OpenAI announced on October 5, 2026, that it will add invisible watermarks to eligible ChatGPT and Codex text outputs in the European Union over the coming weeks. The rollout covers eligible users across all plans, but the company is not making text watermarking a worldwide default at launch. (openai.com)
API customers worldwide can opt in for select models immediately, with watermarking switched off by default. Access to the text detector is more restricted: OpenAI is accepting applications from researchers and expert organizations, with approvals granted case by case rather than opening a public checking tool. (openai.com)
The announcement therefore separates two capabilities: producing marked text and checking text for a mark. It also leaves an important question for developers unresolved. Naming Codex as a covered product does not establish whether, or how, generated source code itself will be watermarked. (openai.com)
What changes for users
OpenAI calls its technology textGrain and describes it as embedding an invisible statistical signal in the model’s word choices. A detector searches for that signal in a passage. The company has published an accompanying technical report, but the mark should not be understood as a visible label telling readers that text was generated by AI. (openai.com)
The announced deployment is prospective, not confirmation that every eligible EU account already produces marked text. OpenAI gives no precise completion date. It also plans availability through cloud partners and an open-source release, without specifying a date for that release. (openai.com)
For organizations using the API, opting into marking does not mean receiving detector access. That distinction limits the immediate use of the rollout as a general-purpose verification workflow: customers can choose marked generation, while checking remains subject to separate approval. (help.openai.com)
Performance claims are not guarantees
OpenAI reports no meaningful performance differences between watermarked and unwatermarked outputs across the benchmarks it uses to assess its Astra model. That is the company’s interpretation of its evaluations, not an independently established finding that every writing or software-development workflow will be unaffected. (openai.com)
Its own detection results illustrate the limits. At a target false-positive rate of 1%, OpenAI reports detecting about 80% of watermarks in 200-token passages and about 95% in 400-token passages for content such as psychology. Detection was substantially lower for mathematics. The target rate is an evaluation setting, not a verified everyday error rate. (openai.com)
Editing also weakened detection in the company’s tests. For 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. Those results describe particular tests, not universal outcomes for edited documents. (openai.com)
Codex coverage is not a code-quality finding
The announcement’s benchmark table includes software-related tasks, but those scores do not establish which portions of Codex output receive a watermark. Nor do they demonstrate that source-code quality is unchanged. The supported claim remains narrower: eligible Codex text output is included in the announced EU rollout. (openai.com)
Separate research provides relevant context, rather than validation of textGrain. In a July 2026 ICML paper, Pierre Fernandez and colleagues studied watermarking applied while a language model rewrites existing text. They reported strong results on open-ended material such as books, but difficulties with verifiable text such as code. Their different method cannot establish the behavior of OpenAI’s deployment. (proceedings.mlr.press)
The regulatory context
OpenAI presents the rollout as a response to the EU AI Act. Separately, Article 50(2) requires providers of systems generating synthetic content, including text, to make outputs machine-readable and detectable as artificially generated or manipulated. It qualifies the required effectiveness and reliability by technical feasibility, content limitations and the state of the art; it does not promise infallible detection. (openai.com)
European Commission guidance says Article 50 applies from August 2, 2026. It also identifies a limited transition until December 2, 2026, for the marking-and-detection obligation covering systems placed on the market before August 2. Those general dates do not, by themselves, establish which deadline applies to each OpenAI system. (digital-strategy.ec.europa.eu)
Provider marking and publisher disclosure are distinct obligations. Article 50 separately addresses AI-generated text published to inform the public on matters of public interest, with an exception involving human review or editorial control and a person or organization holding editorial responsibility. An invisible mark is therefore not a blanket answer to a publisher’s disclosure obligations. (eur-lex.europa.eu)
OpenAI’s own warnings set the practical boundary: a detected watermark does not identify the user, measure human contribution, establish ownership or verify accuracy. Conversely, an undetected watermark does not prove human authorship. For enterprise buyers and publishers, the rollout adds a limited provenance signal—not a conclusive verdict on who wrote a document or whether its contents are trustworthy. (openai.com)
