OpenAI textGrain Watermark Arrives for EU ChatGPT Text

OpenAI has started marking machine-written text to comply with European law. The company announced on Monday, October 5, 2026, that the OpenAI textGrain watermark, an invisible statistical signal embedded in word choices, will roll out across ChatGPT and Codex in the European Union over the coming weeks. API customers worldwide can switch it on for select models, but it remains off by default outside the EU, and access to the detector that reads it is restricted.

What happened

According to OpenAI's help center page on provenance signals, textGrain "subtly adjusts how the model randomly chooses between possible words or word pieces as it writes." At each step, the system creates several adjusted sets of word probabilities that average out to the model's original likelihoods; a secret key selects which set to use. A detector holding the same key can then test whether a passage follows that hidden pattern more often than chance would allow. The watermark adds no hidden characters, invisible spaces or unusual punctuation.

ActuIA reported that OpenAI published a 20-page technical report alongside a blog post titled "Our approach to EU text provenance rules," and that the report is co-signed by five OpenAI researchers and four academics from the University of Pennsylvania and Yale. API customers can enable watermarking for a project or for an entire organization through a "Text provenance" setting and choose which models are covered. OpenAI says it plans to extend coverage to legacy models in the coming weeks and to release textGrain as open-source technology.

Detection has clear limits, which OpenAI published. Mashable reported that in testing the detector caught watermarks in about 80% of 200-word passages, with lower rates for subjects with rigid vocabularies such as mathematics. Replacing 10% of words in a 400-word passage cut detection to 66%, and replacing 25% lowered it to 17%. ActuIA said a 400-token passage on a psychology topic was detected about 95% of the time, at a detector setting targeting a 1% false-positive rate. Across the EU's 24 official languages, OpenAI's help page shows detection ranging from 69.0% in Spanish to 42.2% in Romanian at that threshold, with the watermark strengthened for languages below 60%.

Why it matters

The driver is Article 50(2) of the EU AI Act, which requires providers of generative AI systems to ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated." Most of the act's provisions have applied since August 2, 2026, and systems already on the market before that date have until December 2, 2026, to comply with Article 50(2), according to Mashable and ActuIA. Mashable said violations can draw fines of up to 3% of worldwide annual turnover. The OpenAI textGrain watermark is the company's answer for text, alongside existing watermarks and metadata for images and audio.

The design choices reveal the trade-offs regulators accepted. The EU's Code of Practice on transparency does not require watermarks for outputs under 200 tokens, about 150 words of English, or for code snippets, because short passages and code leave little room for varied word choices. OpenAI says watermarking had no measurable effect on model quality in its benchmarks and a negligible effect on speed.

Access to detection is the most contested element. OpenAI is initially limiting its text detector to approved researchers and expert organizations, citing reliability limits; current partners include Cornell's John Thickstun, ETH Zurich's Martin Vechev and researchers at Slovakia's Kempelen Institute of Intelligent Technologies. ActuIA noted that the code allows such restrictions for text but expects controlled access for regulators, media and fact-checkers, among others. OpenAI also stresses that a watermark does not establish authorship, ownership or accuracy, and that the absence of one "does not prove human authorship."

Competitors chose differently. ActuIA reported that Anthropic announced in August a watermark applied globally rather than only in the EU. For broader context on the pace of model releases, see Called It's coverage of Google's Gemini 4 Argon launch and of privacy-preserving access to AI services in the Ethereum Foundation's zkAPI.

What's next

The EU rollout across ChatGPT and Codex is due over the coming weeks, well before the December 2 compliance deadline. Companies that build products on OpenAI's API and sell them under their own names should check whether they need to switch the setting on, since the default outside the EU is off and EU guidance treats such firms as providers in their own right. Watch for the promised open-source release, wider detector access for regulators and newsrooms, and whether other AI developers follow OpenAI's EU-only approach to the OpenAI textGrain watermark or Anthropic's global one.