## 1) [[Reasoning models]] can produce reasonable, persuasive arguments - Modern LLMs have demonstrated sustained, stepwise reasoning and persuasive argumentation - Early “reasoning model” paradigm: ChatGPT-01 introduced explicit chain-of-thought style and multi-step planning - Iterative improvements: ChatGPT-03 (stronger reasoning scaffolds; better adherence to premises and goal states) - [[Current state]]: Gemini 2,5 Pro is positioned as a “reasoning model,” optimized for long-horizon inference and multi-modal synthesis - Anthropic’s models (e.g., Claude family) are likewise optimized for extended reasoning, debate-style structure, and citation discipline - In practice, these models can: - Assemble coherent cases for and against policy positions - Anticipate basic objections and offer counter-considerations - Tailor [[argumentative structure]] (premise–support–warrant–conclusion) to audience constraints and time budgets - Bottom line for the introduction - These systems can produce “reasonable, persuasive arguments” for a chosen standpoint - This empirical capability is the backdrop for assessing how such arguments function within rational persuasion and public discourse ## 2) Compatibility with McKenna’s account of rational persuasion - McKenna’s framing: treat AI as a “persuasive technology,” with human purveyors bearing intentions - The fact that LLMs can generate reason-giving discourse is fully compatible with McKenna’s account - McKenna adopts the Mitchell–Douglas definition of rational persuasion (thin, reasons-based) - 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 recognising some of the reasons given by A - (iv) A intends each of (i)–(iii) - Minimal conception of epistemic rationality (McKenna) - Rationality = appropriately responding to evidence, facts, and reasons - Source neutrality: nothing in principle disqualifies AI-generated reasons from counting as reasons - Therefore - LLM-produced arguments can, in principle, participate in rational persuasion as McKenna understands it ## 3) McKenna’s “best case” for AI-mediated rational persuasion - McKenna’s stipulation (verbatim block quote) > “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.” - [[Key takeaways]] from the best case - Reasons-forward persuasion (not deception/manipulation) is the operative method - Being persuaded by such reasons can be epistemically rational (on McKenna’s minimal standard) - This sets a generous informational assumption for ethical/political analysis ## 4) McKenna’s major worry: “sophistry on steroids” (flooding the marketplace of arguments) - Scale and concentration, not superhuman persuasiveness, drive the concern - A small number of actors can mass-produce arguments at negligible marginal cost - High barriers to model training/operation concentrate capability in few hands - The “marketplace of arguments” can be distorted by: - Quantity: overwhelming supply of one-sided, reason-giving content - Access/control: privileged placement, default integration, infrastructural chokepoints - Political upshot - Even if each item is “reasonable,” the aggregate deployment can crowd out rival standpoints - This raises democratic and distributive worries: who decides which arguments are most visible, when, and to whom? ## Part Two — Propaganda as Epistemic Defectiveness, and why McKenna’s “best case” still qualifies ### The Epistemic Defectiveness Model (EDM) of propaganda - Propaganda = a communicative act that is epistemically defective and deployed for political ends toward a socially significant audience, with awareness (knew/should have known) - Key conditions (crisp): - Message (Defect): The act conveys an epistemically defective message in a broad sense - [[Not just]] lying; includes: - Misleading omissions (failing to handle salient counter-considerations that bear on the claim’s decisiveness) - Unwarranted presuppositions or implicatures (suggesting decisiveness, consensus, or completeness that the act has not earned) - Inapt or tendentious connections (rhetorical framing that induces unwarranted inferences) - Illocutionary/perlocutionary defects (the way the act functions induces unwarranted beliefs) - Sender (Political End): Produced by or on behalf of a political/social institution/organization/cause, with intent to persuade for direct political ends - Receiver (Socially Significant Group): Aimed at large publics or politically relevant constituencies - Awareness: Producer is aware, or should have been aware, of the epistemic defect - Important elaboration: EDM does not equate propaganda with simple falsehood - An act can be propagandistic even if many or all its sentences are true - The defect resides in how the act induces an unwarranted picture of decisiveness, balance, or consensus ### You might think: “In McKenna’s best case, this isn’t propaganda” - Best case (recap): AI is used to present “genuine arguments, evidence, and reasons … to rationally persuade, not simply deceive or mislead” - Temptation: If there’s no lying and the discourse is reasons-forward, then it can’t be propaganda - (it's a problem whether or not it is propaganda full) ### The six-premise argument: why McKenna is wrong if he thinks his best case escapes propaganda - P1. A communicative act is non-deceptive only if it does not induce unwarranted beliefs in its audience about the status of the claim (e.g., its decisiveness over live alternatives). - P2. In a live, contested dispute, omitting engagement with salient alternative views that would defeat the claim induces an unwarranted belief of decisiveness (i.e., that the claim’s reasons are superior to rivals) in the audience. - P3. Therefore, in live disputes, non-deceptive persuasion requires engaging salient alternatives (at least those that would, if true, negate the claim). - P4. McKenna’s “best case” adds a non-deception constraint (“not simply deceive or mislead”) to a thin, reasons-based model of rational persuasion, but does not require engaging salient alternatives. - P5. Therefore, McKenna’s “best case” fails to satisfy its own non-deception constraint in live disputes; it is confused/incomplete. - P6. When a small number of powerful purveyors of AI persuasion mass-produce and disseminate such “best-case” arguments at scale—using their privileged resources and distribution channels to dominate the “marketplace of arguments”—this concentrated deployment induces unwarranted beliefs about decisiveness and the balance of reasons across large publics. - C. Hence, contrary to initial appearances, concentrated, large-scale deployment of such arguments by powerful purveyors constitutes a communicatively defective, propagandistic practice (propaganda by omission), even without outright lying. ### Short clarifications and implications - “Engaging salient alternatives” does not mean exhaustive cataloguing; it means addressing those rival positions that would, if true, defeat the conclusion you are presenting as settled - The defect in EDM is typically omission/implicature: inducing audiences to treat a claim as decisively supported where that decisiveness has not been established vis-à-vis live rivals - McKenna’s own structural worry magnifies the defect: when few actors dominate production and dissemination, the unwarranted inferences (decisiveness, consensus, evidential balance) are propagated at scale - Therefore, even on McKenna’s generous informational premise (reasons-forward, no outright deception), his best case does not escape propaganda once we apply the epistemic defect standard to live disputes and concentrated deployment