On October 5, 2026, OpenAI announced a phased plan to add its textGrain watermark to eligible ChatGPT and Codex text in the European Union, across plans. The announcement set out a rollout over the coming weeks. For US users, it did not announce a ChatGPT or Codex watermarking default: the separate worldwide option is an opt-in for select API models, and the API feature is off by default. OpenAI’s announcement describes the two plans separately.
OpenAI’s textGrain plan is EU-specific; API opt-in is worldwide
The planned ChatGPT and Codex rollout applies to eligible text output in the EU. OpenAI described it as phased and covering plans across those products. That scope is different from the API option: OpenAI said API customers worldwide could opt in to text watermarking for select models, while the feature would remain off by default.
In practice, the announcement did not make textGrain a global default for ChatGPT or Codex. The worldwide provision it described applies to API customers who choose the option for eligible models.
What textGrain changes in generated text
OpenAI describes textGrain as an invisible statistical signal embedded in a model’s word choices. It is not a visible label or an inserted hidden character. A detector looks for the pattern; OpenAI said access to its text detector would initially be limited to approved researchers and expert organizations, with applications considered case by case.
OpenAI also reports no meaningful performance difference with and without textGrain across its Astra benchmarks. That is the company’s benchmark result, not a claim that every task or response will be identical.
OpenAI’s reported detection figures depend on the test
OpenAI’s published evaluations give different results for different passage lengths and editing conditions. For psychology passages, the company reported about 80% detection at 200 tokens and about 95% at 400 tokens, with a 1% target false-positive rate. Those figures are specific to that evaluation.
A separate test used 400-token English responses to ELI5 questions. OpenAI reported about 92% detection before editing, 66% after replacing 10% of the words with synonyms, and 17% after replacing 25% of the words. The edit conditions drew on the same pool of 600 original responses, with length eligibility assessed separately for each condition.
| Evaluation and content | Passage condition | Detection reported |
| Psychology passages | 200 tokens; 1% target false-positive rate | About 80% |
| Psychology passages | 400 tokens; 1% target false-positive rate | About 95% |
| English responses to ELI5 questions | 400 tokens; before editing | About 92% |
| English responses to ELI5 questions | 400 tokens; 10% of words replaced with synonyms | 66% |
| English responses to ELI5 questions | 400 tokens; 25% of words replaced | 17% |
OpenAI says detection is substantially lower for constrained content such as mathematics, where there is less flexibility in word choice. It does not give one general detection rate for math. The reported figures are results from OpenAI’s evaluations, not universal performance guarantees.
A watermark is not an authorship verdict
A positive result does not identify the person, account, organization, prompt, or conversation associated with a passage. OpenAI says the signal also does not establish ownership, responsibility, accuracy, or how much human work contributed.
A negative result does not prove that a person wrote the text. OpenAI says short, edited, or translated passages, text from unsupported models, older output, and output from other providers may lack a detectable OpenAI signal. And because detector access is initially limited to approved researchers and expert organizations, the text detector is not a public self-checking tool.
