Brussels — OpenAI to Watermark ChatGPT Text by Default in EU

Brussels — OpenAI will begin automatically watermarking text generated by ChatGPT for users inside the European Union, the company said Oct. 6, a step it is taking to comply with the European Union's artificial intelligence law, which took effect in August, according to Ars Technica.

The feature will also be offered to users outside the bloc, but it will stay switched off by default anywhere regulators do not require it, Ars Technica reported. OpenAI is rolling the change out first in ChatGPT and Codex, the company's coding tool; the watermark will remain off by default in OpenAI's developer API.

What does the EU AI Act require of OpenAI?

The EU AI Act, in force since August, requires that content produced by AI models be marked in a way that a separate tool can detect, according to Ars Technica's reporting. The law does not mandate a specific technical standard, and no watermarking method now in wide use — including the SynthID system and the C2PA industry project — meets a fully reliable bar, Ars Technica said. OpenAI's answer is a proprietary method it calls textGrain, described in a technical paper the company has published alongside the rollout.

How does OpenAI's textGrain watermark work?

textGrain embeds patterns in an AI model's word choices that are invisible to an ordinary reader and do not noticeably change the quality of the output, Ars Technica reported. Someone holding a cryptographic key can run a specialized detector against a piece of text to check whether those patterns are present. OpenAI said it will give that detector to a limited number of researchers and organizations at first, with a request-for-approval process for others to gain access over time.

How reliable is the detection, by OpenAI's own numbers?

OpenAI's internal tests found a 92 percent successful detection rate for textGrain-marked text, Ars Technica reported — a figure the outlet called "respectable but not entirely reliable." The same tests showed the system degrades quickly once text is edited.

By the numbers, per OpenAI's tests as reported by Ars Technica:

• 92 percent: baseline successful detection rate for unmodified textGrain output
• Editing 10 percent of a passage cuts the detection success rate by almost 30 percent
• Editing 20 percent of a passage cuts the detection success rate by almost 75 percent
• Detection is also weaker on shorter passages and on translated text, according to the tests

"Changing just 10 percent of the text in an output reduces the successful detection rate by almost 30 percent, and changing 20 percent of the text can lower the success rate by almost 75 percent." — OpenAI's own testing results, as reported by Ars Technica

Ars Technica noted that this pattern is consistent with other watermarking tools the outlet has reviewed in the past: effective against casual copying, easy to defeat with modest editing by anyone with basic technical know-how.

Why is Anthropic's approach different?

Anthropic introduced its own text watermarking in August and turned it on globally, not just where regulation requires it, Ars Technica reported. OpenAI has taken a narrower path, making the feature the default only in jurisdictions — currently the EU — where regulators require it, at least inside the ChatGPT and Codex apps. Outside those jurisdictions, and inside OpenAI's API everywhere, the watermark stays off unless a user or developer turns it on.

Timeline

• August 2026: EU AI Act's content-marking requirement takes effect; Anthropic turns on global watermarking for its models
• Oct. 6, 2026: OpenAI announces textGrain and EU-default watermarking
• "Coming weeks": Rollout to ChatGPT and Codex users inside the EU, per OpenAI's statement to Ars Technica
• Ongoing: Limited researcher access to the textGrain detector, expanding through an approval process OpenAI has not detailed publicly

What to watch

• Whether EU regulators treat a 92 percent baseline detection rate — one that falls sharply after light editing — as sufficient compliance with the AI Act's marking requirement
• Whether OpenAI widens default watermarking to the API or to users outside the EU, as Anthropic has already done globally
• How quickly, and to whom, OpenAI expands access to the textGrain detector beyond an initial set of researchers and organizations
• Whether EU authorities publish guidance on what counts as adequate detection reliability under the AI Act

The underlying trade-off is not new to OpenAI. Every watermarking scheme now deployed by a major AI lab trades resistance to tampering against the demand, in both law and ordinary use, that AI-generated text still read naturally — and Ars Technica's review of OpenAI's own test results suggests textGrain has not escaped that trade-off any more than its predecessors have.

AlterEgo — A keyboard for character voices. Five free persona rewrites a day.