OpenAI plans to add invisible text watermarks to eligible ChatGPT and Codex output in the European Union over the coming weeks, while API customers worldwide can opt in now for select models. That is the split in OpenAI’s Oct. 5 announcement, also covered by TechCrunch and 9to5Mac. In our reading, the split addresses the EU transparency rule while leaving API adoption to developers elsewhere.
The regulatory backdrop is the EU AI Act’s transparency requirement for AI-generated content, which the reporting says took effect on Aug. 2. OpenAI’s approach is built around a statistical signal called textGrain rather than a visible label. According to the company’s description relayed in the reporting, the system nudges word choice in ways that a detector can later identify. In plain operational terms, the mark lives in the text itself, so a copied and pasted passage still carries the signal.
That design matters for two different user groups. For people using ChatGPT or Codex in the EU, watermarking will appear automatically over the coming weeks on eligible accounts and on all plans, according to TechCrunch’s reporting of OpenAI’s announcement. For developers using the API, by contrast, watermarking remains something they choose to enable for select models. The split may help OpenAI address European transparency rules while allowing developers elsewhere to decide whether provenance signals fit their workflows.
OpenAI is also limiting who can inspect the detector. 9to5Mac reported that access will initially be restricted to approved researchers and expert organizations while the company evaluates the technology. The restricted launch means most organizations cannot use OpenAI’s detector for routine checks yet. More fundamentally, its stated error rates mean a detector result should not become a one-step verdict on whether a passage is human or machine written.
The company’s own caveats make that limitation even sharper. OpenAI says the detector can produce both false positives and false negatives, especially on short passages and on text in domains with less flexibility in wording, such as mathematics. It also said editing can weaken detection substantially: in one test described in the reporting, replacing 10% of words with synonyms reduced detection from about 92% to 66%. That makes watermarking more useful for relatively untouched output than for material that has already been translated, summarized or lightly rewritten.
OpenAI’s framing also avoids overpromising what the watermark can prove. The company says a missing watermark does not prove human authorship, because the text may be too short, too heavily edited or generated by another model entirely. It also says the signal does not identify the user, does not measure human contribution, does not establish responsibility and does not verify accuracy. That is the core decision boundary for readers: this is a provenance hint, not a chain-of-custody system.
The rollout also lands in an industry environment where watermarking is becoming more common but not uniform. TechCrunch noted that Anthropic introduced a worldwide Claude watermarking plan two months earlier, a move that drew criticism from some users who argued they supplied the instructions and decisions while the model was only the tool. OpenAI’s version is narrower in both geography and default behavior. It is EU-only for ChatGPT and Codex output, while API customers elsewhere must actively switch the feature on. Our assessment is that this balances regulatory demands against adoption friction, while leaving provenance practices uneven across vendors.
OpenAI says the technical report for textGrain was published alongside the announcement and co-written with researchers from the University of Pennsylvania and Yale, which gives the method some published structure even as the company acknowledges its limits. For businesses, compliance teams and product builders, the immediate implication is straightforward: if you need provenance signals, plan for watermarking as one input into review, not the final answer. If you do not need the signal, the API default staying off means there is no need to redesign every workflow around it on day one.
If you govern AI content, treat OpenAI's watermark as a provenance signal, not proof of authorship, and test how your review process handles short or edited text.
