OpenAI published explanations and supporting material on its official announcement page. AUXER links to the original source rather than republishing imagery without explicit reuse permission.
OpenAI announced textGrain on October 5, 2026, a text-watermarking system designed to make AI-generated writing machine-identifiable. Eligible ChatGPT and Codex text in the European Union will receive the signal automatically over the coming weeks.
API customers get a different rollout: watermarking is available globally as an opt-in setting and remains off by default. The move is tied to transparency requirements under the EU AI Act.
No hidden characters
textGrain does not insert invisible spaces, secret characters or odd punctuation. Instead, the system changes the probability distribution used when the model selects words and word pieces during generation.
Across a long enough passage, those choices create a statistical pattern. A detector that knows what to look for can then estimate whether the text contains an OpenAI watermark.
Copy and paste does not automatically remove it
Because the watermark lives in the wording pattern rather than file metadata, copying text between apps does not inherently strip the signal. OpenAI also stresses the limitations: short passages, heavy editing and constrained text can make detection less reliable.
A missing watermark does not prove human authorship. A detected watermark also does not identify the user, establish ownership, measure how much a person contributed or prove that the content is accurate.
The detector will not be public at launch
OpenAI is initially limiting detector access to approved researchers and expert organizations. The company says false positives and false negatives are still possible, so it does not want the system treated as an automatic verdict on authorship.
What changes for developers
For API users, watermarking becomes a new compliance option. Products that generate text for customers, education, media, support or regulated workflows can add a machine-readable provenance signal without visibly changing the output.
OpenAI says performance impact is negligible and benchmark quality remains within normal run-to-run variation. It also plans to release textGrain as open-source technology so other organizations can build on it.
Why it matters
Most previous “AI detection” tools tried to infer origin after the text had already been generated. textGrain changes that model by embedding a planned statistical signal during generation itself.
It will not settle debates about authorship, disclosure or fraud. It does, however, create a more concrete technical layer for provenance than simply guessing from writing style.
Sources
- OpenAI — Our approach to EU text provenance rules, Oct. 5, 2026.
- OpenAI Help Center — Provenance signals in OpenAI-generated content, updated Oct. 5, 2026.
- OpenAI — Advancing content provenance.