Skip to content
All articles
EU AI Act 8 min read· by Lars Zimmermann

Watermarks in AI text: who can actually check them?

Providers mark AI text invisibly in the choice of words. What Article 50 has required since August 2026, where the method fails and who owns the evidence.

In short

Text watermarks sit in the wording itself: for each next word the model is steered towards candidates matching a secret key. Only the key holder, the provider, can check this. Article 50(2) of the EU AI Act has required machine-readable marking since 2 August 2026, but does not say who may read it. Your own records therefore carry further.

Auf Deutsch lesen: deutsche Fassung

Someone sends me a text and asks whether it comes from an AI. I cannot tell him. Not because the tool is missing, but because the key to it belongs to someone else. Since 2 August 2026 the EU AI Act has required providers to mark generated text in a machine-readable way. On the other half of the matter, who is actually allowed to read this marking, it does not say a word.

Invisible means: in the choice of words, not in the file

Most people think of a watermark as something that sits in the file. A field in the metadata, an entry in the document header. Something like that would be removed in two seconds. With text it works differently, and this difference is the reason why the topic belongs in an audit and not in the IT department.

Google DeepMind published the method in Nature in October 2024, Dathathri and others, volume 634, pages 818 to 823. It is called SynthID-Text and starts at exactly the point where a language model selects the next word. There the model almost always has several equally usable options. The method has these candidates compete against each other in several rounds, controlled by a random function that is tied to a secret key. Whichever matches the key wins.

The result is a text that reads completely normally to you but carries a statistical pattern in its choice of words. The pattern does not lie alongside the text, it is the text. Copy it into another document and it remains. Delete a paragraph and it remains. Google describes the method as robust against cropping pieces of text, changing individual words and light paraphrasing.

This is not a laboratory state. Google states that it has shipped the method in Gemini and compared around 20 million responses in the process, without user ratings differing appreciably between marked and unmarked responses. Anyone working with one of the major providers today therefore in all likelihood already has marked texts in the house without knowing it.

The verification key belongs to the provider, not to you

Now comes the part that counts in an audit. A statistical pattern can only be found by someone who knows the function that generated it. Without the key, the marked text cannot be distinguished from any other text.

In its own documentation Google describes three levels of how a provider can handle the detection tool. Fully private, in which case only the provider itself checks. Semi-open, in which case there is an interface. Open, in which case the detector can be downloaded. Which level applies is decided by the organisation that deploys the method.

Feel free to read the last sentence a second time. Whether you can verify that a text came from a machine is not a technical question. It is the business decision of a company in which you have no stake.

That turns the usual narrative around. The watermark is readily sold as protection for the public. In practice it is first of all a means of evidence that belongs to the provider. It can recognise its own outputs. You cannot, neither with other people’s texts nor with your own.

Evidence that only one side can produce is not evidence in an audit. It is a statement.

On top of that, the methods are not compatible with one another. A detector finds the pattern it was built for, and no other. Anyone with three providers in the house would need three verification routes. The fact that Article 50(2) expressly requires the technical solutions to be interoperable is therefore not praise for the state of the art. It is a description of what is missing.

Where the method becomes weak, and why that is in the law

The method needs room for manoeuvre. It can only weave something in if the model is allowed to choose between several suitable words. Exactly this room is missing wherever the text is strongly determined.

Google writes this in the documentation itself: the marking works less well on factual responses, because there is less opportunity there to influence generation without degrading accuracy. That is an honest statement, and it reaches further than factual questions alone.

  • A short answer carries too little text for a pattern to be statistically secured.
  • A series of numbers, a formula, an address: hardly any freedom of choice, so hardly any signal.
  • Source code is determined over long stretches, otherwise it does not run. The comments in it are normal language and therefore the part in which something is most likely to stick.
  • Anyone who rewrites thoroughly or translates into another language lowers the hit probability considerably. That, too, is in Google’s documentation.

What is remarkable is that the Regulation knows this weakness. Article 50(2) requires the marking only as far as this is technically feasible. And it expressly requires providers to take into account the specificities and limitations of the various types of content, plus the costs of implementation and the generally acknowledged state of the art. The legislator has therefore written into the text that the method has gaps.

Two things follow from this for you. A missing watermark proves nothing. And a provider who promises you marking owes you information on where its method reaches its limit. This information is rarely in the data sheet. You have to ask for it, and you should have the answer in writing.

What Article 50 requires of whom, since 2 August 2026

The obligation hits two different roles, and in conversation they are almost always confused.

The provider is in Article 50(2). Anyone who supplies an AI system that generates synthetic audio, image, video or text content must ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. That is the manufacturer’s obligation, not yours. A system that merely performs an assistive function for standard editing is exempt, and so is one that does not substantially alter the input data or its semantics. One of these exemptions is sufficient. A third concerns systems authorised by law for law enforcement.

The deployer is in Article 50(4). For text a narrow case applies there: anyone who uses an AI system to generate or manipulate text that is published with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated. Then follows the half-sentence that almost everyone overlooks: the obligation does not apply where the content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for the publication.

