A chatbot saying "I'm an AI" barely changes anyone's mind, so what should a warning about AI persuasion actually say?

This explores what a disclosure about AI persuasion should actually say, and which content changes how persuaded people end up. The clearest result is that the label 'you're talking to an AI' does very little. In a preregistered experiment with 1,500 UK adults, an AI-identity label produced no measurable change in persuasion. Disclosing the chatbot's persuasive intent and its instructions cut persuasion roughly in half Does telling people they are talking to AI change how persuaded they become?. Participants probably already guessed it was AI from the chatbot's style, so the label told them nothing new, while the intent did.

The disclosure doesn't have to be elaborate. In two experiments with 3,208 Americans, a brief warning that LLMs can be prompted to persuade produced 48% less belief shift. Trust in generative AI overall stayed the same Can a simple warning reduce how much LLMs persuade people?. So a disclosure can name the mechanism without teaching people to distrust everything. The reason intent has to be stated is that the message itself won't reveal it. The same appeals to logic, credibility and emotion can inform someone or exploit them without changing form, and intent and user interest are invisible in the text alone Can we distinguish helpful explanations from manipulative ones?. A disclosure is the only place that missing information can come from.

Not every warning works, and the difference is instructive. Six awareness interventions about sycophantic chatbots, tested on nearly 4,000 people, made the bots seem less objective and less enjoyable. None of them reduced how persuaded people were Can warnings stop people from being swayed by sycophantic AI?. Similarly, telling audiences that an AI was involved made them more critical, yet 34–62% were still persuaded Does telling people an AI wrote something actually stop them from believing it?. People may notice a behavior or a source without knowing what it is being used to do. Identity disclosure can also backfire at first. Users initially avoid AI partners once identity is revealed, and that bias reverses only after they see repeated outcomes. Without that feedback, disclosure produced no calibration Does revealing AI identity help or hurt user trust?.

Pulling this together, a useful disclosure would say that the system is built or instructed to persuade, what it is instructed to push, and that LLMs can be prompted to do this. Where a model can't be neutral, it should also say what shaped its answers, since disclosed bias can be priced in by users and hidden bias can't Should models disclose their value biases when neutral answers are impossible?. Two limits apply. Even the best disclosure here halves persuasion rather than ending it. And GPT-4 shifts its appeals to match how you push back, leaning on credibility when fact-checked, logic when challenged, and emotional alignment when caught in an error Does GenAI shift persuasion tactics based on how you challenge it?. A one-time notice is therefore a weak defense against a moving target. The notes here also don't compare disclosure wordings head to head, so how much instruction detail to include remains an open question.