OpenAI rolls out text watermarking to meet EU AI Act rules
It defaults on for ChatGPT and Codex, but only in the EU.
The facts
- API customers globally can opt in to text watermarking for select models starting now; it stays off by default in the API
- Within weeks OpenAI will add an invisible watermark to eligible ChatGPT and Codex text output in the EU only
- OpenAI's detector found watermarks in ~80% of 200-token passages and ~95% of 400-token passages at a 1% false positive rate, with much lower rates for math content
- Replacing 10% of words with synonyms cut detection from about 92% to 66%; replacing 25% cut it to 17%
- Detector access opens today but only to approved researchers and expert organisations, not the public
- OpenAI reports no meaningful benchmark performance differences between watermarked and unwatermarked output from its Astra model
OpenAI is switching on text watermarking in response to the EU AI Act. The law requires generative AI providers to make generated text machine-identifiable. The company's method, called textGrain, embeds an invisible statistical signal in a model's word choices. A matching detector then checks a passage for that signal. Starting today, API customers everywhere can opt in to watermarking for select models, though it stays off by default. Over the coming weeks, ChatGPT and Codex will watermark eligible text output automatically. But only for users in the EU.
OpenAI's own numbers explain the caution. At a 1% target false positive rate, its detector caught the watermark in about 80% of 200-token passages and about 95% of 400-token passages, for content like psychology. Math content leaves less room for word-choice variation, and scored substantially lower. Editing erodes the signal fast. Swapping 10% of words for synonyms in a 400-token passage dropped detection from about 92% to 66%. Swapping 25% dropped it to 17%. OpenAI says textGrain matched or beat other approaches it tested, including Google's SynthID for text. It also says benchmark results on its Astra frontier model show no meaningful quality cost from watermarking.
Detector access opens today too, but only to approved researchers and expert organisations vetted case-by-case under the EU's Code of Practice. The public gets nothing at launch. OpenAI plans to open-source the watermarking technology. It says it will expand access as it gains confidence results can be interpreted responsibly. The company also lists what a positive or negative detection cannot tell anyone. It does not measure how much of a passage is human-written. It does not establish ownership or legal responsibility. It does not identify the user who generated it. And it does not verify that the content is accurate.
The regional split means a EU-only default and a global opt-in regime will coexist in the API for the foreseeable future. Cloud partners are adding watermarking support in the coming weeks.
Why it matters
The low detection rate for short, technical or lightly edited text means the watermark will catch casual misuse. It will not catch a motivated actor who rewrites even a quarter of the output. Platforms leaning on it for compliance evidence should treat a clean detection result as corroborating, not conclusive. By keeping the API default off and the detector gated to vetted researchers, OpenAI satisfies the EU's letter now. It avoids exposing a public tool its own data shows can be fooled, or can falsely flag human writing as machine-made.
Primary source
OpenAI — openai.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.