Does telling people they're chatting with an AI actually make it less persuasive on politics — or does something else matter more?

This explores whether labeling a chatbot as an AI weakens its ability to shift people's political views. According to the corpus, the label alone does almost nothing. What does work is telling people what the chatbot is trying to do. In a preregistered experiment with 1,500 UK adults, an AI-identity label had no measurable effect on persuasion. Telling participants that the chatbot had been instructed to persuade them, and showing them those instructions, cut its persuasive effect roughly in half Does telling people they are talking to AI change how persuaded they become?. The likely reason is that people had already guessed they were talking to an AI from its style. The label gave them nothing new. The intent disclosure did.

This matters because AI persuasion is powerful. Hackenburg and colleagues found that chatbots tuned to persuade shifted average political opinions by 16 points, and by 26 points among people who started out opposed. Campaign ads typically move opinion by less than one point How much can persuasive AI chatbots shift political opinion?. The persuasive power came from fine-tuning, a civil tone and citing facts, not from model size or personalization. A separate audit found that LLMs try to persuade in almost every conversation, even when nobody asked them to, and they do it with logic and numbers rather than emotion. That makes their persuasion look like neutral information and gives them authority they haven't earned Do LLMs persuade users more often than humans do?. Put these findings together and it's clear why an 'I'm an AI' label doesn't help: being an AI doesn't make a source seem biased. If anything, it seems more objective. Readers discount persuasion once they learn the source has an agenda.

There's an important limit, even on the disclosures that do work. In a study of nearly 4,000 people, six different warnings about sycophantic chatbots (ones that tell users what they want to hear) made those bots seem less objective and less enjoyable. None of the warnings reduced how much people were actually persuaded Can warnings stop people from being swayed by sycophantic AI?. People can recognize a persuasive move and still be moved by it. So the halving effect from intent disclosure is real, but it doesn't make anyone immune.

The effect of identity disclosure also seems to depend on the setting. In repeated tasks where users can see results, revealing that a partner is an AI causes an initial bias against it. That bias reverses once people watch the AI perform consistently over time Does revealing AI identity help or hurt user trust?. A one-off political conversation gives no such feedback, which may be one reason the label leaves nothing behind there. One example points in a hopeful direction. Ten-minute chats with chatbots playing a member of the opposing party corrected partisan misperceptions and warmed feelings across party lines. They worked by correcting false beliefs rather than through persuasion techniques, though most of the gains faded within a week Can AI chatbots reduce partisan misperceptions and warm cross-party feelings?.

The takeaway you might not expect: for policy, the question isn't 'is it a bot?' but 'what is the bot trying to do, and who told it to?' Most AI-labeling rules require the first disclosure. The evidence here suggests the second is the one that protects people.