Research opportunity · Persuasion, Intermediation & the Information Order

Do AI-content labels change belief and sharing on a live social platform?

Question

On a live, open social platform, when labels saying that a post is AI-generated, or that it is AI-generated and potentially misleading, are applied at random to real political and news posts in users' feeds, how do they change actual reposts, likes and replies, and users' belief in the labelled claims, compared with unlabelled posts?

Wrong or missing validation targetOne PhD chapter or paperNeeds a budget (medium or higher)Partly addressed

Why it matters

EU transparency rules now require labelling of deepfakes and some AI-generated public-interest text, but evidence on whether labels work comes almost entirely from survey experiments with stated sharing intentions. The one study found with real platform data measures engagement on entertainment posts, not belief or political content. Before labelling rules are judged effective or useless, labels need testing where people actually scroll, share and forget.

The debate lists this among what we still do not know: natural-field effects of AI labelling

Evidence that motivates it

Label effects are context-dependent: one 2026 experiment found simple AI-authorship labels did not reduce persuasion of policy messages, while experiments on misleading synthetic images found warning labels reduced belief.
Label effects are context-dependent (simple AI-authorship labels did not reduce persuasion of policy messages; warning labels reduced belief in misleading synthetic images); field trials are named as what would change it, and it licenses neither "labels work" nor "labels fail".
EU AI Act Article 50 transparency obligations applying from 2 August 2026 require disclosure for chatbot interaction and clear labelling in key cases including deepfakes and AI-generated/manipulated public-interest text; generative AI systems already on the market before that date have until 2 December 2026 to meet the Article 50(2) machine-readable marking obligation; the voluntary Code provides compliance guidance.
EU AI Act Article 50 labelling obligations apply from 2 August 2026; the claim explicitly does not license any conclusion about their effectiveness.

A possible design

Use Bluesky's open moderation layer, where independent labelling services attach labels to posts and each user chooses which services to subscribe to and whether each label hides, warns or is ignored. Recruit consenting users (about 1,500), randomly assign each to one of three research labelling services (one that applies no labels, an "AI-generated" label, or an "AI-generated, may mislead" label), ask them to subscribe to it with a fixed "warn" display setting, and install a companion survey app; because subscription is the user's own choice, verify compliance and analyse by assignment. Identify AI-generated political or news images and text in their feeds with a detector plus human confirmation; each service labels the same confirmed posts in its own wording, and a random half of eligible posts are left unlabelled by every service for a within-person comparison. A labelling service is an ordinary account that anyone can find and subscribe to, so its labels also reach non-participant subscribers; state the research purpose on each service's profile, monitor subscriber lists and check for contamination. Measure real reposts, likes, replies and dwell (where available through the client), and use short in-app prompts a few hours later to measure belief in a random subset of labelled and unlabelled claims. Compare effects by label wording, by whether the content is actually misleading, and by user partisanship. Contrast results with survey-experiment estimates for the same items.

Data needed

Three Bluesky labelling services and a custom client or feed for outcome logging; an AI-content detector with human verification; participant recruitment and payment; ethics approval covering labels shown on real content.

Feasibility

One PhD chapter or paper. Medium: hosting and detection a few hundred US dollars per month; participant payments about US$10,000 to 25,000, which needs a grant or project budget. Engineering effort is substantial but the platform's tools are open. Bluesky's users are not representative of social-media users generally, so the study shows whether labels change real behaviour, not population averages. Mislabelling real content as AI is an ethical risk; restrict labels to confirmed cases and debrief participants.

Has this been done?

Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).

Survey experiments show labels reduce belief in misleading AI images (Wittenberg et al.; Höltervennhoff, Ricker et al., CHI 2026, who also find overall discernment falls) and do not reduce persuasion of AI-written policy messages (Gallegos et al.); online vignette experiments show labels reduce engagement (Seeger et al.). A Canadian experiment injected labelled AI posts into participants' own Facebook feeds in a single session and found small labels ineffective and no effect of any label on accuracy judgements or sharing (The Dais, 2025). Real platform data exist only for engagement with entertainment posts after TikTok's disclosure policy (Carney et al.), without randomization or belief outcomes, and a methods paper proposes researcher-run field experiments on Bluesky without results. A randomized field test of AI labels on organically occurring political or news posts on a live platform that measures both real sharing and belief was not found.

