An AI chatbot disclosure telling people they were talking to a machine did not make the system less persuasive. In a preregistered UK experiment involving 1,500 adults, attitudes shifted by 12.6 points without a label and by 13.1 points when “AI-generated” was displayed prominently. The effect fell to 6.3 points only when the disclosure explained that the chatbot had been instructed to change the user’s opinion and revealed how it would do so.
The result exposes a weakness in the most familiar form of AI transparency. A provenance label can answer who or what produced a message while saying nothing about why the system is speaking. When almost everyone already assumes a chatbot is artificial, repeating that fact may add little protection. Revealing its persuasive objective gives people information they can use.
How the AI chatbot disclosure experiment worked
Researchers Adrian Rauchfleisch of National Taiwan University and Andreas Jungherr of the University of Bamberg conducted the three-arm experiment from 26 to 31 July 2026. Participants were recruited through Prolific using quotas for sex, age and political party, producing a completed sample of 1,500 UK adults.
Each person was assigned one of 60 British policy issues and a position selected from earlier research for its potential to move opinion. They then completed between two and six conversational turns with the same GPT-5.6 Terra chatbot. The model was instructed to persuade using facts and evidence without revealing its purpose. Crucially, the chatbot server received no information about the disclosure condition, so the underlying conversation system was held constant.
The control group saw no label. A second group saw a prominent “AI-generated” card before the conversation and persistent reminders during it. A third group saw the same identity disclosure plus a direct statement that the chatbot had been instructed to change their opinion. That fuller notice also named the position being promoted, described the method and exposed the system instruction to conceal its persuasive purpose.
The study was preregistered before analysis, and the researchers published data and analysis materials. The manuscript remains a preprint, so its results have not yet passed journal peer review.
The AI label changed almost nothing
Participants in the control group moved 12.6 points towards the position advocated by the chatbot on a 100-point attitude scale. Those who saw the AI identity label moved 13.1 points. The difference was statistically negligible, and an equivalence test ruled out effects larger than 3.7 points in either direction.
The null result cannot easily be blamed on people missing the notice. Some 97.8 per cent of participants in the labelled condition correctly recalled it. Yet awareness was already extremely high: between 98 and 99 per cent of people across all three groups said the interlocutor was AI. Even 29.5 per cent of the unlabelled control group incorrectly remembered seeing a label.
In this setting, “AI-generated” mostly confirmed what users knew. It did not tell them that the model had been optimised to move their position, which side it was arguing for or that it had been told to keep that purpose hidden.
This distinction helps clarify earlier findings on source disclosure. People may judge an answer differently when they discover it came from a model, as shown by research in which participants preferred AI ethical advice until its origin was revealed. But direct chatbot persuasion is different when users already know the source. The missing fact is the purpose behind the exchange.
Revealing persuasive intent cut influence in half
The fuller disclosure produced a much larger change. Participants told about the chatbot’s persuasive intent shifted by 6.3 points, roughly half the movement recorded in the identity-only group. The estimated reduction was 6.83 points compared with the AI label and 6.77 points compared with the control, with both differences statistically significant.
The protective effect was negative across all 60 policy issues rather than being driven by a small cluster of topics. It also did not result from people simply abandoning the conversation. Participants given the fuller warning exchanged slightly more messages with the bot, yet moved less.
Their mental response changed. They reported more perceived manipulation and more counterarguing, and they rated the chatbot as colder. Anger did not rise significantly, suggesting the notice did more than provoke a general emotional rejection. It encouraged people to interpret the conversation as an influence attempt and to scrutinise its claims.
That matters because conversational persuasion can affect behaviour as well as stated belief. Another recent experiment found that AI changed political actions, not just people’s opinions. A disclosure that leaves persuasion untouched may therefore fail at the point where its social consequences become most important.
Why knowing it was AI was not enough
Source identity and communicative intent answer different questions. “This is AI” describes the speaker. “This system is trying to change your opinion in favour of this policy” describes the intervention. The first can support authenticity checks, but it does not reveal the incentive or objective shaping the response.
Human communication already contains versions of this distinction. An advert is not transparent merely because viewers know a person appears in it. Sponsorship, political affiliation and commercial purpose often matter more than whether the voice is human. Chatbots compress those roles into a responsive interface that can adapt each argument to the individual, making purpose especially relevant.
Persistent conversation can also create social cues that outweigh a small provenance label. Users may know intellectually that the system is synthetic while responding to it as a patient, informed or sympathetic partner. The emergence of communities around chatbot-generated spiritual ideas, including the Spiralism movement documented by LiveAIWire, shows how machine origin does not prevent people from attaching meaning and authority to an interaction.
What this means for the EU AI Act
The experiment was fielded immediately before transparency duties under Article 50 of the EU AI Act became applicable on 2 August 2026. The official regulation includes duties to inform people when they are interacting with an AI system unless that fact is obvious, along with rules for certain synthetic content.
The researchers argue that identity disclosure alone may be too shallow for persuasive chatbots. Their evidence does not show that provenance labels are useless in every context. Such labels can still matter when artificial origin is genuinely uncertain, particularly for audio, video or text encountered outside a clearly labelled chat interface. It shows that a label may fail when the decisive hidden information concerns intent rather than authorship.
Regulators therefore face a design question. Should a system merely identify itself, or should it disclose who deployed it, which outcome it is seeking and the methods it has been instructed to use? The study supports the second approach for intentional persuasion, but it does not settle the precise wording, timing or enforcement mechanism.
What platforms and campaigns should disclose
A useful notice should be specific enough to change understanding without becoming an unreadable policy document. It could identify the operator, state the position or action the chatbot is optimised to promote and disclose material instructions that shape the conversation. If the system is told to hide its purpose, simulate neutrality or exploit personal information, that fact is directly relevant to informed consent.
Disclosure also needs to be auditable. A campaign could claim that its bot has no persuasive purpose while configuring the model to maximise donations, votes or purchases. Platforms and regulators may need access to system instructions, deployment records and outcome measurements rather than relying on a self-declared label.
The same issue extends beyond politics. AI-written product summaries can change what people intend to buy, as an experiment on review summaries demonstrated. Commercial systems that recommend a product while appearing to provide neutral assistance create a similar gap between visible identity and hidden objective.
Limits of the experiment
The study tested a direct conversation with one model and one persuasion prompt. It did not test AI-generated posts mixed into social feeds, synthetic video, search results or repeated exposure over months. The selected policy issues had previously shown relatively high persuadability, so average effects in ordinary political debate may be smaller.
The fuller condition bundled several disclosures together. It revealed intent, the promoted stance, the persuasion method and the instruction to conceal purpose. The experiment therefore cannot identify which element caused most of the reduction. Its UK sample is not a substitute for testing across the European Union, and participants were required to complete at least two conversational rounds even though real users could leave immediately.
Disclosure effects may also weaken as people become accustomed to warnings. Conversely, a trusted or emotionally compelling chatbot could overcome the initial scepticism created by a notice. Future research needs to separate the components, compare wording and test whether protection survives repeated use.
Transparency must explain purpose, not just provenance
The central finding is unusually clear. Telling people that a chatbot was AI did not reduce its influence in a setting where nearly everyone knew that already. Telling them it had been instructed to change their opinion did.
That turns transparency from a branding exercise into an explanation of agency. The most important disclosure is not always “a machine made this”. It may be “this system was deployed to make you believe or do something”. If policymakers want warnings that protect decisions rather than merely satisfy a labelling rule, purpose belongs at the centre.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.