textGrain detects long output, but 25% editing cuts it to 17%

OpenAI will roll out the invisible textGrain watermark to eligible ChatGPT and Codex text across all plans in the EU over the coming weeks. The technology changes word choices so that a statistical signal remains in the text. API customers worldwide can opt in for select models, but the company is not making it the global default.

The detector will not be public at launch. Approved researchers and expert organizations can receive access on a case by case basis. It reports only whether an OpenAI watermark was detected and is not intended to reveal the user, prompt or conversation.

The EU gets a mark while the rest of the world gets an option

The regional rollout responds to EU rules requiring machine readable identification of generated content. OpenAI is keeping API watermarking optional for select models and plans to offer it through cloud partners as well.

The company says textGrain matched or exceeded the other methods it tested, including SynthID for text. Its published benchmarks found no meaningful quality difference between watermarked and ordinary output.

Longer and less constrained text gives the detector more signal

At a target false positive rate of 1%, the detector found watermarks in about 80% of 200-token psychology answers and 95% of 400-token answers. Results were substantially worse for mathematics, where the model has less freedom in word choice.

That defines the practical use. A watermark can help a provider establish that a longer, lightly edited passage came from a supported model. It cannot identify the author, verify accuracy, measure the human contribution or establish ownership of the text.

Ordinary editing dismantles the statistical trace

In a test on 400-token passages, replacing 10% of the words with synonyms reduced detection from about 92% to 66%. Replacing 25% cut it to 17%. Translation, shortening and deeper rewriting can weaken the signal in other ways.

A negative result therefore does not prove human authorship. A positive result is not a verdict on a person either, because OpenAI explicitly warns about false positives and false negatives. Restricted detector access reduces some misuse, but it also makes independent checks harder in an ordinary school, newsroom or company.

Handling uncertainty will matter more than the presence of a mark

The important operational question is how approved organizations report a score, passage length and reliability limits. Without that context, a binary answer can easily turn weak statistical evidence into an unjustified accusation.

OpenAI says it will expand the technical report and later release the technology as open source. Reproducible tests across languages, domains, translations and real editing workflows will show whether textGrain becomes an interoperable provenance layer or mainly a regulatory label on untouched output.

Lilith's verdict

A passage goes through two editors and the signal shrinks from 92% to 17%. Anyone treating that trace as a stamp of guilt is confusing a laboratory fingerprint with the author's ID.