European AI Act makes machine-readable marking mandatory, but watermarks alone fall short

No. Since 2 August 2026 the European Commission mandates machine-readable marking of AI content under Article 50, but OpenAI and research show that watermarks have detection limits and prove neither accuracy, ownership nor human contribution. For confidential content you must keep your own internal provenance record with model, version, source and human review.

On 19 October 2026 the European Commission published an updated Code of Practice on Transparency of AI-generated Content, describing it as a voluntary practical instrument to help providers and deployers demonstrate compliance with the binding transparency obligations of Article 50 of the AI Act. Those obligations have applied since 2 August 2026. For professionals, the core question thereby shifts from "was this made by AI?" to "what evidence makes provenance, editing and human responsibility verifiable?"

That distinction is no play on words. A marking or label is intended to signal that AI was involved, but it does not by itself establish the extent or nature of that involvement. A marking or label does not by itself establish whether the content is correct, who is ultimately responsible or how much a human contributed to it. That is precisely where the risk lies for anyone working with confidential or highly confidential information.

What exactly does Article 50 of the AI Act require around marking and labelling?

Article 50 places two different obligations on two different parties. According to the guidelines of the European Commission, providers must furnish AI-generated or manipulated content with a machine-readable marking, while deployers must visibly label certain deepfakes and AI-generated texts on matters of public interest.

• Machine-readable marking (provider): a signal that software can detect, for example a watermark or metadata, indicating that content was made or edited by AI.
• Visible labelling (deployer): a human-readable notice for specific categories of content, such as deepfakes.
• Exceptions and equivalence: The Commission’s guidance and the Code of Practice describe relevant exceptions and, under the applicable conditions, allow alternative measures that are demonstrably equivalently adequate.

Our reading: in practice, compliance does not come down to one universal standard label, but to being able to demonstrate appropriate technical and organisational measures within the applicable obligation.

Why does a watermark or AI label prove neither accuracy nor human contribution?

Because a marking only signals provenance, not truth or authorship. OpenAI describes in its explanation of the EU text provenance rules a phased approach with opt-in text watermarks and invisible watermarks in EU text. In doing so, OpenAI itself names the limits: detection depends on text length and text type, editing can weaken the signal, and a watermark establishes no human contribution, ownership, user identity or factual accuracy.

Research points the same way. In a study published in the Proceedings of IJCAI-26, Qing Zhang, Yifei Huang, Juyoung Lee, Thad Starner and Jun Rekimoto tested with 81 participants an interface that makes provenance density visible instead of a binary "Made with AI" label. Participants distinguished truthful from fabricated text better than with no signal or a bare label. The authors warn explicitly, however, against false certainty, domain bias and the temptation to use provenance as a universal arbiter of truth; their technical audit was limited and provenance density alone was insufficiently distinguishing.

Our interpretation: two facts must never be confused here. 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 policy on a signal that the makers themselves describe as fallible.

What does this mean for directors, lawyers and CISOs who must account for confidential AI output?

Our analysis is that a machine-readable marking is insufficient as the sole internal accountability evidence for highly confidential content. We therefore advise directors to release external AI publications only after a linked provenance and review record, not merely after switching on a watermark. Because OpenAI reports that editing can weaken the watermark signal, our recommendation is that an organisation should not rely on an externally detectable feature as its sole proof; the CISO should log the chain of model, version and editing internally at the moment of generation, rather than reconstruct it afterwards. Because the Code of Practice permits alternative, equivalently adequate measures, our analysis is that lawyers should set out concretely in supplier and usage agreements which provenance evidence the organisation itself keeps and can produce. In our recommendation, organisations should consider requiring AI contracts to include an enforceable audit right with concrete inspection evidence rather than a general undertaking. Because the IJCAI study warns against false certainty, our recommendation is that no label should replace professional review and that the responsible handler should continue to test every factual claim at source level, as also applies with unchecked AI citations in legal decision-making. These recommendations sit within the broader framework of the EU AI Act and compliance.

What belongs in an internal provenance record for AI content?

A workable record captures, per piece of AI-influenced content, what was done, with what and under whose supervision. The Centers for Disease Control and Prevention’s guidance on disclosing generative AI use in scientific work offers a useful reference point for this record: 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.

1. Which content was influenced and what action was performed (generated, edited, summarised)?
2. Which model and which version were used, and via which route?
3. For what purpose and in what context the output was deployed.
4. Which source or provenance signals exist (marking, metadata, consulted sources).
5. Which human review took place and who took the release decision.
6. Which uncertainties or limitations remain, for example a weakened watermark or a non-detectable fragment.

The question all sources leave open is how you keep this record when the AI output itself contains confidential information. In our assessment that is precisely the crux: the provenance evidence must be verifiable for those permitted to inspect it, without making the non-public source content broadly available. Such a verifiable decision record instead of loose activity logs is what marking alone does not deliver: the professional final judgement remains with the human.

Sources: The article relies on guidelines and the Code of Practice from the European Commission, OpenAI's explanation of the EU text provenance rules, an IJCAI-26 study and a CDC guideline on disclosure.