to control what information we have access to, but to control the marketplace of arguments.” So: **he’s close to “control of access”** but **not** really doing surveillance/chilling effects or state-mandate scenarios. --- ## 5\. “Conversion-optimised timing / exploiting vulnerability while staying truthful” This is **only weakly connected** to McKenna as written here. He does discuss *tailoring* and the way personal information can make chatbots more persuasive (via the Salvi et al. study), but he does not develop the more specific “exploit vulnerability windows” or “bypass agency” story. The closest he gets is the empirical point that persuasion increases with personal data: > “However, in the second setup, where the chatbots were given basic demographic information … they found that the chatbots were significantly more persuasive. The researchers conclude that ‘not only are LLMs able to effectively exploit personal information to tailor their arguments, but they succeed in doing so far more effectively than humans’.” But that’s still about *effectiveness via tailoring*, not about *ethically objectionable timing/manipulation without defectiveness*. So I’d call this: **adjacent, not central**. --- ## Bottom line If you want examples that are *genuinely in McKenna’s line of fire*, the strongest matches are: 1. **Flooding / “armies of arguers-for-hire”** and the procedural consequences of that. 2. **Concentrated capacity in a small number of actors** who can shape the “marketplace of arguments.” 3. **Reinforcing existing inequalities** through asymmetric resources and access. The other examples I gave (surveillance/chilling; mandatory interface; conversion-timing manipulation) are not really what he’s doing in this paper, even if they are in the same galaxy of concerns about AI and democratic life. If you want, I can now do the next thing your paper needs: pull **one or two “anchor” passages** that make it unmistakable that McKenna does *not* regard his best case as propaganda (because he brackets deception/misinformation), and then use those anchors to craft the end of Section 1 so the reader arrives at your dilemma naturally. flooding = propaganda on the tuttle ross account, no? if not explain the example to me again whereby the zone is flooded but in such a way that does not lead to epistemic defectiveness. If it is flooded in a way that means the populace have all the facts available, and all the arguments available, and they are all presented in a fair way = not propaganda but also benign. if the flood does lead to some defect in epistemology, then it is propaganda. what am i missing here? i think a similar argument could be fun against oyur other examples too You’re not missing anything *in the core conditional you’ve stated*. If we grant the revised Ross-style account as the definition, then: - **Flooding + epistemic defectiveness** (misleading by omission/implicature/presupposition, etc.) ⇒ **propaganda** (assuming the other conditions: political ends, targeting, awareness). - **Flooding without epistemic defectiveness** ⇒ **not propaganda** (again, on Ross). So the question is: **is “flooding without epistemic defectiveness” a coherent possibility**, or does flooding *by its nature* produce epistemic defectiveness? That’s the real disagreement. You’re leaning towards: flooding *will* always make things epistemically defective unless it’s trivially benign (balanced, complete, fair). Two things to separate: 1. “Flooding” as such (high volume) 2. “Flooding in order to dominate / skew / crowd out” (a particular strategic use of high volume) McKenna is worried about (2). Ross-propaganda kicks in when (2) introduces epistemic defectiveness. Your argument is: (2) essentially always does; if it doesn’t, it collapses into (a benign information service), in which case there’s no problem and no propaganda. That is a perfectly coherent line. What I can do now is: give you the *best* non-propaganda flooding example that’s still recognisably “flooding,” and then show you why you might still think it’s not the situation McKenna has in mind. ## A non-propaganda flooding scenario (coherent, but not McKenna’s target) **Example: “The Open Flood”** A public institution runs an AI “argument generator” for civic topics. It produces a *huge* volume of arguments, but it is designed with explicit constraints: - For any contested question, it generates arguments for *all major sides*. - It labels arguments by side, strength, and assumptions. - It forces symmetry: for every argument it outputs on side A, it outputs a comparable argument on side B (or it flags that no comparable argument exists). - It includes a “representativeness module”: it explicitly tells you whether the set you are seeing is a representative sample of the best arguments known in the literature, and it points you to where dissenting views can be found. - It includes *equal prominence* presentation: the interface does not push one side more than the other; the default view is balanced. - It is transparent that it is giving you a *large* set; it does not license the inference “this is all there is” without warrant. In fact, it explicitly warns against that inference. - It