ChatGPT Text Watermark Guide: textGrain Explained

Last updated: October 6, 2026

OpenAI has started adding an invisible signal to some AI-generated writing, and it will change how publishers, teachers and businesses talk about disclosure. The system is called textGrain. Announced on 5 October 2026, it rolls out first to eligible ChatGPT and Codex users in the EU, with an opt-in for API developers. This guide explains what textGrain does, who receives it, why the EU is first, and why a watermark check should not replace clear disclosure. This article is desk-validated from OpenAI’s announcement and reputable coverage, not hands-on tested.

Quick answer: OpenAI's textGrain adds an invisible watermark to ChatGPT and Codex text in the EU. Here is who gets it, how detection works, and what to disclose instead.

What is textGrain and how does the ChatGPT text watermark work?

textGrain is an invisible statistical watermark. Rather than adding visible marks or hidden characters, it gently biases word choices as text is generated. Over a few hundred words, those small choices form a pattern that a detector holding a secret key can recognise. OpenAI developed the approach with researchers from the University of Pennsylvania and Yale. The writing should read normally, but the distribution of words carries a faint fingerprint. That differs from metadata or labels, which can be stripped easily, and it is probabilistic: the detector judges likelihood across a passage rather than spotting one tell-tale word. See our AI guides hub for similar explainers.

Who gets textGrain, and when does it apply?

Eligible ChatGPT and Codex users in the EU across all plans receive textGrain over the coming weeks after 5 October. API developers worldwide can opt in for select models from 5 October; it is off by default for the API, and cloud partners will follow. So a ChatGPT user in the EU may soon produce watermarked text with no setting change, while a developer elsewhere must deliberately enable it. Two similar outputs can therefore carry different signals depending on product, region and model. Check plan and region before assuming a passage is covered.

Why is OpenAI launching textGrain in the EU first?

The driver is transparency law. OpenAI points to Article 50 of the EU AI Act, whose obligations apply from 2 August 2026, with pre-existing systems required to comply by 2 December 2026. Those rules push providers to make machine-generated content identifiable in a machine-readable way. A built-in watermark helps show regulators a measure that travels with the text itself. Anthropic introduced a similar Claude watermark in August 2026, so OpenAI is not alone in treating provenance as a compliance and trust issue.

Does textGrain identify you or change writing quality?

OpenAI says the watermark does not identify the user. The signal indicates that text likely came from a covered OpenAI model; it is not designed to name an account holder or reveal a prompt. Provenance is therefore not personal attribution. OpenAI also says testing showed no meaningful quality change. Watermarking does not improve accuracy, so verify names, dates and figures against primary sources before publishing. Our AI Guides category has more drafting checklists.

How reliable is detection if text is edited?

Detection weakens as text changes. Attributed to OpenAI’s announcement and coverage, in a 400-token test replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. Short passages, translated text, mathematics and heavily edited drafts are also harder to judge because the detector needs enough original word choices. A negative result does not prove human authorship, and a positive result is one signal among several, not a standalone verdict.

Can you rely on detection?

Only within narrow limits. This original table summarises when a check helps, based on the launch information above. At launch the detector is not public; access is initially limited to approved researchers and expert organisations, so most publishers and educators cannot run an authoritative check yet.

SituationWhat detection can tell youVerdict
Long, lightly edited EU ChatGPT or Codex passagePattern may be measurableUseful supporting signal; confirm context
Short note or headlineToo few tokens for a stable patternDo not rely on it
Rewritten or translated draftSignal degrades sharplyNegative result is inconclusive
Maths or code-heavy textFewer natural word choicesLow reliability
Misconduct claim based only on a checkProbabilistic, limited accessNever sole evidence

What should publishers and educators do instead?

Rely on disclosure and process, not a hidden signal. This checklist works whether or not a watermark is present.

Frequently asked questions

Is the textGrain detector available to the public?

No, not at launch. Access is initially limited to approved researchers and expert organisations, so most readers cannot get an authoritative verdict today. Be wary of third-party sites claiming to replicate the secret-key detector, and lean on disclosure and source checks instead.

Does a negative result prove a person wrote the text?

No. It only means no strong pattern was found in that passage. Short, translated, mathematical or heavily edited text can lose the signal, and uncovered models may never have carried it. Absence of evidence is not evidence of human authorship.

Can developers turn textGrain on outside the EU?

Yes, via the API opt-in. Developers worldwide can enable it for select models from 5 October, off by default. Document whether it is enabled, which models are covered, and how users are informed.

Will textGrain appear in every ChatGPT answer?

No. Coverage depends on product, region, model and rollout timing. EU ChatGPT and Codex users across all plans are first, phased over coming weeks, while API use needs an opt-in. Very short outputs may carry too little text for reliable detection.

Sources and methodology

Desk-validated on 6 October 2026 from OpenAI’s 5 October 2026 announcement and reputable coverage; not hands-on tested, and the detector was unavailable to us. We summarised the mechanism, rollout, EU AI Act dates and editing limits in our own words and attributed test figures to OpenAI’s announcement and coverage. Sources: B&T coverage; Search Engine Journal; OpenAI community post on EU text provenance.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!