Can You Tell If a Text Was Written by AI?

Whether a text was written by AI can be reliably established only if the provider watermarked it and the watermark is verified with its key. The watermark is invisible when you read the text, and post-hoc AI detectors are no substitute: they are not accurate enough to base decisions on. Under Article 50 of the EU AI Act, providers have been required to mark AI-generated content in a machine-readable format since 2 August 2026, and the large model providers watermark: Gemini since 2024 and Claude for models launched from August 2026.

In this article we show how a text watermark works and why you cannot see it, who watermarks in production today, why post-hoc AI detectors are inaccurate, what the AI Act requires, and what follows from all this for your company.

Why can't you see the watermark in the text?

A language model produces text word by word: at every step it chooses among several candidates, and in many places more than one word would continue the sentence equally well. Watermarking uses these small decisions. Among the near-equal candidates the model chooses based on a secret key instead of chance, so the sequence of word choices carries a statistical pattern — invisible to the reader, verifiable with the key.

That is why there is nothing to look for in the text itself: the watermark is not in characters or formatting but in the way the words follow each other as the key dictated. Detection does not need the model either — with the method Google DeepMind published in Nature (SynthID-Text), the detector needs only the text and the key. The pattern is statistical, so certainty grows with length: a few sentences carry little evidence, a longer text carries a strong signal. The same fact sets its limits: heavy editing and rewording with another model weaken the pattern, translation replaces the word choices entirely, and constrained output — factual statements, code — leaves few choice points for the watermark to live in.

Who watermarks in production today?

• Google. SynthID-Text has been watermarking Gemini's responses in production since October 2024. DeepMind published the method in Nature, released a reference implementation as open source, and in a live experiment covering nearly 20 million Gemini responses found no statistically significant quality difference against the unwatermarked model.
• Anthropic. Claude watermarks its text output from August 2026 on, using a variant of SynthID-Text, worldwide. The company states the watermark does not affect response quality, a detection API is in the works, and models released before 2 August 2026 fall under a transition period.
• OpenAI. By its own account it developed an accurate text watermark but did not deploy it: translation and rewording with another model can defeat it, and in the company's judgement it would disproportionately affect some groups, such as non-native English writers.

The backdrop is European: the Commission published a Code of Practice on the transparency of AI-generated content, signed by around 190 organisations — among them several major model providers — before the marking obligation entered into application on 2 August.

Why are post-hoc AI detectors unreliable?

A post-hoc detector does something different from watermark verification: without a key, it guesses from style whether a machine wrote the text. That guess is not enough for a decision.

• OpenAI shut down its own detector in July 2023 over low accuracy. By the company's own account, the tool recognized 26% of AI-written texts while falsely flagging 9% of human-written ones as AI.
• A study published in the journal Patterns measured seven widely used detectors: they falsely labeled an average of 61.3% of essays by non-native English writers as AI-generated, while performing near-perfectly on native writers' texts.

The practical rule: an accusation or a sanction cannot be built on a detector result — not for submitted work, not for job applications, not for supplier material. What a detector flags is suspicion, not evidence.

What does the EU AI Act require?

Article 50 creates two separate obligations, and they do not fall on the same actor:

• As a provider — if you develop or offer under your own name an AI system that generates text, images, audio or video — you must mark the output in a machine-readable format so that its artificial origin is detectable. By the regulation's text, the solution has to be effective and reliable "as far as this is technically feasible" — the law reckons with the technology's limits.
• As a deployer — if your company merely uses AI — a disclosure duty for text arises when you publish it to inform the public on matters of public interest. By the regulation's text this does not apply where the text has undergone human or editorial review and a natural or legal person holds editorial responsibility for its publication.

These obligations have applied since 2 August 2026. The legal assessment belongs to your legal counsel; the engineering side is ours: where your systems emit generated content, and with what marking, can be assessed and designed. We describe the preparation steps in our EU AI Act preparation guide.

What follows from this for your company?

It can touch you in three roles, and the three call for different things:

• You write content with AI. Assume that the big providers' output carries a watermark: what you write with AI today and publish under your company's name may later be shown to be machine-written. You handle that risk with process: documented human review and editorial responsibility — the same thing the regulation's exception expects. For public-facing material, owning your AI use is a smaller risk than denying it.
• You ship an AI system to your users. If your product generates text, the marking obligation is yours. That is a system-design question: which model provider gives you what marking — SynthID-Text is open source, Anthropic is building a detection API — and what your own system has to add. In our generative AI integration projects, the EU AI Act risk classification happens at the start of the project.
• You make decisions about texts. HR material, coursework, supplier bids: do not accuse and do not sanction based on a detector result — the false-positive rate is documented, and it hits non-native writers hardest. If the question matters, handle it with process: declarations, version history, documentation of how the work was done.

And in procurement, ask every AI vendor: how does the system support marking under Article 50? The answer shows how prepared they are.

Our team has been building production enterprise systems since 2013 and has closed 46+ projects, and EU AI Act preparation is part of our services. You tell us where and what you use AI for, and we tell you where the marking obligation touches you and what belongs in your processes.