SDSignal Desk

OpenAI’s EU-default text watermark fades fast under light edits, its own data shows

· Updated · Signal Desk newsroom

Analysis of reporting by Ars Technica. How we work

Image: Ars Technica

OpenAI will turn on its textGrain watermark by default for ChatGPT and Codex text in the EU to meet the AI Act, keeping it opt-in elsewhere. Its testing found about 92% detection, falling almost 30% after 10% of text changes and almost 75% after 20%.

Key takeaways

  • textGrain becomes default for EU ChatGPT and Codex text in the coming weeks; it is opt-in elsewhere, including the API.
  • The EU AI Act requires machine-detectable marking of AI-generated content.
  • Detection was about 92%; editing 10% of text cut it by almost 30%, and 20% by almost 75%.
  • Shorter and translated text are harder to detect.
  • Anthropic turned on text watermarking globally in August.

The policy

Ars Technica reports that OpenAI will turn on its textGrain watermark by default for ChatGPT and Codex text in the European Union in the coming weeks, and offer it as an off-by-default option elsewhere, including the API. The trigger is the EU AI Act, in force since August, which requires AI-generated content to be marked in a way other tools can detect. OpenAI explained its approach in its own post (EU provenance rules). Like other language-model watermarks, textGrain nudges word choices in patterns a reader will not notice but a keyed detector can find. Detector access is limited to researchers and approved organizations for now.

The durability numbers

OpenAI’s own testing found the watermark was detected about 92 percent of the time. Changing 10 percent of the text cut the detection rate by almost 30 percent, and changing 20 percent cut it by almost 75 percent. Shorter and translated text are harder to catch. Ars notes existing approaches such as SynthID and C2PA are also relatively easy to circumvent with basic know-how.

A signal that fades sharply after a fifth of the words change will survive casual copy-and-paste but not deliberate editing, paraphrasing or translation. Detection will mostly flag people who left output largely untouched, such as hurried students or honest users, while sophisticated actors pass through. That makes it reasonable for measuring broad patterns of synthetic text but a poor basis for discipline, grading or accusations; a detection hit should be treated as a clue rather than proof.

Regional split

Anthropic switched on text watermarking for its models globally in August. OpenAI is making it the default only where regulators require it, at least for its apps, a defensible business decision that signals it sees the value mainly in compliance. If the main argument for watermarking is a healthier information ecosystem, EU-only defaults leave most of that ecosystem unmarked. OpenAI deserves credit for publishing a technical paper and numbers that expose the weaknesses. How EU regulators judge compliance given published circumvention rates, and whether watermarks are combined with other provenance signals, are open questions. Google has expanded public watermark checking for media (SynthID checker). More in our OpenAI coverage and regulation coverage.

Who feels it

EU ChatGPT and Codex users
Default watermarking within weeks; edited or translated text may not carry a signal.
Educators and employers
OpenAI’s data argues against treating detection as proof.
API developers
Opt-in, so EU-obligated teams must decide deliberately.
Regulators
The AI Act’s marking rule meets tools with documented limits.

What to watch

  1. Independent tests from researchers with detector access.
  2. Whether OpenAI makes watermarking a global default.
  3. EU regulators’ compliance judgments.
  4. Combining watermarks with other provenance signals.

FAQ

Where is the watermark on by default?

In the EU, for ChatGPT and Codex text; elsewhere it is opt-in.

How reliable is it?

About 92% detection, dropping sharply when 10–20% of the text is edited.

Why only the EU?

The EU AI Act requires machine-detectable marking of AI content.

Who can use the detector?

Researchers and approved organizations, for now.

Read the original

Continue at the source.

Ars Technica

Companies: OpenAI

More on OpenAI

All OpenAI coverage →