AI Disclosure and Political Persuasion: Labelled AI Still Persuades: Disclosure doesn't disarm the message

In brief

An AI authorship label does not measurably reduce how much a political message shifts people's policy views. In two experiments, including one with 1,601 Americans, support moved about the same whether the message was credited to an AI, a human expert, or nobody. The label does bite on belief: it lowers how accurate people judge flagged claims and content to be, and the strongest evidence comes from false or misleading images. The null on attitude change is a limit of the designs, not proof that no effect exists, and in the main study the label described an expert AI model rather than a bare badge.

How to use this

Treat an AI label as a credibility tool, not a way to disarm a political argument. Before relying on a label to blunt a message's pull on policy views, test the message's factual claims instead: that is where labels measurably move belief. When you design a label, write the wording deliberately, because readers infer full automation from a badge that may only mean a touch-up, and reword it when that inference misfires. Check the outcome you actually care about. If it is belief in a specific claim, expect a small effect; if it is attitude change, expect none within the range these studies could detect, and do not claim more. Remember that forced exposure is not scrolling, so a labelling rule still has to work where people can skip.

What the story is about

A label saying a political message was written by AI is meant to tell readers something about where the message came from. Whether that makes the message less persuasive depends on which response you mean. A message can shift what you think the government should do about something. It can also change whether you believe the specific claims inside it, or how far you trust whoever sent it. Those are separate effects, and a label can move one without moving the other.

A large pre-registered survey experiment on this question, with 1,601 Americans, recruited on the online platform Prolific and fielded in 2024 (Gallegos et al., 2026). Everyone read the same policy message, on one of four topics. The message was said to come from an expert AI model trained in US policy, from a policy expert trained in US policy, or it carried no attribution at all. On a 0 to 100 scale of support for the policy, the message worked: support moved 9.74 points. The authorship label made no measurable difference to that shift. The gap between the AI label and the human label was about one point on the same scale, too small to be statistically distinguishable from no difference at all. Between the AI label and no label at all it was 0.11 points. The labels also made no difference to how confident people were, how accurate they judged the message, or whether they intended to share it (Gallegos et al., 2026).

An independent team tested the same question with a nationally representative probability sample of 3,861 people (Wang, Sturgis and de Kadt, 2026). Support for the policy was again unaffected by the AI label. The same label did lower how accurate people judged the article to be, and it lowered their interest in the policy, neither of which the first study had found. The authors read their own results as limited in scope and dependent on context.

The clearest evidence that a label does something comes from false images. Two pre-registered experiments, with 7,579 Americans between them, tested labels on real social media posts whose images journalists or fact-checkers had already flagged as both AI-generated and misleading (Wittenberg et al., 2025). The first experiment used 14 posts, the second 29. A label of any kind lowered how much people believed the post's central claims and lowered the image's credibility. Labels that warned the content could mislead showed a larger tendency. Putting both cues in one label added nothing beyond the two cues (Wittenberg et al., 2025).

What differs from study to study is the outcome being measured. A label works on belief in a factual claim. It does not work on the change in attitude that comes from following an argument, which is what a policy appeal is built to produce. On attitude change, labels do nothing (Gallegos et al., 2026; Wang, Sturgis and de Kadt, 2026). On perceived accuracy, trust, willingness to read, sharing and clicks, they do something small to moderate (Altay and Gilardi, 2024; Toff and Simon, 2025; Zoizner et al., 2025; Wang, Jung and Bapna, 2025). On belief in a claim that fact-checkers have already flagged, they do something real (Wittenberg et al., 2025).

The main finding of no effect has a limit. The study could reliably detect a difference of about 3 points or more on the 0 to 100 scale, and nothing smaller than that (Gallegos et al., 2026). The differences it measured sat well inside that range, so the honest reading is that no effect was large enough for this design to see, not that no effect exists. A label effect of one or two points cannot be ruled out. The AI arm described information generated by an expert AI model trained in US policy. That is a specific and flattering description, not the bare badge a platform might attach. The authors chose that wording to mirror how developers describe their models, and they note it captures a general expert-AI label with no other source cues. That difference is the likeliest reason this finding sits alongside the accuracy penalties found elsewhere. One more limit applies to almost all of these forced-exposure attitude-change experiments. In almost all of them, participants could not scroll past. A labelling rule has to work in a world where people can skip.

In the experiments run so far, a label saying a political message was written by AI does not measurably reduce how much the message shifts people's policy views. It can lower belief in specific factual claims, especially claims that have already been flagged as false or misleading. It can also change how accurate, trustworthy or worth reading people judge the content to be.

