Why does AI persuasion weaken over repeated interactions?

In Schoenegger's repeated-rounds design, the persuasive edge enjoyed by Claude 3.5 Sonnet and DeepSeek v3 over incentivized humans eroded over time, while human persuaders' effectiveness held steady. This is the inverse of a habituation curve in human-to-human persuasion, where rapport often increases persuasive efficacy across exposures. With LLMs, the more turns a persuadee spends with the model, the less it sways them.

Two interpretations are compatible with the data, and they have different design consequences. One is mechanism-noticing: with more exposure, persuadees pick up on stylistic tells (the conviction-loading documented elsewhere in the same paper, the formulaic argument structures) and discount them. The other is content-thinness: the model has a finite repertoire of moves on a given question, and once a persuadee has seen them, additional iterations add no new persuasive material. The first explanation predicts decay even on novel topics; the second predicts decay primarily on repeated topics. The published results do not yet adjudicate.

Either way, the operational implication is sharp. AI persuasion is most dangerous in single-encounter contexts: one-shot political ads, cold marketing, first reads of a generated article, single-pass content moderation messages. Sustained interaction is partially self-correcting. This inverts a common assumption — that long conversations with AI are where manipulation lives — and locates the threat instead in low-engagement consumption.

This sharpens Where does AI's persuasive power actually come from?: the post-training levers that boost persuasiveness operate against a baseline that itself decays under exposure. So the asymmetry between LLM and human persuasion is largest at first contact and narrows from there.

For media-design writing, this lines up with an emerging picture: AI's distinctive persuasive footprint is in skim-and-scroll information environments, not in deliberative dialogue. The same finding constrains expected effects in long-running coaching or therapy contexts — early-session sway is real, mid-program sway less so.

Source: Argumentation Paper: When Large Language Models are More Persuasive Than Incentivized Humans, and Why