Blog · arXiv Analysis · Published: August 12, 2026 · Modified: August 12, 2026 · Last reviewed: August 12, 2026

The Persuasive Intent Becomes the Campaign Receipt

An AI-generated label names the medium. It does not tell a person who commissioned the conversation, which belief it is meant to change, or which instructions organize the attempt.

A new experiment makes that informational difference measurable. Its result supports a practical governance proposal: treat persuasive-purpose disclosure as a verifiable campaign receipt, not a decorative badge.

The Paper

The source is Adrian Rauchfleisch and Andreas Jungherr's Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion, arXiv:2608.11794v1 [cs.CY], cross-listed in cs.AI and cs.HC, submitted August 12, 2026. It reports a preregistered experiment fielded from July 26 through 31 with 1,500 UK adults recruited through Prolific using quotas for sex, age, and party.

Three Arms, One Persuader

Each participant first rated one of 60 UK policy stances, then held a two-to-six-turn conversation with gpt-5.6-terra arguing for that stance, and finally rated it again. The issues were drawn from a prior study and selected for relatively high persuadability. The three-arm design compared no disclosure, a prominent AI-identity notice before and during the chat, and that same notice plus the actual target stance, persuasion method, and instruction to conceal the persuasive goal.

The chatbot and prompt were otherwise identical. Crucially, the server never received the assigned arm, so the model could not change its behavior in response to the notice. The comparison therefore isolates information shown to the participant rather than a different persuader. The authors call the third arm an intent disclosure, but it is a package: goal, method, and concealment instruction arrive together.

The Identity Label Added Little

The raw mean attitude shift was 12.6 points on the 0–100 scale without a notice and 13.1 points with the AI-identity notice. In the preregistered adjusted model, the identity-notice difference was 0.06 points, with a 90% confidence interval from −1.97 to 2.08. The interval fell inside the study's prespecified equivalence bounds of ±3.7 points, so this experiment supports a bounded conclusion: any effect of that notice in this design was smaller than the margin the researchers defined as meaningful. It does not prove that every AI label is behaviorally inert.

The recall checks make a simple visibility explanation unlikely. Of participants shown the identity notice, 97.8 percent recalled it. Yet 98–99 percent in every arm already identified their partner as an AI chatbot, including those shown no notice; 29.5 percent of the control group even misremembered seeing a label. Here, the badge mostly repeated what people thought they knew.

Purpose Changed the Encounter

The additional purpose-and-instructions package produced a different result. Its raw mean attitude shift was 6.3 points. Relative to the identity-only arm, the adjusted difference was −6.83 points, with a 95% confidence interval from −9.26 to −4.40; relative to control it was −6.77 points. In this experiment, the measured shift was therefore roughly halved, not eliminated.

The other outcomes help describe the response without establishing a causal mechanism. Compared with identity disclosure alone, the package increased reported counterarguing and perceived manipulation, reduced warmth toward the chatbot, lowered acceptance of the campaign's methods, and increased support for penalties against its sponsor. The exploratory persuasion-knowledge measure also rose. Anger did not rise significantly, and because these measures followed the attitude outcome, they cannot formally mediate it. Intent-notice recall was 79.1 percent.

The Campaign Receipt

The governance lesson is not simply to make labels longer. Identity answers, “What am I talking to?” A campaign receipt should answer, “Whose organized attempt is this, what change does it seek, and how is it operating?” That is the informational move tested here.

This essay proposes that a receipt for a persuasive chatbot name the sponsor and any controlling organization; the target stance or behavior; the audience and outreach channel; the model and deployment version; the governing persuasive method; any instruction to hide or soften the purpose; the data used for personalization; the campaign period; and a stable identifier for the underlying instruction record. A brief notice can link to the complete, versioned record.

A Receipt Must Be Auditable

A deployer should not be the final judge of its own disclosure. The receipt needs a change log, retention rule, named custodian, and an independent route for comparing the public summary with the instructions actually served. Sensitive prompts need not be dumped into the open; a qualified auditor can inspect protected material while the public record preserves the sponsor, objective, methods, targeting categories, and version history. If the objective or prompt changes, the receipt changes too.

This distinguishes disclosure from absolution. A complete receipt does not make covert targeting fair, an unsupported claim true, or a manipulative campaign acceptable. It creates an object that participants, researchers, regulators, and sponsors can contest after the interface disappears.

The Regulatory Boundary

Article 50(1) of the EU AI Act requires providers of certain systems intended to interact directly with people to ensure that those people are informed they are interacting with AI, unless that is obvious, subject to stated exceptions. The paper describes its identity treatment as testing that disclosure logic; it is not a legal compliance audit.

The authors use Regulation (EU) 2024/900 on political advertising as a different transparency template. That regulation includes sponsor information and, where applicable, information about targeting or ad-delivery techniques. It does not govern AI chatbots as a class. The campaign receipt is this essay's extension by analogy, not a statement of current legal requirements or legal advice.

Limits That Stay Attached

The evidence is narrower than the proposal. The study's limits include one model, one strong evidence-oriented persuasion prompt, short conversations, a UK Prolific sample, and issues intentionally selected because persuasion was likely. Participants had to complete at least two turns, so the design could not measure whether a disclosure changes the initial choice to engage. Repeated exposure and habituation were not tested.

Most importantly, the experiment cannot separate disclosure of the goal from disclosure of the method or the concealment instruction. It also cannot guarantee truthful self-reporting by a real sponsor. The measured result is evidence that this bundled information mattered under these conditions, not a settled effect size for other populations, models, interfaces, issues, or campaigns.

Artifact Audit

For this review, the arXiv record, version-one PDF, experimental HTML, preregistration, and linked data-and-code project were checked. The paper reports University of Bamberg ethics approval, informed consent and debriefing, mixed models with policy-issue random intercepts, Holm correction within preregistered families, and two one-sided equivalence tests with prespecified bounds. This page reports confirmatory and exploratory findings separately and reproduces no table, figure, prompt, or extended passage.

Sources

Return to Blog