OpenAI has chosen an invisible statistical watermark as its answer to Europe’s text-provenance obligations. It calls the method textGrain. Instead of attaching a visible label or metadata to a file, it subtly changes word choices and later looks for a statistical pattern. The company says some API customers will be able to opt in and that it will add the watermark to eligible ChatGPT and Codex output in the European Union over the coming weeks. The detector will not be publicly available at launch: OpenAI will accept applications from researchers and expert organisations. Source: OpenAI
The most revealing part of the announcement is its own data. At a target false-positive rate of 1 per cent, OpenAI says it detects the watermark in about 80 per cent of 200-token passages and 95 per cent of 400-token passages on psychology tasks. Performance falls in mathematics, where there is less freedom to choose synonyms. In a test of 400-token texts, replacing 10 per cent of words with synonyms cut detection from 92 per cent to 66 per cent; replacing 25 per cent brought it down to 17 per cent. These are results from the company’s evaluation, not an independent audit or a guarantee for every text, language or system. Source: OpenAI
That weakness is not a laboratory footnote. If a method depends on a statistical distribution of words, editing, summarising, translating or rewriting can change the signal. OpenAI’s report acknowledges that short text and topics with restricted vocabularies are harder to detect. It also says a watermark cannot reveal who used the system, how much a person contributed, whether the passage is true or who bears legal responsibility. The absence of a detectable mark does not prove human authorship: the text may come from an unsupported model, predate deployment or have changed enough for the detector to miss it.
The European rule now in force does not turn a detector into an authorship judge. Article 50 of the AI Act requires providers to mark generative content in a machine-readable format and make it detectable; the obligations began to apply on 2 August 2026, with a limited transition until 2 December for systems placed on the market before August. For certain texts on public-interest matters, a labelling obligation falls on deployers and is triggered only under defined conditions. The Commission says no label is required where the text has undergone substantive human review or editorial control and an editorial entity takes responsibility. Source: European Commission Source: European Commission
That distinction matters to journalism. A text generated or altered by AI does not become reliable because it carries a mark; a reviewed piece does not become human merely because an editor touched it. The rules separate two questions: whether a system generates detectable synthetic content, and whether a responsible editor examined the text and made the publication decision. The Commission describes review as deliberate examination of substance by people with relevant judgement, or real editorial control with power to alter, approve or reject the content on substantive grounds. Automated proofreading is not enough.
OpenAI’s decision to restrict the initial detector to specialists follows its warning about false positives and missed marks. The remaining questions are who will evaluate the method, under what conditions researchers may publish findings, and how long the watermark will remain useful against deliberate editing. The company plans to release the technology as open source and revisit its approach as evidence changes. For now, the announcement sets out a deployment plan and internal tests; it offers no public, independent comparison across several languages or journalistic uses.
Technical provenance can help identify some AI-generated content. It cannot certify truth, automatically distinguish assistance from authorship or settle responsibility. The useful contribution of this launch may be that the limitation appears beside the promise. A watermark can offer a clue; it cannot do the work of a newsroom, court or reader. When a method loses more than four-fifths of its detection after a quarter of the words are changed, public policy should regulate the purpose and process, rather than treating the signal as conclusive proof.