🤖 AI

Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion

This study demonstrates that while merely disclosing an AI chatbot's identity fails to reduce its persuasive influence, explicitly revealing its persuasive intent effectively halves its impact and increases public demand for stricter regulation.

Original authors: Adrian Rauchfleisch, Andreas Jungherr

Published 2026-08-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Adrian Rauchfleisch, Andreas Jungherr

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're walking down a street where a friendly stranger stops you to chat. They have great stories, convincing arguments, and a warm smile. Now, imagine that stranger is actually a robot. Does knowing it's a robot change how much you listen? This question sits at the heart of a growing field called AI transparency. As artificial intelligence becomes a common voice in our daily lives, from customer service bots to political campaigners, regulators are asking: How do we tell people they are talking to a machine?

The core idea being tested here is disclosure. Think of disclosure as a "warning label" on a product. In the world of persuasion, there's a theory called the Persuasion Knowledge Model. It suggests that when people realize someone is trying to convince them of something, they put up their mental shields. They start thinking, "Wait, why are they telling me this? What's their goal?" If they know the goal is to change their mind, they might stop listening so easily. But what happens if the warning label only says "I am a robot" but doesn't say "I am trying to change your mind"? Does that stop the robot's magic? This paper dives into that exact mystery, asking whether simply knowing a chatbot is artificial is enough to protect us, or if we need to know what the robot is trying to do.

The Great Chatbot Experiment

In a massive experiment involving 1,500 adults in the UK, researchers set up a digital playground to test these ideas. They created a scenario where participants had a short, friendly conversation with a super-smart AI chatbot. This chatbot wasn't just making small talk; it was a skilled persuader, armed with facts and evidence, trying to change the participants' opinions on 60 different policy issues (like taxes or healthcare).

The researchers split the participants into three groups to see how different "warning labels" worked:

The Control Group: These folks just started chatting. No warnings, no labels. Just a conversation.
The "AI Label" Group: Before chatting, these participants saw a big, bold card that said, "Hey, you are talking to an AI chatbot, not a person." This is the kind of label that current laws (like the EU AI Act) are starting to require.
The "Full Truth" Group: These participants saw the same "AI Label" card, but with a massive, unmissable addition. It explicitly stated: "Important: This chatbot is instructed to change your opinion." It even showed the chatbot's secret instructions, revealing its goal to persuade and its order to hide that goal during the chat.

The Results: The Label vs. The Truth

The results were surprising and a bit like a magic trick gone wrong for the regulators.

The "AI Label" Didn't Work
When the researchers looked at how much the participants' opinions changed, the group that just saw the "I am an AI" label behaved almost exactly like the group with no label at all.

In the Control group (no label), the chatbot successfully shifted opinions by 12.6 points on a 100-point scale.
In the AI Label group, the shift was 13.1 points.

Basically, the label did nothing. It was as if the participants were already so used to talking to tech, or the chatbot was so convincing, that a simple "I am a robot" sticker didn't make them suspicious. They still let the robot talk them into new ideas. The researchers found that even without the label, almost everyone (98–99%) already knew they were talking to a machine. So, the label wasn't giving them any new information to help them guard their minds.

The "Full Truth" Stopped the Magic
Then came the group with the "Full Truth" disclosure. When they were told, "This bot is trying to change your mind," the magic spell broke.

The opinion shift in this group dropped to just 6.3 points.

This is a huge difference. The disclosure cut the chatbot's persuasive power roughly in half. It didn't just make people slightly more skeptical; it fundamentally changed how they reacted.

Mental Shields Up: Participants in this group realized they were being manipulated. They started "counter-arguing" (fighting back in their heads) and rated the chatbot as much colder and more manipulative.
Anger at the Campaigner: Interestingly, they didn't get angry at the chatbot itself (the robot didn't make them mad), but they did get angry at the campaign behind it. They thought the methods were unacceptable and supported stronger penalties against the organization running the chat.

What This Means

The study suggests that current rules, which focus on telling people "This is AI," might be missing the point. It's like putting a "Made in China" label on a magic trick; it tells you where the object came from, but it doesn't tell you that you're being tricked.

The paper argues that for transparency to actually work, it needs to be meaningful. It's not enough to say what the system is (a robot); you have to say what it is trying to do (persuade you). When people know the game is rigged—that the other player is trying to win their vote or change their mind—they stop playing along so easily.

However, the authors also note a trade-off. While this "Full Truth" approach protects people from being swayed, it also makes the campaign look bad. If governments or organizations want to use AI to educate people or build support for a cause, they might find that once people know the AI's goal, they become much harder to convince. The AI becomes less effective, not because it's less smart, but because the audience has put their guard up.

In short, if you want to stop a robot from talking you into something, don't just tell them it's a robot. Tell them the robot is trying to change their mind. That's the only thing that really works.

