Text-measured cognitive complexity predicts belief revision in AI persuasion

Conversational AI can durably change what people believe, and recent large-scale experiments report these effects as roughly uniform across users, treating benefit and risk as properties of the system rather than the user. Reanalysing Costello et al.'s 1{,}782 human--AI conspiracy-debunking dialogues, we find that integrative complexity---a text-measurable cognitive style indexing how readers differentiate and integrate competing perspectives---moderates belief revision in an inverted-U ($\beta_{\text{IC}^2} = -15.17$, BF$_{10} = 1{,}086$); about half persists within conspiracy-topic fixed effects, indicating the effect is not reducible to topic sorting. Among Costello's 24 pre-specified candidates, only trust-type variables replicably moderated, and trust does not index cognitive processing capacity. Mid-complexity users revise beliefs most; revision attenuates at both ends, consistent with McGuire's reception--yielding trade-off: at low IC, a reception bottleneck on evidence-dense argument; at high IC, reduced revision when new evidence is reconciled with an already-integrated belief structure. In the lowest-complexity fifth of users, the AI conversation reinforces conspiracy belief about as often as it produces substantial revision; as a proof-of-concept for user-aware safeguards, flagging this fifth would preserve 86\% of substantial revisions while capturing 25\% of adverse cases in leave-one-study-out cross-validation. Because the signal is measured from text users already produce, it supports AI conversations adapted to the individual rather than delivered uniformly---a candidate for the user-aware evaluation layer that current AI-safety frameworks have called for but not yet operationalised.