How strong this evidence is

Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
What was read: 2 of the 8 nearest works below were judged from an abstract, summary or search listing only, and 1 has details that could not be confirmed; each is marked. Not every entry says how much of the work was read.
Where the search could not reach: Screened by web search on the date shown. The journal full text of the Gallegos et al. study was not opened; its findings were read from the arXiv preprint, and the journal version was confirmed through its DOI record. Platform-internal label tests (for example by Meta, TikTok or YouTube) are not public. Google Scholar was not searched. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: The Taylor & Francis article 'How Users Perceive and React to Labeled AI-Generated Content on Social Media: A Longitudinal Study of Cognitive and Emotional Effects' (International Journal of Human-Computer Interaction, 2026, doi:10.1080/10447318.2026.2618553) appeared in results but could not be opened; its design (platform or simulated) is unknown and should be checked. Platform-internal label tests are not public. Google Scholar was not searched. Absence of a hit is not proof that the work does not exist.
Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.

Nearest work found

Wittenberg, Epstein, Péloquin-Skulski, Berinsky & Rand, "Labeling AI-generated media online," PNAS Nexus 4(6), pgaf170 (28 May 2025)
Two preregistered US survey experiments (n = 7,579) with misleading AI images; labels overall reduced belief, and a plain "AI-Generated" label did not significantly reduce stated sharing. Sharing is self-reported, not observed on a platform.
Gallegos, Shani, Shi, Bianchi, Gainsburg, Jurafsky & Willer, "Labeling messages as AI-generated does not reduce their persuasive effects," PNAS Nexus 5(2), pgag008 (10 Feb 2026); preprint arXiv:2504.09865
US survey experiment (n = 1,601): AI-model, human-expert or no label on AI-written policy messages; labels did not change attitude change, perceived accuracy or willingness to share. Survey setting only; findings read from the arXiv preprint, journal details confirmed through its DOI record.
Carney, Riveros & Tully, "Made With AI: Consumer Engagement With Social Media Containing AI Disclosures," Journal of Consumer Research, advance article ucag013 (7 May 2026)
About 1.1 million US TikTok posts after the platform's 2023 disclosure policy, with 7–8% fewer likes for disclosed posts conditional on views, plus eight experiments. Observational platform data on creator content; no belief or misinformation outcomes and no political content focus.
Condom-Tibau, Puccetti, Bacciu, Abrate & Cresci, "Open Platform Field Experiments: Expanding the Design Space of Experimental Research on Social Media," arXiv:2609.21608 (18 Sep 2026)
Proposes researcher-run interventions on Bluesky components, including moderation services, and maps the architecture and governance; the abstract reports no experiment on labels.
Seeger, Wessel & Lehrer, "AI content labeling and user engagement on social media: The role of AI level, content type, and disclosure timing," Electronic Markets 36, article 31 (2026)
Two online vignette experiments (n = 325 and 371) on simulated Instagram profiles; AI labels reduced engagement, more for emotional content. Simulated feed and stated engagement only.
Lockhart & Tessono (The Dais, Toronto Metropolitan University), "Human or AI? Evaluating Labels on AI-Generated Social Media Content," report (March/May 2025)
Randomized experiment with 2,472 Canadians (Leger panel, Sept-Nov 2024, English and French): a Who Targets Me browser extension injected AI-generated posts into the first 50 posts of each participant's own Facebook feed with no label, a small disclaimer or a full-screen blocking label. Small labels had no effect; full-screen labels reduced views; neither label changed accuracy judgements or sharing. Closest to an in-feed test, but the posts are researcher-injected, exposure is one session, and it is a policy report, not peer reviewed; whether sharing was observed or stated was not confirmed.
Höltervennhoff, Ricker et al., "'That's another doom I haven't thought about': A User Study on AI Labels as a Safeguard Against Image-Based Misinformation," CHI 2026 Judged from an abstract, summary or search listing onlySome details not confirmed
Focus groups plus an online survey of more than 1,300 US and European participants with simulated posts crossing AI versus real images and true versus false text: labels reduced belief in false AI content but unlabeled false content was believed more and labeled true content doubted, so overall discernment fell. Simulated posts, no sharing outcome reported. Read from the CISPA news page; full author list not confirmed.
"Labeling Synthetic Content: User Perceptions of Label Designs for AI-Generated Content on Social Media," CHI 2025; arXiv:2503.05711 Judged from an abstract, summary or search listing only
Seen in search results only (title, venue, identifier; authors not checked). A user-perception study of label designs for AI-generated content; not a live-platform field experiment.

