BrokenCtrL Editorial · 30 August 2026 · Label: HAC — human-approved, AI-assisted.
On 2 August 2026, Article 50 of the EU AI Act began requiring deepfakes to carry a label saying they are artificial (European Commission, Article 50 transparency FAQ) — seven months after the only preregistered experiments on the question found that a labelled deepfake keeps persuading a majority of the people who watch it (Communications Psychology, 2 January 2026). The legislation and the evidence are pointing in opposite directions, and both are public documents. This is an instance of a pattern that will recur as AI law hardens: regulating the disclosure of a harm instead of the harm, because disclosure is what legislators can draft.
What the Bristol Experiments Actually Found
Simon Clark and Stephan Lewandowsky of the University of Bristol’s School of Psychological Science ran three preregistered experiments (N = 175, 275, 223) in which participants watched a deepfake video of a man confessing, under varying levels of warning that the footage was fake (Verified — the paper, Communications Psychology, 2 January 2026). The design matters: preregistration means the hypotheses and analyses were locked before data collection, which removes the usual escape hatch of results massaged after the fact. These are, at the time of writing, the only preregistered experiments testing what a transparency warning does to a deepfake’s persuasive effect.
The warning did not do what the law assumes it does. Participants told in advance that the video was fake still rated the depicted man’s guilt at +0.87 on the study’s scale, against −0.40 for the control group that never saw the fake (Verified). In the condition with the most specific warning, 53.3% of participants judged the man’s guilt from the content of a video they had been told was fabricated (Verified).
The fake outperformed its own disclosure.
What the Law Assumes
Article 50’s deepfake provisions rest on a single mechanism: deployers must disclose that content has been artificially generated or manipulated, and the disclosure is what protects the viewer. The obligation applies since 2 August 2026, carries fines up to €15 million or 3% of worldwide turnover, and is supported by a Commission Code of Practice on marking and labelling (Verified — Commission publications). The same assumption runs through the platform policies that pre-date the Act: C2PA content credentials, YouTube’s altered-content labels, Meta’s “AI info” tags. Label it, and the deception is handled.
The strongest defence of the Article is worth stating properly. Transparency obligations were never advertised as a complete remedy; they serve autonomy, provenance and enforcement — a labelled fake can be traced, reported and prosecuted, and an unlabelled one now carries liability on its own, deception or not. On this reading, the Bristol results attack a claim the law never made.
The defence holds right up until you read what the label is for. The Commission’s own guidance frames Article 50’s purpose as protecting people from being deceived by synthetic content — the label exists so the viewer discounts the fake. That discounting is precisely the mechanism Clark and Lewandowsky tested, under conditions more favourable than any platform will ever achieve: the warning was explicit, immediate and attached to the specific video, not buried in a corner icon. The law’s enforcement machinery may be sound. The psychological premise underneath it is the thing the experiments measured, and it failed.
A newsletter that circulated the study, The Slow AI‘s “Epistemic Crisis” (27 August 2026), claimed the label “halves the damage” (Probable — the newsletter, cited here only as the pointer that led us to the paper). The paper says something worse. The influence is not halved into safety; it substantially survives.
The Majority That Knew, and Judged Anyway
The third experiment contains the number that should be pinned above every desk where deepfake policy is drafted. Among participants who explicitly accepted the warning — people who stated they believed the video was fake — 53.8% still relied on that video’s content to judge the man’s guilt (Verified). Not the inattentive, not the credulous: the successfully warned. The label reached them, they believed it, and the fake persuaded them anyway.
The authors draw the conclusion the legislation avoids: transparency “is insufficient to entirely negate the influence of deepfake videos,” and “further measures, such as removing or prohibiting deepfake content, should therefore be considered” (Verified — the paper’s discussion section). The people who produced the only preregistered evidence on the mechanism are saying the mechanism does not carry the weight the law puts on it.
None of this required hindsight. The paper was published 2 January 2026; the obligation took effect 2 August 2026. For seven months the evidence and the implementation timeline ran side by side, and the implementation did not blink. Content generated and published before 2 August does not even need retroactive labels (Verified — Commission FAQ), so the archive of existing fakes sits outside the remedy entirely, while the remedy itself is one a majority of warned viewers walk straight through.
The verdict is short. Europe’s deepfake protection is a label, and the only preregistered evidence on labels says the deepfake wins anyway.