So what

For anyone who writes or regulates political messages, the practical point is that an AI label is not a reliable way to reduce a message's persuasive power over policy attitudes. It can change what people believe about particular factual claims, and how far they trust the source. The authors of the Gallegos et al. study put their finding in those terms: the persuasive effect came out close to equivalent whether the message was labelled as AI-written, human-written or unlabelled. They add that disclosure policies may serve purposes their experiment did not test, such as helping people make informed choices and limiting the spread of AI-generated material (Gallegos et al., 2026).

For political parties

If you are running a campaign, the badge is not where your exposure sits; your factual claims are. Labelling an AI-written message does not measurably weaken its pull on policy views. Where the label bites is credibility. In two pre-registered experiments in the US and UK, with 4,976 people, labelling headlines as AI-generated lowered how accurate people thought they were and their willingness to share them. It did so whether the headline was true or false, and whether it had been written by a person or a machine. The AI label moved perceived accuracy by 2.66 percentage points. A label saying the headline was false moved it by 9.33 points (Altay and Gilardi, 2024). A claim your audience doubts is a bigger problem than a disclosure.

For government

For a regulator, the honest case for mandating AI labels is about informing the public. It is not a way to blunt the persuasive force of political messages. The evidence on AI disclosure comes from experiments plus one commercial field study of virtual influencers, a study that has not yet been peer-reviewed and that measured clicks and plays rather than votes (Wang, Jung and Bapna, 2025). Every label in these experiments was presented as platform or publisher information. What the evidence does suggest is that wording, and the inference readers draw from it, matter more than a label's presence (Altay and Gilardi, 2024; Meta, 2024).

Case studies

In April 2024 Meta moved from removing manipulated media to labelling a much wider range of video, audio and image content, triggered either by industry-standard AI indicators or by the poster's own disclosure. The company's stated reason was that "providing transparency and additional context is now the better way to address manipulated media and avoid the risk of unnecessarily restricting freedom of speech" (Meta, 2024). The badge was called Made with AI. On 1 July 2024 Meta renamed it AI info, because content carrying only minor AI modifications, such as retouching, was tripping the indicators and getting labelled as though a machine had made the whole thing. That is the inference problem from these experiments appearing at platform scale: readers and photographers read the badge as full automation when it often meant a touch-up (Meta, 2024). The platform's answer was to say more, not less.

References

Altay, S. and Gilardi, F. (2024) 'People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation', PNAS Nexus, 3(10), article pgae403. Available at: https://doi.org/10.1093/pnasnexus/pgae403 (Accessed: 10 September 2026).

Gallegos, I.O., Shani, C., Shi, W., Bianchi, F., Gainsburg, I., Jurafsky, D. and Willer, R. (2026) 'Labeling messages as AI-generated does not reduce their persuasive effects', PNAS Nexus, 5(2), article pgag008. Available at: https://doi.org/10.1093/pnasnexus/pgag008 (Accessed: 10 September 2026).

Meta (2024) Our approach to labeling AI-generated content and manipulated media, 5 April (updated 1 July 2024, 12 September 2024 and 23 October 2025). Available at: https://about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/ (Accessed: 10 September 2026).

Toff, B. and Simon, F.M. (2025) '"Or they could just not use it?": the dilemma of AI disclosure for audience trust in news', The International Journal of Press/Politics, 30(4), pp. 881–903. Available at: https://doi.org/10.1177/19401612241308697 (Accessed: 10 September 2026).

Wang, C., Sturgis, P. and de Kadt, D. (2026) 'AI labeling reduces the perceived accuracy of online content but has limited broader effects', Telematics and Informatics, 107, article 102416. Available at: https://doi.org/10.1016/j.tele.2026.102416 (Accessed: 10 September 2026).

Wang, M., Jung, J. and Bapna, R. (2025) The impact of realism and AI disclosure on virtual influencer effectiveness: a large field experiment. SSRN preprint. Available at: https://doi.org/10.2139/ssrn.5202553 (Accessed: 10 September 2026).

Wittenberg, C., Epstein, Z., Péloquin-Skulski, G., Berinsky, A.J. and Rand, D.G. (2025) 'Labeling AI-generated media online', PNAS Nexus, 4(6), article pgaf170. Available at: https://doi.org/10.1093/pnasnexus/pgaf170 (Accessed: 10 September 2026).

Zoizner, A. et al. (2025) 'Can AI-attributed news challenge partisan news selection? Evidence from a conjoint experiment', The International Journal of Press/Politics, 31(4), pp. 926–951. Available at: https://doi.org/10.1177/19401612251342679 (Accessed: 10 September 2026).