is open-source or otherwise contestable: independent groups can audit, fork, and compete. Now imagine it produces 10,000 arguments per topic. That’s “flooding” in the sense of sheer quantity. But it is *not* epistemically defective in the Ross sense because: - it is not omitting counterarguments (it systematically includes them), - it does not create misleading implicatures about balance (it explicitly addresses balance and representativeness), - it does not mislead about completeness (it explicitly warns about limits), - it doesn’t crowd out alternatives (it links and supports contestability). On Ross, that’s **not propaganda**, because there’s no epistemic defectiveness being conveyed. And on your moral reading, it’s **benign** (or at least *not* the kind of thing that should worry us in the propaganda register). So yes: there is a coherent flooding-without-propaganda case. But notice that this “Open Flood” is *not* the kind of flooding McKenna worries about, because it doesn’t enable domination of the marketplace of arguments by one actor with one agenda. It’s closer to a deliberation tool. ## So why does this not resolve your point? Because your point is deeper: you’re basically saying: > Any flooding that is *strategically deployed* (as McKenna fears) will almost inevitably introduce epistemic defectiveness (through skewed availability, pragmatic effects, crowding out, etc.). > If you design it so those defects don’t happen, you’ve designed away the political worry too — you’ve made it benign. That’s a strong claim, and it might well be true. But it isn’t a conceptual truth. It’s an empirical/structural claim about how flooding works in real communicative environments and in real political economies. McKenna’s project is arguably to say: “Given our world — incentives, concentration, attention scarcity — flooding tends to become domination.” You’re saying: “Right, and when it does, it becomes propaganda (on Ross).” So you might actually be converging on a clean thesis for your paper: **McKenna’s ‘best case’ is not best enough: once you cash out the scale-and-concentration worry, the mechanism by which it bites is precisely the mechanism of epistemic defectiveness (broadly construed), and so the phenomenon he wants to treat as distinct from propaganda falls under propaganda on the revised Ross account.** That would be a very crisp argument. ## What you might still be missing (if anything) Only one subtle thing: Even in the “Open Flood” case, someone could argue it’s still problematic, not because it misleads, but because **volume alone** can be a barrier to effective participation (cognitive overload). That would be a *procedural* complaint, not a propaganda complaint. You might respond: “If it overloads citizens, then it is epistemically defective, because it predictably prevents appropriate uptake.” But that’s a choice: it means you are expanding “epistemic defectiveness” to include “epistemically unusable for bounded agents.” Ross-style accounts might or might not want that expansion. So there’s one genuine point of friction: **does ‘epistemic defectiveness’ include ‘designed in a way that predictably prevents good epistemic uptake, even if it doesn’t mislead’?** If yes, then even some benign floods become propaganda-ish. If no, then benign floods can be non-propaganda yet still troublesome for other reasons. But if you want to keep your crisp dichotomy, you can: it’s defensible to say the only relevant worry is the one that creates misleadingness (broadly construed), and that’s where propaganda begins. ## Applying the same move to the other examples You’re right that a similar “either benign or defective” pushback can be applied to several of my earlier examples: - If the central authority is benign and epistemically transparent, no propaganda, perhaps still illegitimate but you might deny that matters. - If the authority is less benign, the trouble shows up as epistemic defect (selection, omission, framing) and becomes propaganda. That suggests a paper-level conclusion that is either: 1. **Ross collapses McKenna’s space**: once you accept the revised Ross account, the interesting “middle” McKenna wants (worry without propaganda) is unstable: either the system is benign and non-propagandistic, or it becomes propagandistic via epistemic defectiveness. or: 1. **McKenna identifies a third kind of wrong** only if you insist that legitimacy/agency matters independently of epistemic defectiveness. If you don’t, then yes, his dilemma bites hard. So: you’re not confused. You’re holding a very tight line. The remaining work is to decide whether you want to defend the “collapse” thesis (which is powerful and elegant), or whether you want to preserve a non-propaganda wrong (legitimacy/agency) and argue that McKenna’s best case is still problematic even if epistemically clean. Okay, good. I think we're on the same page now. Um I'm gonna keep thinking about the last point you make in the section you called what you might still be missing brackets if anything close brackets I'll think about that in just a minute but in the meantime can you think about what we've just been talking about and uh revise the plan on the