Literature check

Stands as first published. The closest new work is a Canadian experiment by The Dais (Lockhart & Tessono, 2025) that injected AI-generated posts into participants' own Facebook feeds through a browser extension and randomized no label, a small disclaimer or a full-screen label (n = 2,472): small labels had no effect, full-screen labels cut views, and neither changed accuracy judgements or sharing. That is a step toward in-feed testing but uses researcher-injected posts in a single survey session, not labels on organically occurring political or news posts with real reposts on a live platform. A CHI 2026 survey study (Höltervennhoff, Ricker et al.) finds labels reduce belief in false AI images but worsen overall discernment, again with simulated posts. No randomized field test on a live platform measuring both real sharing and belief was found.

Searched: web (extended); arxiv.org; osf.io; ssrn.com; academic.oup.com; pnas.org; nature.com; dl.acm.org; journals.sagepub.com; misinforeview.hks.harvard.edu; dais.ca; cispa.de; German-language web. Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.

The screen before this check

Partly done already: related work covers some of this question (first-pass screen of 12 October 2026).

Survey experiments show labels reduce belief in misleading AI images (Wittenberg et al.) and do not reduce persuasion of AI-written policy messages (Gallegos et al.); online vignette experiments show labels reduce engagement (Seeger et al.). Real platform data exist only for engagement with entertainment posts after TikTok's disclosure policy (Carney et al.), without randomization or belief outcomes. A methods paper proposes researcher-run field experiments on Bluesky but no results were found. A randomized field test of AI labels on political or news content that measures both real sharing and belief was not found.

Search queries used

field experiment AI-generated content labels social media platform real users sharing engagement effect
AI authorship label persuasion policy messages experiment 2026 labels do not reduce persuasion
EU AI Act Article 50 labelling effectiveness study deepfake disclosure evaluation natural setting
Bluesky labeler field experiment warning labels misinformation OR AI-generated randomized
browser extension field experiment warning labels AI-generated images social media feed belief sharing real-world
Meta "AI info" label OR "Made with AI" label natural experiment rollout effect engagement causal study

Literature check, 13 October 2026:

Bluesky labeler randomized field experiment labels users engagement belief [allowed_domains: arxiv.org, osf.io, ssrn.com, papers.ssrn.com]
field experiment AI-generated label real social media users randomized reposts belief political content 2026
AI disclosure label effect sharing misinformation observational platform data X Community Notes AI-generated image note [allowed_domains: academic.oup.com, pnas.org, nature.com, dl.acm.org, journals.sagepub.com, misinforeview.hks.harvard.edu]
"Open Platform Field Experiments" Bluesky OR "Labeling AI-generated media online" Wittenberg cited field study real feed AI labels
in-feed browser extension experiment real Twitter X Facebook feeds AI-generated content warning label sharing behavior participants own accounts
Kennzeichnung KI-generierter Inhalte soziale Medien Feldexperiment Wirkung Teilen Glaubwürdigkeit Studie

Provenance

RO-LD10-05 v3. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-10 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.

Changes

v2, 13 October 2026: Literature check added: a second, deeper search for existing work, with its queries, new nearest work and limits. The question and verdict stand. See v1 as published.
v3, 13 October 2026: Cost band corrected from "low to medium" to "medium": participant payments of about US$10,000 to 25,000 need a grant or project budget, in line with other opportunities of that size. Found in the independent review of the opportunity-page fixes. The question, design, literature check and verdict are unchanged. See v2 as published.

All research opportunities