Europe's new AI labels could train the public to trust unlabeled deepfakes

Europe's rule requiring AI-generated images, audio, and video to carry visible labels took effect on August 2. Researchers studying how people react to synthetic media are already worried about an unintended consequence: a label meant to protect audiences could train them to treat content without a label as authentic, which is the exact signal a malicious deepfake maker wants to send.

The rule is Article 50 of the EU AI Act, and it has two layers. Article 50(2) requires providers of AI systems, including general-purpose models, that generate synthetic audio, image, video, or text to mark their outputs in a machine-readable format detectable as artificially generated or manipulated, with technical solutions that are effective, interoperable, robust, and reliable. Article 50(4) requires deployers of deepfake image, audio, and video systems to disclose that the content was AI-generated, and deployers manipulating text published to inform the public on matters of public interest to disclose AI generation unless the content underwent human review with a person or legal entity holding editorial responsibility. The European Commission has also published a set of official icons for the disclosure; using the EU icons is optional, but disclosure in line with the act is mandatory.

The mechanism researchers flag is the boomerang effect, a pattern that shows up in public-health, education, and now media-literacy research. The idea is straightforward: when an intervention becomes the rule, the visible marker that was supposed to protect people starts to invert. Audiences get used to seeing the label, the label becomes normal, and content without the label starts to feel like evidence of authenticity. A peer-reviewed study on AI content perception found that when people are shown sophisticated synthetic media, their ability to distinguish real from fake can be little better than a coin-toss, which means the cue people will actually lean on is the label, not their own judgment.

Research on deepfake-ad disclosure found that the presence and timing of AI disclosure shape how people respond to a piece of content, and a separate study on influencer misinformation found that social cues can compete with or override authenticity signals entirely. The boomerang argument pulls these threads together: if the law normalizes the icon and the watermark, then unlabeled content stops being "I don't know" and starts being "must be real", which is the opposite of what the rule is trying to do.

The rule does try to close that gap. Article 50(2)'s machine-readable marking obligation is meant to outlast the visible icon; a watermark embedded in the file should survive cropping, screenshot, and re-upload, and it should be detectable by tools the platforms themselves run. Non-compliant companies face fines up to EUR 15 million or up to 3% of total worldwide annual turnover (roughly USD 16-17 million at recent exchange rates, with the turnover cap applying to larger providers), and the law reaches extraterritorially to companies outside the EU whose AI-generated content is visible inside the union. Anthropic has told The Conversation it will ensure all Claude chatbot models launched in the EU after August 2 will support machine-readable marking, and is working to add watermarks to earlier Claude models, per The Conversation's summary of the company's compliance posture.

The rule also has narrow exceptions. Article 50 does not apply to AI systems authorised by law to detect, prevent, investigate, or prosecute criminal offences, subject to appropriate safeguards, and 50(2) does not apply where AI performs an assistive function for standard editing or does not substantially alter the input data or its semantics. Individuals creating content in a personal or non-professional capacity are carved out. Australia, by contrast, is taking a softer line and only recommends that AI developers mark AI-generated content.

None of that resolves the boomerang. Watermarks and platform checks are layered signals; the visible icon is the cue a casual viewer sees, and the cue shapes the default assumption. A label describes the labeled item. The absence of a label is not proof of real, and treating it that way is the trap the rule could end up training the public to fall into. Trust, going forward, comes from layered signals: a watermark, a platform check, a source you can verify, a context that fits. It does not come from the presence or absence of a single icon.