canvas okay so I think we really got to the nub of the issue there in terms of the dilemma I think for McKenna. So can you give me a structure which sort of lays out all of this stuff we've been talking about in the most sort of elegant, plausible, coherent way for a paper? Paper Plan: Mc Kenna, Persuasive Ai, And Ross-style Propaganda 1. Working thesis and payoff This paper argues that, once we adopt the revised Ross-style account of propaganda defended by Bonard, Contesi, and Marques (“The defectiveness of propaganda”, Philosophical Quarterly), McKenna’s attempt to isolate a politically troubling “best case” that is not propaganda becomes unstable. The core dilemma is: Either (i) McKenna’s “best case” is idealised so far that it is epistemically non-defective in the broad Ross sense (no misleading omission, implicature, presupposition, etc.), in which case it is not propaganda and (arguably) benign; or (ii) the “best case” remains politically troubling in the way McKenna wants (scale + concentration dominating the marketplace of arguments), but then it predictably produces epistemic defectiveness at the campaign/system level (via omission-by-crowding-out, Quantity-style misleadingness about balance/completeness, and related pragmatic effects), in which case it is propaganda after all. The upshot is a “collapse” result: McKenna’s purported middle category (worry without propaganda) either dissolves into the benign or collapses into propaganda on the revised Ross account. A secondary, optional strand (kept brief) notes a residual possibility: some arrangements might be politically illegitimate even if epistemically pristine (domination without defect). If that route is pressed, it supports a different conclusion: McKenna’s worry may be better captured by an additional category (epistemic domination / deliberative capture) rather than by propaganda. This strand is optional and can be used to hedge if needed. 1. Section 1 — McKenna’s project, properly developed 2.1 Background: influence, dependence, and the suspicion of persuasion Explain McKenna’s starting point: our intellectual lives are shaped by external factors; propaganda/indoctrination are obvious bad cases; persuasion is supposed to differ from overt coercion/manipulation but remains philosophically suspicious. 2.2 The AI persuasion target and the “middle ground” ambition Present the two extremes he rejects: hyper-persuasion alarmism vs scepticism that AI changes nothing. Emphasise his main claim: the key worry is not superior persuasiveness but scale and concentration of capacity. 2.3 What current evidence suggests (overview) Summarise the empirical picture: LLMs are often as persuasive as humans; personalisation can increase effectiveness; persuasion effects are typically modest; there are plausible limits on persuadability. 2.4 The “best case” move: bracketing misinformation and focusing on reasons Explain his conceptual clarifications: persuasion is intentional; treat AI as persuasive technology used by human purveyors; then explicitly bracket misinformation/deception and focus on rational persuasion via arguments and reasons. 2.5 The three arguments and where the weight lies Briefly cover: the rationality argument (largely deflated), the sophistry/marketplace-of-arguments argument (central), and the political argument (regulation as a new battleground). 2.6 End Section 1 with the interpretive landing Make explicit that McKenna mentions propaganda but does not treat his “best case” as propaganda. The point of his “best case” is to show that we should worry even when the outputs are framed as genuine argumentation rather than deception. Transition sentence: the question is whether McKenna’s “best case” can remain both (a) non-propagandistic and (b) troubling once we apply an epistemic-merit account of propaganda that treats defectiveness broadly. 1. Section 2 — The revised Ross account: conditions and the role of defectiveness 3.1 State the revised conditions Present the refined Ross-style conditions (in your own words, faithfully): propaganda requires (i) epistemically defective message (broadly construed), (ii) produced by/on behalf of an institution/cause with intent to persuade for direct political ends, (iii) targeted at a socially significant group, and (iv) awareness (or culpable ignorance) of defect. 3.2 Clarify “epistemically defective” beyond falsity Explain the central innovation for your purposes: defectiveness includes omission, implicature/presupposition, Quantity-style misleadingness, and (more generally) communicative architecture that predictably induces unwarranted beliefs or unwarranted connections among beliefs. 3.3 The “easy verdict” and the dilemma set-up State the key conditional clearly. If McKenna’s “best case” is genuinely non-defective in this broad sense, then it is not propaganda on this account. Therefore, anyone who wants to classify McKenna’s troubling “best case” as propaganda must show that the troubling features (scale + concentration) generate defectiveness at the system/campaign level. 1. Section 3 — The McKenna dilemma as a structured choice This section articulates the dilemma cleanly and motivates why it matters. 