Does an AI chatbot that argues with you or acts 'off-script' actually change your mind weeks later, or does that effect just fade?

This explores whether a chatbot that breaks users' expectations, by pushing back, taking an unusual stance or simply acting unlike a typical assistant, leaves effects that last beyond a month. The short answer: the corpus has no study that tracks expectation-violating stances past the one-month mark. What it does have is a set of findings about how chatbot effects change over time, and almost all of them point the same way. Early effects tend to fade.

The clearest pattern is decay. Longitudinal work with the Mitsuku chatbot found that the social pull behind relationship formation drops in a predictable way as novelty wears off, so results from a single session can't be stretched to cover weeks or months Do chatbot relationships lose their appeal as novelty wears off?. Persuasion follows the same curve. Claude and DeepSeek began with a strong persuasive edge over human persuaders, but that edge shrank over repeated rounds while the humans stayed steady Does AI persuasiveness fade across repeated conversations with the same person?. That is the reverse of what happens between people, where rapport usually builds. A surprising chatbot stance is partly a novelty effect itself, so the default expectation should be that its impact weakens with exposure.

There is a twist, though. Expectations don't just get violated, they get rebuilt. Research on personalization shows each interaction raises the baseline of what users expect, which makes later failures more disappointing Does chatbot personalization build trust or expose privacy risks?. So a lasting effect may not be a durable change in belief. It may be a durable change in what users now count as 'normal', which makes the next surprise harder to pull off. Related work on sycophancy shows a gap between how users feel about a chatbot and how they are influenced by it. Warnings made flattering chatbots seem less objective and less enjoyable, yet users were just as persuaded Can warnings stop people from being swayed by sycophantic AI?. Any long-term study of unexpected stances would need to measure both, because one can fade while the other stays.

The therapeutic chatbot literature adds a warning about method. Trials that compare a chatbot against a waitlist mostly measure the effect of having someone to talk to, which is why ELIZA can match Woebot Do chatbot trials against waitlists measure real therapeutic value?. A study claiming that a contrarian or unexpected stance has lasting effects would need to rule out that same 'any contact helps' explanation. It would also need to separate the stance from consistency, since users respond more strongly to chatbots that share emotions consistently than to ones that adapt Do chatbots trigger human reciprocity norms around self-disclosure?.

What you may not have expected to learn: the more useful question may not be whether the effect lasts, but which effect lasts. On the corpus's evidence, enjoyment and novelty fade, persuasive edge fades, and baseline expectations ratchet upward. Persuasion that survives users becoming aware of it may be the stickiest of all. Direct evidence on expectation-violating stances beyond one month is a real gap in the collection.