Europe is building an AI disclosure icon. A label needs evidence behind the click

THE SIGNAL IN ONE SENTENCE

A tiny icon is being asked to carry a very large argument. It may need to tell a voter that a video was manipulated, tell a reader how artificial intelligence shaped a public-interest article, tell a screen reader user the same thing without sight and tell a platform which machine-readable signal survived three exports and a screenshot. That is a lot of civic infrastructure to squeeze into one polite symbol. The European Commission says more than 90 signatories to its Code of Practice on Transparency of AI-generated Content established two task forces on September 30. One covers marking and detection. The other covers labelling. The work list is unusually concrete. It includes interoperable marking systems, standardized interfaces, provider-neutral detection tools, public benchmarks, red-team exercises, an interactive second layer behind an EU disclosure icon, an audio version called an earcon and sector-specific practices. Some of that work is already organized. Much of it is not yet finished. The Commission says an interoperability group is due to launch in October or November. Groups for the interactive second layer and audio earcon are scheduled for January 2027. Minutes from the September 30 and October 1 plenary meetings are still forthcoming. The plain signal is that Europe is building the right pieces, but the public value will depend on what happens after someone notices the label. A label without evidence is a warning light with no dashboard. The legal setting matters. Article 50 of the EU AI Act applies transparency obligations to several kinds of AI systems and content. Providers of generative AI systems are expected to make synthetic audio, images, video and text detectable in machine-readable form as far as technically feasible. Deployers have disclosure duties for deepfakes and certain AI-generated or manipulated text published to inform the public on matters of public interest. Those obligations began applying on August 2, 2026. The code of practice is voluntary, but the law is not. The Commission says signatories can use the code as a recognized route for demonstrating compliance. Organizations that use another route must show that their measures are adequate. That distinction is easy to lose in the confetti around an icon. The icon is not the law. It is one visible interface into a larger system of marking, disclosure and accountability. The first thing behind the click should be provenance. Who published the item? Which organization or person is taking responsibility for it? Which AI system contributed? Was the entire artifact generated, or did a person use a model for one part? When was the disclosure attached? Has the file changed since then? These questions sound basic because they are basic. They are also more useful than a generic message saying AI was involved. A photograph with a removed lamppost is not the same thing as a fabricated photograph of a minister accepting a suitcase. A translated paragraph is not the same thing as an invented news report. A synthetic voice reading an approved script is not the same thing as a voice clone used to impersonate a real person. The label needs to disclose the nature of the involvement, not merely the existence of a tool. The Commission's labelling task force explicitly lists that goal. Good. The hard work is turning it into categories ordinary people can understand without requiring a law degree or a tour of a metadata standard. The second layer should therefore answer five questions in plain language: What was generated or changed? Which system did it? Who made the decision to publish it? What evidence supports the label? How can a person challenge it? That last question matters because labels can be wrong in both directions. A missing label may allow deceptive content to travel as authentic. A false label may discredit genuine reporting, art or evidence. The harm is not symmetrical in every situation, but neither error should disappear into a support form designed to exhaust the person using it. The second thing behind the click should be the confidence of any detector. Detection is not a magic scanner that turns red when a lie walks through the door. Different media leave different traces. A watermark may be cropped. Metadata may be stripped by a messaging app. A synthetic image may be resized, filtered or printed and photographed. Text detection can struggle with short passages, editing, translation and ordinary writing that happens to resemble a model's output. The Commission's task force page does not publish detector accuracy, error rates or a universal interoperability result. It says the groups will work on benchmarks, best practices, red teaming and standardized interfaces. That is an agenda, not a scorecard. The distinction should be visible every time a detection result reaches the public. If a platform applies a label because a trusted provenance credential is intact, say that. If it applies a label because the uploader disclosed AI use, say that. If a statistical detector produced a probability, show the method, the confidence range and the conditions that can break it. Do not turn three different forms of evidence into one cheerful checkmark. The third thing behind the click should be an editing history. Provenance is often described like a chain. That is useful until real media start behaving like real media. A reporter records a video. An editor trims it. A tool reduces noise. Another tool translates speech. A producer adds synthetic narration for accessibility. A social platform compresses the file. A user screenshots one frame. The chain becomes a kitchen drawer full of adapters. The