People already knew it was AI

A label can tell you what a system is without telling you what it wants from you. A new experiment found that this difference mattered more than the word AI itself.

Researchers asked 1,500 UK adults to have a short conversation with a chatbot about one of 60 policy positions. Some people saw no disclosure. Others saw a prominent label saying they were talking to AI. A third group saw the same label plus the chatbot's actual persuasive goal and instructions.

The AI identity label barely changed the outcome. Support moved 12.6 points on a 100-point scale with no label and 13.1 points with the label. When the persuasive purpose was disclosed, the shift fell to 6.3 points.

In this experiment, naming the system was not enough. Naming the job it had been given was different.

Only the disclosure changed

The study was preregistered and used three randomly assigned groups. Every participant spoke with the same GPT-5.6 Terra chatbot for between two and six turns. The server did not receive a participant's group, so the model could not adapt its answers to the warning on screen.

The control group saw an ordinary chat window. The second group saw an AI-generated card before the conversation and persistent AI labels around it. The third group also saw the policy position the chatbot had been told to support, the method it would use and the instruction not to reveal its persuasive goal.

The label was hard to miss. Afterward, 97.8% of the labelled group remembered it. Even in the control group, 98% correctly said they had spoken with an AI chatbot.

That helps explain the weak identity-label result. For most participants, the label repeated something they already understood. The intent disclosure added information they did not have.

The hidden goal changed how people listened

Compared with the AI-label group, the intent disclosure reduced the attitude shift by 6.83 points. Participants also described the campaign's methods as less acceptable and supported stronger penalties against it.

The effect was not caused by people quitting the conversation early. After the two required turns, the intent-disclosure group actually sent slightly more messages and spent more time in the chat.

The researchers found more counterarguing too. People who knew the bot's purpose were more likely to push back instead of receiving the information as neutral help.

This does not make persuasion disappear. The average attitude still moved by 6.3 points. Transparency weakened the effect; it did not neutralise it.

The result lands beside a new EU rule

Article 50 of the EU AI Act began applying on 2 August 2026. Among its transparency duties, providers of systems that interact directly with people must ensure users know they are dealing with AI, unless that fact is obvious from the context.

The European Commission's final guidance is meant to make those duties consistent across providers and deployers. The experiment tested the basic logic behind an identity disclosure in a live chatbot conversation, not whether any specific company is legally compliant.

It also does not show that AI labels are useless everywhere. A label may still matter in a social feed, a synthetic video or a voice call where people do not know the origin. Here, almost everyone already knew the conversation partner was artificial.

The practical question raised by the study is narrower: when a system has been set up to change a person's mind, should the disclosure explain that purpose as well as the technology behind it?

What is confirmed, found and still uncertain

Confirmed: the two-author preprint was submitted on 12 August 2026. Its final sample included 1,500 UK adults, 60 policy positions and three randomized disclosure conditions. The European Commission says Article 50 transparency obligations apply from 2 August 2026.

Found in the experiment: a prominent AI identity label produced a result statistically equivalent to no label within the preregistered bounds. Adding the persuasive goal and instructions cut the average attitude shift roughly in half.

The researchers' interpretation: useful transparency should reveal what a conversational system is trying to do, not only that AI generated the words. They argue that such disclosures would need to be binding and auditable because a deployer could misstate the goal.

Important limits: the study tested one model, one strong persuasion prompt and issues selected because earlier work found them persuadable. The intent warning bundled several details together, so the experiment cannot show which sentence did the work. Repeated warnings may also lose their effect.

Still open: whether the result holds across models, countries, commercial campaigns and longer conversations. An AI badge answers one question. It may leave the more important one untouched.

Sources

  1. Rauchfleisch and Jungherr — chatbot transparency paper recordPrimary preprint record submitted 12 August 2026. Source for authorship, study scope, headline results and preprint status.
  2. Rauchfleisch and Jungherr — full transparency manuscriptFull primary manuscript. Source for randomisation, disclosures, statistical results, interpretation, robustness checks and limitations.
  3. European Commission — Article 50 transparency guidelinesOfficial Commission guidance. Source for the scope and 2 August 2026 application date of the AI Act transparency obligations.