OpenAI has announced its approach to text watermarking under the EU AI Act, saying it will use an invisible statistical watermark to help identify text generated by its models. The company said on October 5 that it will begin adding the watermark to eligible ChatGPT and Codex text outputs in the European Union over the coming weeks. It will also let API customers globally opt in to watermarking for select models, while keeping the feature off by default.

OpenAI said the watermark, called textGrain, changes the statistical pattern of the words selected by its models without adding visible marks or hidden characters. Its detector then checks whether a passage contains that pattern.

The company will initially restrict access to the detector to approved researchers and expert organisations. OpenAI said the detector will indicate only whether it detects an OpenAI watermark; it will not identify the user or reveal their prompts or conversations.

Watermarking has clear limits

OpenAI said its tests found that textGrain can detect watermarks more reliably in longer passages, but performance falls for shorter, highly constrained or heavily edited text.

At a 1% target false-positive rate, OpenAI said its detector identified watermarks in about 80% of 200-token psychology passages and about 95% of 400-token passages. Detection was lower for mathematics, where there is less flexibility in word choice.

Editing can also weaken the signal. OpenAI said replacing 10% of words in a 400-token passage with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%.

The company said its tests showed no meaningful difference in model performance between watermarked and unwatermarked outputs across the benchmarks it evaluated.

But OpenAI also stressed that a watermark cannot establish that a person did not contribute to the text, who owns it or whether it is accurate. It cannot identify the person who generated the text either. Importantly, the absence of a detected watermark does not prove that a human wrote the text. OpenAI said these limitations are one reason it is not making the detector publicly available at launch.

What does a text watermark actually prove?

MediaNama founder and editor Nikhil Pahwa raised similar concerns in August while writing about Anthropic’s text watermarking system. Pahwa described such systems as a statistical signal rather than a conventional watermark: “So the ‘watermark’ is closer to a statistical fingerprint than a label stamped onto the text.”

He also questioned what practical problem text watermarking is meant to solve. “What’s the ‘job to be done’ for AI watermarking of text?”

Pahwa argued that the distinction between AI-generated language and human thinking can be important when judging whether someone actually created an original idea. “Writing is compliance; thinking is not. What you’re trying to challenge really is unoriginal thought, not whether something was written with AI.”

He also pointed to cases where AI can help people express their own ideas, including when language is a barrier. “Someone who perhaps can’t express themselves well, but can think smartly and sharply, especially those who are not native English speakers, they can still use AI to express their thoughts better than they could on their own because of a language barrier. Should they be penalised?”

Pahwa also raised questions about translation and editing. If a person writes something in one language and uses AI to translate it, a detector could identify AI involvement even though the underlying ideas came from the person. “How will automated translation be impacted? If I had written this article in Hindi and used AI to translate it to English, would it be penalised as AI writing?”

He summed up the distinction as: “Just as translation isn’t transliteration, AI writing isn’t AI thinking.”

Editing and multiple AI tools can weaken detection

Pahwa noted that statistical watermarking does not necessarily provide a permanent record of AI involvement. He cited paraphrasing and the use of multiple AI systems as potential complications. “Only Claude can prove that an AI text output was written by it, because its statistical patterns are its own.”

He also wrote: “What happens when I use two different platforms to generate an output? Because each has its own patterns, and if I mix two or three AI generation tools to generate an output, the variance in the output by itself will mean that none of them will be able to detect it.”

Pahwa also warned against treating AI detection as definitive because of false positives. “These are imperfect tools, and AI writing detection is full of false positives. We need to be very careful about this.”

OpenAI’s rollout reflects the same caution. The company is introducing the EU watermarking as a regional measure rather than a global default, and it will initially limit detector access to researchers and expert organisations.

The move is also tied to the EU’s rules on AI-generated content. OpenAI said the EU AI Act requires providers of generative AI systems to make generated text identifiable in a machine-readable way. Anthropic had similarly linked its own text watermarking system to compliance with the EU AI Act.

For now, OpenAI’s approach makes the watermark a signal that an OpenAI system may have generated or processed text, rather than proof of AI authorship. The company said it plans to study how the watermark performs after editing and translation and may expand detector access as the technology and evidence develop.

Read more: