AI Act machine-readable marking but watermarks fall short

You must now keep your own internal provenance record for any AI-generated or manipulated content you release externally, documenting the model, version, source data and human review at the moment of generation. Machine-readable marking alone satisfies the technical requirement but does not establish accuracy, ownership or human contribution.

The prompt is an analysis of 28 October 2026 of machine-readable marking for AI-generated content under Article 50 of the EU AI Act, which argues that watermarks have detection limits and cannot by themselves prove accuracy, ownership or the extent of human involvement. The European Commission published an updated Code of Practice on Transparency of AI-generated Content on 19 October 2026 to help providers and deployers demonstrate compliance with binding transparency obligations that have applied since 2 August 2026. In our assessment, the practical shift for anyone handling confidential information is from asking "was this made by AI?" to being able to demonstrate verifiable provenance, editing history and human responsibility through documented controls rather than relying on external marking alone.

What does Article 50 actually require of you?

Article 50 places distinct obligations on two parties. Providers must furnish AI-generated or manipulated content with machine-readable marking. Deployers must visibly label certain deepfakes and AI-generated texts on matters of public interest. For most professionals, the operative constraint is that you cannot satisfy the requirement through marking alone. A label signals that AI was involved but establishes nothing about the accuracy of the output, who bears ultimate responsibility, or how much human judgment shaped the final product. That distinction matters most when the content is confidential or highly confidential.

Why do watermarks not establish what the regulation requires?

Watermarking technology has known limits that the vendors themselves describe. Detection depends on text length and text type. Editing can weaken or destroy the signal. A watermark establishes no human contribution, no ownership, no user identity and no factual accuracy. Research testing provenance density as an alternative to binary "made with AI" labels found that participants distinguished truthful from fabricated text better with provenance information than without, but the authors explicitly warned against false certainty and against treating provenance as a universal arbiter of truth. The technical audit was limited and provenance density alone was insufficiently distinguishing.

Two facts must never be confused. A label does not prove that content is incorrect, and the absence of detection does not prove that content was made by a human. Anyone who equates either with "reliable" or "unreliable" builds accountability on a signal that the makers themselves describe as fallible.

What controls must you be able to demonstrate?

1. Record the model and its version — document which model, which version and which deployment environment generated each piece of content at the moment of generation, not afterwards.
2. Log the source data and training context — capture which source data, which retrieval systems and which prompts fed the model, so the chain is reconstructible.
3. Document human review at each stage — record who reviewed the output, what they checked, what they changed and when they approved release.
4. Preserve the editing history — maintain a record of any edits, refinements or corrections made after initial generation, since editing can weaken external detection signals.
5. Establish an audit trail for inspection — ensure the record is verifiable by those permitted to inspect it without making non-public source content broadly available.
6. Include enforceable audit rights in supplier agreements — require AI contracts to specify which provenance evidence your organisation keeps and can produce, with concrete inspection rights rather than general undertakings.

How should you structure the record itself?

A workable record captures, per piece of AI-influenced content, what was done, with what tool and under whose supervision. The CDC's guidance on disclosing generative AI use in scientific work offers a practical reference: disclose which tool, model and version were used, where the tool was deployed and what human review took place. The CDC additionally advises C2PA metadata where possible and warns against entering sensitive or non-public data into public tools. The challenge all sources leave open is how you keep this record when the AI output itself contains confidential information. The provenance evidence must be verifiable for those permitted to inspect it, without making the non-public source content broadly available. That is precisely where marking alone fails: it does not deliver the decision record that accountability requires.

What stays your professional responsibility?

No label should replace professional review. The responsible handler must continue to test every factual claim at source level, as also applies with unchecked AI citations in legal decision-making or high-stakes analysis. Tooling can capture and preserve the chain of model, version and human decision. What tooling cannot do is substitute for your own judgment about whether the content is fit for its purpose and whether the human review was adequate. The machine-readable marking satisfies the technical requirement. The provenance record satisfies the accountability requirement. Neither satisfies the professional requirement, which remains yours alone.