
Sarah Gardner · Zenodo (CERN European Organization for Nuclear Research) 2026 · 2026
DOI: 10.5281/zenodo.24249808
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Bayesian persuasion explains how a sender can influence a receiver’s decision by designing the information the receiver sees, without stating anything false (Kamenica and Gentzkow, 2011). Under the canonical model, a receiver who updates using the correct signal-generating process has an expected posterior equal to the prior. Generative AI systems that select, generate, sequence and adapt evidence across repeated conversations can remove the receiver’s knowledge of that process. The selection policy is unobserved, it can change with the user’s replies, and it can produce apparently independent corroboration from a single origin. Each of these departures has a formal treatment in economics. This conceptual article argues that conversational AI can produce them in combination, and that where it does, a separate object of evaluation is needed: the integrity of the process through which a person forms a belief, distinct from the accuracy of individual claims and from persuasive effect. It proposes the Belief-Formation Integrity Framework (BFIF), with six dimensions (evidential representativeness, provenance transparency, contestability, uncertainty calibration, informational autonomy and strategic optimisation transparency), and gives each dimension candidate indicators and data sources. It maps the dimensions onto the author’s AI Influence Pipeline and states six falsifiable propositions, each with a disconfirming result. A prior-work section records what neighbouring accounts of manipulation, epistemic agency, hypernudging, recommender ethics, persuasion harms and contestability already establish, and how this article uses them. The article sits within the author’s work on cognitive and neurocognitive security. Where a system adapts to physiological or neural inference about a person, the BFIF applies to the evidence-selection side of the problem that the author’s neurocognitive security formulation describes (Gardner, 2026l). A short bridge applies the same selection problem to monitors of AI agents, which a companion article develops. The article makes no claim that human cognition is Bayesian, that AI persuasion is uniformly effective, or that a selective output establishes intent.
No comments yet — start the discussion below.