4.1 The benign horn: the “Open Flood” / idealised best case Describe an idealised scenario: AI produces large volumes of arguments but is transparent, symmetric across sides, explicitly flags limits, ensures exposure to counterarguments, and is contestable/auditable. Conclude: this is not propaganda (no defectiveness) and it is plausibly benign. 4.2 The troubling horn: marketplace domination as a mechanism Re-state McKenna’s core worry: a small number of actors can deploy persuasive AI at scale, creating armies of arguers-for-hire and dominating what arguments circulate. 4.3 Thesis preview: why the troubling horn tends to be defective Introduce the guiding claim you will defend in the next section: once flooding is deployed strategically to dominate the marketplace of arguments, the campaign predictably becomes epistemically defective (even if individual arguments are locally sound), because it skews availability and invites false inferences about balance, completeness, and decisiveness. 1. Section 4 — Converting McKenna’s troubling “best case” into propaganda This is the core argumentative section. The strategy is to treat the relevant communicative act at the level of the campaign/ecosystem (not individual utterances), then show how defectiveness emerges. 5.1 The unit of evaluation: from utterance to campaign architecture Argue, using the defectiveness framework, that what is conveyed includes pragmatic implications and omissions that arise from distribution, prominence, repetition, and crowding-out. 5.2 Three routes to defectiveness under “flooding + concentration” Route 1: Omission-by-crowding-out Even if counterarguments exist somewhere, flooding can make them practically inaccessible, producing a systematically incomplete evidential basis for belief formation. Route 2: Quantity-style misleadingness about representativeness Saturation on one side predictably licenses false inferences such as “these are the main considerations” or “there are no serious alternatives,” even without explicit denial. Route 3: Defective maps of salience and evidential weight Controlling what becomes salient predictably distorts what audiences take to be decisive or important, generating unwarranted connections among beliefs and attitudes. 5.3 Meeting the other conditions Briefly show how McKenna’s target cases plausibly satisfy: institutional/cause provenance (big tech actors, campaigns, interest groups), intent to persuade for direct political ends (broadly construed), targeting of socially significant groups, and awareness/culpable ignorance of these predictable effects. Interim conclusion: the troubling version of McKenna’s “best case” is propaganda on the revised Ross account. 1. Section 5 — Objection handling and refinement 6.1 Objection: ‘If it’s fair and complete, it’s benign’ Agree. Clarify that this is the benign horn. The paper’s claim is not that all large-scale argument generation is propaganda, but that McKenna’s specific worrying scenario (dominating the marketplace) requires, or tends to produce, defectiveness. 6.2 Objection: ‘Flooding need not mislead’ Respond by narrowing the thesis: Moderate version: Not all flooding is propagandistic, but flooding used to dominate the marketplace of arguments typically becomes defect-involving given attention scarcity and pragmatic inference. Strong version (optional): Under realistic conditions, domination-by-flooding is structurally unstable without defectiveness; if it avoids defect, it ceases to be the worrying case. 6.3 Optional: the residual non-propaganda wrong Very briefly note that if one insists there can be domination without defectiveness, that wrong is better captured as legitimacy/epistemic-domination rather than propaganda on Ross. This is a hedge, not the main line. 1. Conclusion Restate the collapse result. McKenna’s attempt to identify a politically troubling “best case” distinct from propaganda faces a dilemma once we apply the revised Ross account. Either the scenario is genuinely epistemically non-defective and benign (and so not what should worry us), or the scenario remains troubling in McKenna’s sense because it dominates the marketplace of arguments, in which case it becomes propaganda via campaign-level epistemic defectiveness. 2. What to quote (for drafting) From McKenna: (1) The “quantity not quality” passage (scale + concentration). (2) The “marketplace of arguments” + “arguers-for-hire” passages. (3) The passage where he brackets misinformation and focuses on genuine arguments. From Bonard/Contesi/Marques: (1) The refined conditions. (2) Defectiveness beyond falsity (especially omission/Quantity). (3) ‘Message’ understood broadly beyond locutionary content. 1. Immediate next drafting step Draft a full Section 1 (2–3 pages) that culminates in: McKenna’s best case is designed to be non-propagandistic, yet he thinks it remains politically troubling; the next section asks whether that middle category survives the revised Ross account or collapses into either benignity or propaganda. Done — I revised the canvas plan to make the *dilemma* the spine of the paper and to build the most elegant route to your “collapse” conclusion (benign horn vs propagandistic horn).