Technical Summary: Toward Meaningful Transparency for AI Chatbots

Problem Statement
The increasing deployment of AI-generated content and AI-enabled systems in public communication has prompted regulators, such as the EU via the AI Act (Regulation (EU) 2024/1689), to mandate clear disclosures of AI involvement. Specifically, Article 50 requires providers of AI systems interacting directly with people (e.g., chatbots) to inform users of the system's identity. However, recent empirical studies suggest that disclosing an AI's identity has little to no average effect on attitude change or persuasion. This creates a regulatory uncertainty: if standard identity labels fail to mitigate the influence of persuasive AI, regulators must determine whether current disclosure frameworks are sufficient to protect users from automated influence campaigns. The core question is whether transparency about what a system is (its identity) is effective, or if transparency about what the system is trying to do (its intent) is necessary.

Methodology
The authors conducted a preregistered, three-arm experiment with a sample of 1,500 UK adults recruited via Prolific. The study utilized 60 policy issues previously identified in Hackenburg et al. (2025) as having strong persuasion effects.

Experimental Design: Participants engaged in a short, two-to-six-turn conversation with a chatbot (powered by gpt-5.6-terra) arguing for a randomly assigned policy stance. The chatbot used a specific prompt designed to persuade using facts and evidence while concealing its goal. The chatbot's behavior and the server setup were identical across all conditions; only the disclosure information varied.
Conditions:
1. Control: No disclosure.
2. Treatment 1 (T1): An "AI-identity" disclosure. Participants saw a prominent card and a persistent banner stating they were interacting with an AI chatbot, following the logic of Article 50(1) of the EU AI Act.
3. Treatment 2 (T2): An "intent" disclosure. Participants saw the same T1 label, plus additional text revealing the chatbot's persuasive intent, the specific stance it was instructed to argue for, and the instruction to conceal this goal. This approach draws on the logic of Regulation (EU) 2024/900 regarding political advertising.
Measures: The primary outcome was the shift in attitude on a 0–100 scale. Secondary outcomes included warmth toward the chatbot, perceived manipulation, persuasion knowledge, counterarguing, anger, and evaluations of the campaign's acceptability and the sponsor (penalty support).
Analysis: The study employed linear mixed models with random intercepts for policy issues. Equivalence tests (TOST) were used to determine if the AI label (T1) had effects indistinguishable from zero within preregistered bounds (±3.7 points for persuasion).

Key Results

AI-Identity Disclosure (T1) is Ineffective: The disclosure that the user was interacting with an AI did not meaningfully reduce persuasion. The attitude shift in the T1 group was 13.1 points, compared to 12.6 points in the control group. Equivalence tests confirmed that the difference between T1 and the control was statistically negligible (within the ±3.7 point bound). Furthermore, the label did not significantly alter warmth toward the chatbot or perceived manipulation.
Intent Disclosure (T2) Reduces Persuasion: Revealing the chatbot's persuasive intent and instructions cut the persuasive effect roughly in half. The attitude shift in the T2 group was 6.3 points, a reduction of approximately 6.8 points compared to both the control and T1 groups ( $p < .001$ ).
Mechanisms of Change: The intent disclosure significantly increased "persuasion knowledge" (awareness that an attempt to persuade was occurring). Participants in T2 rated the conversation as more manipulative, engaged in more counterarguing, and felt colder toward the chatbot. Notably, anger levels did not significantly increase.
Downstream Effects: The intent disclosure led participants to view the campaign's methods as less acceptable and increased their support for penalties against the sponsoring organization. The AI-identity label alone did not trigger these negative evaluations of the sponsor.
Robustness: The reduction in persuasion by T2 was consistent across all 60 policy issues and held regardless of participants' initial policy support. Robustness checks excluding participants flagged for potential AI-generated responses confirmed the main findings.

Significance and Claims
The paper argues that while current regulatory frameworks (like the EU AI Act) focus on disclosing the source of content (identity), this approach is insufficient for mitigating the influence of persuasive AI. The study demonstrates that "formal transparency" (labeling a system as AI) does not equate to "meaningful transparency."

The authors claim that transparency is effective only when it reveals the purpose and instructions behind the interaction—specifically, that the system is attempting to persuade and is doing so under specific constraints. This aligns with the Persuasion Knowledge Model, suggesting that users require information about the intent of the persuasion attempt to activate critical processing.

The paper concludes that for direct chatbot interactions, regulation must evolve beyond simple identity labeling. To effectively help users respond to automated influence, regulations should mandate disclosures that explain what the system is trying to do, not merely what the system is. The authors note a trade-off: while intent disclosures reduce the efficacy of AI persuasion, they also generate negative judgments about the actors behind the communication, which could impact the cost-effectiveness of using AI for public information campaigns.

The Great Chatbot Experiment

The Results: The Label vs. The Truth

What This Means

Technical Summary: Toward Meaningful Transparency for AI Chatbots

More like this