disclosure system needs to survive that mess gracefully. When a machine-readable mark is lost, the system should not pretend the history never existed. It should preserve the last verified step, identify the gap and avoid making claims it cannot support. That is why the task force's focus on interoperable interfaces matters more than a single fashionable detector. Interoperability means a signal created by one tool can be read and interpreted by another without forcing a person to visit five separate services. The Commission says the planned group aims to let ordinary people assess whether content is AI-generated or manipulated without trying several detectors independently. That is the right test. If the disclosure only works inside the company that created it, the public has not received a standard. It has received a loyalty program. The fourth thing behind the click should be accessibility. The labelling task force plans work on an audio earcon, a short sound that can communicate AI generation or manipulation. It also mentions spoken language and accessibility for visually impaired people. This should not become the audio equivalent of a cookie banner jingle. The sound needs a consistent meaning, enough distinction to be recognized and a way to access the fuller record. It should work with screen readers, captions, transcripts, broadcast audio and quiet public settings. It should not require a person to memorize a different chirp for every platform. Accessibility is not a final polish applied after the visual icon receives applause. It is part of whether the disclosure exists at all. The fifth thing behind the click should be responsibility. The public needs to know who attached the label, who can correct it and who is accountable when a disclosure is missing or misleading. An AI provider may create a machine-readable mark. A media editor may decide how much AI involvement to disclose. A platform may detect a signal and display an icon. A regulator may assess whether the system meets the law. Those roles should not collapse into a foggy statement that the ecosystem handled it. The evidence layer should name the responsible party at each step. It should also include a challenge path with a clock. If a journalist's authentic video is falsely labelled during an election, a response in 30 business days is not a remedy. If an unlabeled deepfake is spreading, a maze of automated replies is not accountability. The disclosure interface should record when a challenge was filed, who reviewed it, which evidence was considered and whether the label changed. Aggregate statistics should show error rates, response times and reversal rates without exposing private complainants. This is where public benchmarks and red-team exercises can become genuinely useful. A benchmark should not only ask whether a detector recognized pristine laboratory files. It should test crops, screenshots, compression, dubbing, translation, mixed human and synthetic work, accessibility transformations and deliberate attempts to remove or fake provenance. It should publish false-positive and false-negative rates by medium and condition. It should show how the system behaves when evidence conflicts. It should separate known provenance from statistical suspicion. A red team should include people who have reasons to break the system, and people who will be hurt when it breaks itself. That means researchers and security specialists, but also journalists, artists, disability advocates, election officials, teachers, civil-society groups and people whose real images have been falsely called fake. The task forces say independent experts and stakeholders may be invited. The quality of that participation will matter. There is also a governance problem hiding inside the signatory count. The broader code had about 190 signatories by the end of July, according to the Commission. More than 90 established the two implementation task forces in September. Those numbers show interest. They do not prove consistent compliance. The Commission should publish who participates in each working group, attendance, decisions, unresolved disagreements, benchmark methods, test data limitations and declared conflicts of interest. Meeting minutes being forthcoming is understandable at the start. It cannot become the permanent condition of a transparency project. The project needs its own second layer. There are practical uses for getting this right. A newsroom could attach a record showing that a photograph is authentic while noting that an AI tool translated the caption. A museum could identify synthetic reconstruction separately from original archival material. A court or insurer could distinguish intact provenance from a detector's guess. A social platform could present the same core disclosure through a visual icon, audio cue and machine-readable interface. A teacher could see whether a student disclosed translation help without pretending a detector can identify authorship from prose alone. The system will not eliminate deception. Bad actors may remove marks, forge records or publish through channels that ignore the rules. Legitimate creators may lose metadata through ordinary editing. Small organizations may struggle with implementation. Different Member States and market-surveillance authorities may interpret evidence differently. And a beautiful icon can still become wallpaper. That is the caution. Europe's task forces are promising because they connect visible labels to standards, benchmarks, red teaming, accessibility and machine-readable evidence. The danger is that the visible symbol arrives first and the evidence stays in a meeting agenda. The public does not need a tiny badge that says trust us. It needs a tiny door that opens.