Your quotation, your job advertisement, your internal inspection report do not fall under this. That is not informing the public on matters of public interest. Anyone in a small or mid-sized company who uses AI text in day-to-day business as a rule has no labelling obligation under Article 50(4). It becomes different as soon as you as a company publish on a topic of public interest and the text goes out unchecked. Then editorial control is more than good style. It is the reason why the obligation does not hit you.

On the deadline. Article 50 has applied since 2 August 2026 and was not postponed by the Digital Omnibus. What was postponed were the obligations for high-risk systems. For systems that were already on the market before that date, amending Regulation (EU) 2026/1744 provides a transition of four months for the marking obligation under paragraph 2, that is until 2 December 2026. For everything that came onto the market afterwards it applies without a grace period.

Who owns the evidence about your own text?

This is where it becomes uncomfortable, and this is where the real point for governance in the company lies.

Suppose a client accuses you of having had an inspection report generated by machine. You say a human wrote it. How do you prove that? A missing watermark does not help you, because it is also missing in short texts, in strongly determined texts and with every provider that does not mark at all. A watermark that is present, in turn, you cannot make visible yourself. You are left without means in both directions.

The same applies the other way round. You buy in texts and want to know which part of them came from the machine. You cannot determine it. You can have it warranted, and then you have ended up at the contract and not at the technology.

That is the reason why I lead customers away from marking and towards their own records. What is created in your own house you can also prove yourself, and better than any third-party detector ever could:

  • Which model in which version was involved, and at which point in the process?
  • Who approved the draft, and what is this approval tied to?
  • Is the input from which the draft was created still available?
  • Is it noted in the document itself what was created by machine and what was edited?

These are four questions that you answer in your own records. They do not depend on a key that someone else keeps. And in a dispute they carry further than any watermark, because they show the path and not merely the origin.

A warning to close this section. The watermark answers the question of whether a machine generated the text. It does not answer the question of who owns the text. Those are two different things, and they are constantly mixed up. Provenance is not an allocation of rights.

Where this connects to the standards

If you operate a management system, this is not a new topic. It is a familiar topic in a new place.

  • ISO/IEC 42001:2023 A.8.5, Information for interested parties. The control requires you to determine and document your obligations to report information about your AI systems. Article 50 is exactly such an obligation and belongs in this register with role and cut-off date.
  • ISO/IEC 42001:2023 A.8.2, System documentation and information for users. The limit of the method belongs here. Anyone who tells their people that the outputs are marked without saying when they are not has informed them wrongly.
  • ISO/IEC 27001:2022 A.5.33, Protection of records. The four questions above only become evidence once the records remain unchanged and retrievable.
  • ISO/IEC 27001:2022 A.8.32, Change management. A model change at the provider silently changes the behaviour and possibly also the marking. Anyone who does not record the change loses the attribution.
  • ISO/IEC 27001:2022 A.5.32, Intellectual property rights. The right place for the question of who owns a text. It is not in the watermark, after all.

The audit question I actually ask is, in the end, not a technical one. It is: who in your organisation knows which of the AI services you use mark their outputs, and how does that person know? The honest answer is mostly that nobody has checked. That is exactly where the work begins.

Share: LinkedIn E-Mail

Frequently asked questions

Do I have to label my AI text as a company?+

In most cases, no. The machine-readable marking under Article 50(2) hits the provider of the AI system, not you as a user. The disclosure obligation for deployers under Article 50(4) applies only to text that is published to inform the public on matters of public interest, and it does not apply if the text has been reviewed by a human and someone holds editorial responsibility. Quotations, job advertisements and internal documents do not fall under it.

Can I check myself whether a text comes from an AI?+

As a rule, no. The statistical pattern can only be found with the matching key, and whether the detector is made accessible to anyone at all is decided by the provider. It can keep it entirely to itself, offer it via an interface or make it publicly available. Commercially available AI detectors do not work with the watermark but guess on the basis of stylistic features, and this produces false hits in both directions.

What happens on 2 December 2026?+

Article 50 itself has already applied since 2 August 2026 and was not postponed by the Digital Omnibus. For AI systems that were already on the market before 2 August 2026, Regulation (EU) 2026/1744 grants a four-month transition for the marking obligation under paragraph 2. This period ends on 2 December 2026. For systems that came onto the market later, the obligation applies without a grace period.

Does the watermark survive editing?+

Partly. Google states that the method is robust against shortening pieces of text, changing individual words and light paraphrasing. A thorough rewrite or a translation into another language, by contrast, considerably lowers the detector’s confidence. Short and strongly determined texts carry hardly any evaluable signal from the outset.

Does a watermark count as evidence in an audit?+

Only to a limited extent. Evidence must be traceable and verifiable. A pattern that only the provider can read out does not meet that on its own. Records from your own organisation are more robust: model version, input, approval step and labelling in the document itself. This evidence belongs to you and is independent of whether a provider opens up its detector.

Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)

Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.

Questions about your own case?

In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.

Continue reading