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

The paper Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion by Adrian Rauchfleisch and Andreas Jungherr investigates whether different types of disclosure reduce the persuasive influence of AI chatbots. The study was motivated by the EU AI Act's Article 50, which requires that users be informed when interacting with an AI system, and by prior research suggesting that such AI-identity disclosures have little effect on attitude change.

The authors conducted a preregistered three-arm experiment with 1,500 UK adults recruited via Prolific. Participants had a short conversation with a persuasive chatbot (gpt-5.6-terra) about one of 60 policy issues. The chatbot was identical for everyone, and only the disclosure varied: a control group received no disclosure, a first treatment group (T1) received a prominent AI-generated content label, and a second treatment group (T2) received the same label plus a disclosure of the chatbot's persuasive intent and its verbatim instructions.

The key results are as follows:

• The chatbot shifted attitudes by 12.6 points on a 100-point scale in the control group.

• The AI-identity disclosure (T1) was practically equivalent to no disclosure, with a 13.1-point shift. The preregistered equivalence test ruled out effects larger than ±3.7 points (d = 0.15; pTOST =.002). The label also did not significantly increase persuasion knowledge (b = −0.14, 95% CI [−0.32, 0.03], p =.111).

• The additional intent disclosure (T2) cut the persuasive effect roughly in half to 6.3 points. Compared with T1, it reduced persuasion by 6.83 points (95% CI [−9.26, −4.40], p <.001). The contrast with control was nearly identical (b = −6.77, 95% CI [−9.19, −4.36], p <.001; d = 0.25).

The intent disclosure also changed participants' evaluations of the interaction and the campaign behind the chatbot. Compared with T1, participants in T2 rated the conversation as more manipulative (b = 0.62 on a 7-point scale, 95% CI [0.41, 0.83], p <.001), counterargued more (b = 0.49, 95% CI [0.31, 0.67], p <.001), and rated the chatbot 5.02 points colder on a 0–100 feeling thermometer (95% CI [−7.78, −2.26], p <.001). However, anger did not rise significantly (b = 0.13, 95% CI [−0.02, 0.27], p =.083). The largest shift appeared for persuasion knowledge (b = 0.71 on a 7-point scale, 95% CI [0.53, 0.89], p <.001, d = 0.49; exploratory), indicating that participants were more aware that someone had tried to persuade them.

The disclosure also influenced evaluations of the campaign: participants in T2 viewed the campaign's methods as less acceptable (b = −0.34, 95% CI [−0.51, −0.17], p <.001) and supported stronger penalties against it (b = 0.35, 95% CI [0.20, 0.51], p <.001). The AI-source disclosure alone did not lead to such a penalty.

The authors explain why the source label fails: 98–99% of participants in every arm, including the control, correctly identified that they had talked to an AI chatbot. Additionally, 29.5% of control participants falsely remembered having seen an AI label. The chatbot's polished, information-rich style was likely a clear indicator that it was not human, so the label disclosed little new information.

The paper concludes that transparency works when it reveals what the system is trying to do, not merely what the system is. The authors argue that current regulatory efforts, which focus on content authenticity and factuality, may be the wrong template for interactive AI chatbots. They suggest that the logic of Regulation (EU) 2024/900, which requires disclosure of persuasive intent and targeting logic in political advertising, might provide a more promising template for regulating persuasive AI.

The authors also note several limitations: the intent disclosure is a package of elements (intent, method, and concealment instruction), so the design cannot isolate the contribution of each; the effect may weaken with repeated exposure; the causal chain cannot be formally established because some measures were assessed after the attitude outcome; and a purpose disclosure depends on what the deployer reports, arguing for making such disclosures binding and auditable.

Improvements to AI Systems:

1. Implement intent-based disclosure mechanisms in chatbot interfaces. Instead of only labeling content as AI-generated, the system should explicitly state its persuasive goal (e.g., I am designed to change your opinion on [topic]) before or during the interaction. This can reduce attitude shifts by up to 50% compared to no disclosure.

2. Add a persuasion awareness module that tracks user counterarguing and perceived manipulativeness in real time. If these metrics rise above a threshold, the system can dynamically adjust its tone or offer a neutral mode to reduce unintended influence.

3. Develop a purpose audit trail for AI deployers. The system should log the exact persuasive instructions, target outcomes, and user-facing disclosures, making them machine-readable and verifiable by third parties. This enables regulators to audit compliance with transparency laws (e.g., EU AI Act, Regulation 2024/900).

4. Create a disclosure effectiveness predictor that estimates whether a given disclosure will actually increase persuasion knowledge. The system can test different disclosure phrasings (e.g., I am trying to persuade you vs. This is AI-generated) and select the one with the highest predicted counterarguing, based on user interaction patterns.

5. Integrate a post-interaction debrief feature that asks users if they felt persuaded and shows them the chatbot's original instructions. This reinforces transparency and helps users calibrate trust for future interactions.

6. Build a coldness calibration control for persuasive chatbots. Since users rated the chatbot as colder when intent was disclosed, the system can adjust its warmth or empathy to avoid backlash while still maintaining transparency—ensuring the disclosure doesn’t harm user experience unnecessarily.

What the Improved AI System Can Do:

• Reduce its own persuasive impact by up to 50% when required, simply by disclosing its intent, without needing to change its argumentation style.

• Self-monitor and report whether its disclosures are actually increasing user awareness (e.g., via user feedback or behavioral proxies like counterarguing), and adapt if they are not.

• Provide regulators and users with auditable proof of what the AI was instructed to do, what it disclosed, and how users responded—enabling meaningful oversight.

• Offer a transparency mode for sensitive domains (politics, health, finance) where the system proactively reveals its persuasive design and allows users to opt into a neutral, non-persuasive version.

• Generate personalized disclosure messages that maximize persuasion knowledge without triggering anger, based on user traits (e.g., prior suspicion, cognitive style).

The growing role of AI-generated content and AI-enabled systems in public communication has led regulators to demand clear disclosure of content provenance and AI involvement. But the effects of such disclosures remain uncertain. We test two disclosure approaches in their impact on an AI chatbot's persuasive appeal. In a preregistered experiment, 1,500 UK adults held a short conversation with a persuasive chatbot about one of 60 policy issues. The chatbot was identical for everyone. We randomized the disclosure that people received: nothing (control), a prominent disclosure that they were interacting with an AI (T1), or that disclosure plus the chatbot's persuasive intent and instructions (T2). The chatbot shifted attitudes by 12.6 points on a 100-point scale in the control group. The AI-identity disclosure was practically equivalent to no disclosure, with a 13.1-point shift, whereas the additional intent disclosure cut the persuasive effect roughly in half to 6.3 points. It also made participants view the campaign's methods as less acceptable and support stronger penalties against it. For direct chatbot interactions, transparency about AI identity alone does not meaningfully impact its influence. While current rules emphasize what a system is, our results show why the regulation of persuasive AI must also address what the system is trying to do.