OpenAI says its textGrain method can add an invisible statistical watermark to eligible ChatGPT and Codex text in the EU, then detect the pattern in a passage. But the announcement is a plan for a rollout over the coming weeks, not proof that the method can reliably establish where everyday text came from.
The evidence supports a narrower claim: OpenAI reports that its detector finds the watermark in some tested passages. Whether it remains dependable across ordinary writing, code, translation and editing is unproven; the reported results come from OpenAI, and no independent evaluation of textGrain has been reported.
A watermark identifies a signal, not an author
OpenAI’s announcement says textGrain shapes a model’s word choices to embed a statistical pattern. Its technical report describes generation tied to keyed randomness, with a detector that checks a passage for the resulting signal. The mark is not a visible label attached to the text.
That distinction matters. A detector can report that it found a watermark; that does not tell a reader who used a model, how much a person contributed, who owns a passage or whether it is accurate. OpenAI says a missing mark does not prove human authorship either: text may be too short, changed or translated, or may come from an unsupported model or output made before watermarking.
OpenAI plans to apply the watermark to eligible ChatGPT and Codex text output in the EU, but has not named a start date or specified which outputs qualify. The company says the approach will not become a global default. API customers worldwide could opt in for select models, while access to the detector would initially be limited to approved researchers and expert organizations.
OpenAI’s results show how editing can matter
The company reports that detection depends on passage length and subject. At a target false-positive rate of 1%—the rate at which unwatermarked text is incorrectly flagged—its detector found watermarks in about 80% of 200-token psychology passages and about 95% of 400-token passages. OpenAI said results were substantially lower for mathematics, without giving a corresponding figure in its announcement.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Editing reduced detection in another company-reported test. For 400-token passages, replacing 10% of words with synonyms lowered the reported detection rate from about 92% to 66%; replacing 25% lowered it to 17%. The disclosed figures concern text passages and do not establish how well the detector performs on code.
These numbers are useful evidence that the signal can survive some conditions, and that modest rewriting can weaken it. They are not independent confirmation of real-world accuracy. A peer-reviewed ICLR 2024 study of a different watermark method found that, after strong human paraphrasing, detection in its test setting required an average of 800 tokens at a false-positive rate of 1e-5. That result cannot be transferred to textGrain; it underscores why performance depends on the method and conditions tested.
EU compliance is not the same as provenance proof
The European Commission says Article 50 transparency obligations apply from Aug. 2, 2026. Systems placed on the market before that date have until Dec. 2 to meet the marking and detection requirement. The Commission says providers must make AI-generated text machine-readable and detectable as far as technically feasible; its Code of Practice is voluntary, while the underlying legal obligations are not. The AI Act Service Desk sets out the timing.
The Commission’s technical report treats effectiveness, robustness, reliability and interoperability as relevant measures of text-marking approaches. Those are the right tests for a system meant to help identify generated content. A watermark that works on longer passages but weakens after editing may still provide a useful signal, but it cannot settle attribution on its own.
The verdict is supported, but limited, for the claim that textGrain can mark some output in a detectable way under tested conditions. Reliable everyday provenance remains unproven. Before anyone treats a detector result as evidence of authorship or misconduct, independent evaluations should test false positives and missed marks across text lengths, subjects, languages, code and substantial edits. OpenAI’s restricted detector access may allow such scrutiny; the company’s announcement does not yet provide it.