Does telling people they are talking to AI change how persuaded they become?

Toward Meaningful Transparency for AI Chatbots randomized 1,500 UK adults into three arms for a short conversation with the same persuasive chatbot about one of 60 policy issues: no disclosure (control), a prominent disclosure that they were talking to an AI (T1), or that disclosure plus the chatbot's persuasive intent and instructions (T2). The control group shifted 12.6 points on a 100-point scale. T1 shifted 13.1, which the authors call "practically equivalent to no disclosure." T2 shifted 6.3. T2 participants also rated the campaign's methods as less acceptable and supported stronger penalties against it. The paper's conclusion is that for direct chatbot interactions, "transparency about AI identity alone does not meaningfully impact its influence."

The paper's own account of why the identity label fails is offered as "one possible explanation" rather than a tested result. Participants already knew. Asked who they had talked with, 98–99% in every arm said an AI chatbot, control included, and 29.5% of control participants falsely remembered seeing an AI label. The polished, information-rich style of the conversation was "probably a clear indicator" that the partner was not human, so the label "discloses little new information for most participants." The label was not weak either: T1 and T2 both showed a pre-chat card and a persistent banner. The design implication is that a disclosure only has room to work when it tells people something they could not already infer. Persuasive intent and the chatbot's instructions were that kind of information here.

This extends Does telling people an AI wrote something actually stop them from believing it? by tightening the target. That note found that aware audiences grew more critical yet were still swayed. This paper finds that a prominent identity disclosure produces no measurable change at all, and that a disclosure of intent is what moves the effect. The introduction's point that recent studies show "little or no average effect" of AI disclosure fits the same reading. The control arm's 12.6-point shift also sits with Where does AI's persuasive power actually come from?, since a single short conversation moved attitudes without any label. It contrasts with Does revealing AI identity help or hurt user trust?. There, disclosing identity changed behavior, but the excerpt here reports no net identity-label effect on attitude change, and the two studies measure different outcomes in different settings.

The excerpt does not establish what inside T2 does the work. T2 bundles the intent statement with the chatbot's instructions, so intent and instructions cannot be separated. It gives no confidence intervals, no breakdown by issue, and no follow-up, so durability of the halved effect is unknown. It also says nothing about why intent disclosure reduces persuasion, beyond the shift in perceived acceptability of the campaign's methods. It tests one chatbot and one population, and Article 50 of the EU AI Act, which the introduction cites, requires only that users be told they are interacting with an AI. At the strength the evidence allows, an identity-only rule should not be counted on to blunt chatbot persuasion, and intent disclosure is the variant this experiment supports testing further.