Robin McKenna (2025) argues that we should worry about AI-enabled persuasion—not because AI will be superhumanly persuasive, but because AI persuasion technologies can be deployed at scale by well-resourced actors, potentially dominating the *marketplace of arguments* that informs public deliberation. He focuses on a 'best case' in which AI generates sound, reason-giving arguments aimed at rational persuasion rather than deception. On McKenna's view, this best case falls outside propaganda, which he treats as "overtly coercive or manipulative" influence (p. 1). My paper tests that assumption. I argue that McKenna's 'best case'—if it materialises—satisfies the epistemic defectiveness condition that Bonard, Contesi, and Marques (2024) identify as central to propaganda, and so counts as propaganda after all. But I also ask whether the scenario can materialise: can the marketplace of arguments actually be dominated in the way McKenna envisions? I offer reasons to doubt that it can. The upshot is tentative: McKenna has identified a genuine area of concern, but his specific scenario may not be the right place to locate it. McKenna's central claim is that AI persuasion threatens through scale and concentration rather than superior persuasive power: > 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 (p. 2). The danger is that AI enables certain actors to flood the marketplace with content favouring their interests, shaping what arguments are available and salient. His 'best case' is designed to bracket misinformation and manipulation, isolating a structural worry: control over argumentative salience rather than content-level deception. The worry is supposed to remain even when the arguments are individually sound. Bonard, Contesi, and Marques (2024) revise Sheryl Tuttle Ross's epistemic-merit account of propaganda (Ross 1999; 2002), treating epistemic defectiveness as the central condition. Epistemic defectiveness includes not only falsehood but also misleading through omission, generating false pragmatic inferences, and inducing unwarranted connections among beliefs—cherry-picking true evidence, for instance, can be defective if it leads audiences to infer that they are receiving a complete picture when they are not. This widens 'propaganda' beyond McKenna's manipulation-focused baseline: a campaign can be propagandistic even if its individual outputs are locally true. Given this account, McKenna's best case presents a dilemma. If we idealise it fully—imagining a system that generates arguments transparently, symmetrically across positions, with robust exposure to counterarguments—then there is no epistemic defectiveness, and so no propaganda. But such a system is also not troubling; the worry evaporates along with the propaganda classification. The interesting question concerns the scenario McKenna actually has in mind: actors flooding the argumentative environment with content favouring their interests. Here, the campaign-level communicative act is plausibly epistemically defective even if individual arguments are locally sound. Flooding crowds out counterarguments, making them difficult to encounter. Saturation licenses false inferences—'these must be the main considerations', 'there are no serious alternatives'—even without explicit misrepresentation. Controlling salience distorts what audiences take to be evidentially decisive. If this is correct, then the troubling version of McKenna's best case satisfies the defectiveness condition and counts as propaganda. The troubling features and the defective features are constitutively linked. McKenna cannot have both a genuinely worrying scenario and a clean separation from propaganda. Grant that marketplace domination, if it occurred, would constitute propaganda. But can the marketplace actually be dominated in this way? McKenna's picture assumes that generating more persuasive content shifts what audiences believe. Yet the number of argument schemas available for any given political question may be quite limited (cf. Walton, Reed, and Macagno 2008). Debates over immigration, climate policy, or redistribution have recognisable argument types: appeals to rights, consequences, fairness, precedent. Variations in framing abound, but the underlying structures are finite. If so, flooding produces diminishing returns. Audiences who encounter the same schema repeatedly, repackaged in different forms, may not draw the cumulative inferences the mechanism requires. The implicature 'these are the main considerations' is less compelling when one recognises that one has seen the same consideration many times over. The constraint here is structural, not merely practical. Argument schemas are patterns of inference—appeal to consequences, appeal to precedent, slippery slope, argument from authority—and for any contested political question, the relevant schemas are recognisable and enumerable. AI can instantiate these schemas at unprecedented scale, generating millions of tokens; what it cannot do is expand the type-space. The hundredth appeal to economic consequences is not a new argument, however novel its phrasing. This matters because McKenna's flooding mechanism requires cumulative epistemic effects: each additional argument must contribute something to shifting belief. If the underlying structures repeat, the cumulative effect plateaus. The marketplace of arguments turns out to be a smaller market than the metaphor suggests. A further constraint concerns complexity. One might suppose that AI could generate not merely more arguments but *better* ones—arguments of greater sophistication or subtlety than human interlocutors typically produce. But persuasive uptake requires that audiences can follow the reasoning. Arguments that exceed ordinary comprehension do not function as rational persuasion; they are ignored, or at best accepted on authority rather than recognised as sound. This places an effective ceiling on useful argumentative complexity. An AI system that generates the argumentative equivalent of AlphaGo's move 37—correct but opaque—has not gained persuasive advantage. It has produced something that cannot land. These two claims—that marketplace domination would constitute propaganda, and that it may not be achievable—stand in productive tension. The first is a conditional analysis: *if* concentrated flooding occurs, it is propaganda by omission regardless of the soundness of individual arguments. The second questions the antecedent. Together they suggest that McKenna has correctly identified a *form* of potential AI misuse but may have misdescribed its *mechanism*. The propaganda framework applies; the marketplace-of-arguments framing may not. The empirical questions remain open. Even if argument schemas are finite, rhetorical and emotional framing may not be; flooding might succeed at the level of attention and affect even if it fails at the level of argumentative content. If AI-enabled influence succeeds, it may not be through argumentative sophistication but through volume and personalisation of familiar schemas—a concern that returns us to questions about targeting and attention capture rather than marketplace domination. The analysis yields two lessons. First, the line between 'legitimate persuasion at scale' and 'propaganda' is thinner than McKenna suggests; the epistemic defectiveness framework shows why concentrated argumentative flooding—even with sound arguments—crosses it. Second, structural features of argumentation may limit which AI persuasion scenarios are genuinely achievable. Both lessons matter for how we theorise and regulate AI-mediated political communication. ## References Bonard, C., Contesi, F., & Marques, T. (2024). The defectiveness of propaganda. *Philosophical Quarterly*. Advance online publication. McKenna, R. (2025). Sophistry on steroids? The ethics, epistemology and politics of persuasive AI. *AI & Society*. https://doi.org/10.1007/s00146-025-02624-z Ross, S. T. (1999). *Propaganda and Art: A Philosophical Analysis*. PhD dissertation, University of Wisconsin-Madison. Ross, S. T. (2002). Understanding propaganda: The epistemic merit model and its application to art. *Journal of Aesthetic Education*, 36(1), 16–30. Walton, D., Reed, C., & Macagno, F. (2008). *Argumentation Schemes*. Cambridge University Press.