ChatGPT text is getting an invisible watermark in Europe to comply with the EU AI Act. Here’s how it works

OpenAI is adding invisible watermarks to eligible ChatGPT and Codex text in the European Union to comply with new transparency requirements under the EU AI Act.

The rollout will begin over the coming weeks and apply across all ChatGPT and Codex plans in the EU. OpenAI is keeping text watermarking off by default elsewhere, though API customers worldwide can opt in for select models starting today.

The move comes less than two months after Anthropic introduced invisible watermarks for supported Claude models under the same EU rules. Anthropic says its watermark is embedded directly into generated text, allowing the signal to travel with the text when it is copied and pasted elsewhere and potentially survive some editing.

“We’re expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology,” OpenAI said in a post on X.

The company calls the technology textGrain. Rather than inserting visible labels, hidden characters, or metadata, textGrain changes the statistical pattern of word choices the model makes during generation. A detector can then analyze a passage and estimate whether it carries an OpenAI watermark.

The key word is estimate.

“Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model. In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Customers using our API can turn on text watermarking for select models worldwide today.”

However, OpenAI says text watermarking still has major limits, especially once text is shortened, rewritten, translated, or heavily edited. That makes the rollout as much an experiment in AI provenance as a compliance measure.

The watermark can disappear quickly after editing

OpenAI’s own tests show how fragile the signal can become.

For 400-token passages, its detector identified watermarked text about 92% of the time under the test conditions. Replacing just 10% of the words with synonyms pushed the detection rate down to 66%. Replacing 25% dropped it to 17%.

Text length matters too. At a target false-positive rate of 1%, the detector found watermarks in about 80% of 200-token psychology passages and roughly 95% of 400-token passages.

Impact of text length and type on detection rate

Performance was much weaker for mathematics, where models have less freedom in choosing alternative words and phrases.

That limitation matters far beyond academic testing. Users, editors, translation tools, and other AI systems routinely rewrite AI-generated text before publication. A watermark that works well on untouched output may look very different after several rounds of revision.

As a result, OpenAI is taking a cautious approach to detection. The company is opening applications for its detector, but access will initially be limited to approved researchers and expert organizations.

The public will not get an OpenAI text-watermark checker at launch.

OpenAI says that decision reflects the risk of both false positives and false negatives.

A watermark does not prove who wrote the text

OpenAI is drawing a sharp line around what its watermark can establish.

A detected signal can indicate that an OpenAI system generated or processed part of a passage. It cannot reveal who used the model, which account generated the text, what prompt was entered, or which conversation produced it.

It cannot establish copyright ownership or legal responsibility. It cannot determine whether the text is accurate. It cannot measure how much of the final work came from a human writer or editor.

The reverse matters too.

If OpenAI’s detector finds no watermark, that does not prove the text was written by a human. The passage could have been edited, translated, shortened, generated by an unsupported model, created before watermarking began, or produced by another AI provider.

OpenAI says textGrain did not cause meaningful declines across its internal benchmark testing of Astra, its latest frontier model. Scores moved slightly in both directions depending on the benchmark.

The company plans to publish more technical details and eventually release the watermarking technology as open source.

For now, the EU rollout creates an unusual split in how OpenAI-generated text will be treated. ChatGPT and Codex users in Europe will begin receiving eligible text with the invisible signal built in. Users elsewhere will not, unless developers choose to turn watermarking on through the API.

That difference reflects where AI regulation is heading.

Europe is moving from asking AI companies to disclose how their systems work to requiring machine-readable signals inside some of the content those systems produce.

OpenAI is complying, but its own numbers make one thing clear: an invisible watermark may help answer whether a model likely touched a piece of text. It is far from a definitive test of who actually wrote it.