OpenAI to watermark ChatGPT and Codex text in the EU

OpenAI introduces text watermarks for the EU

OpenAI will begin watermarking text produced by ChatGPT and Codex in the European Union to comply with the EU AI Act, TechCrunch reported. The company announced the change on 5 October 2026, setting out a regional approach rather than making watermarking standard worldwide.

The system embeds a detectable pattern through small adjustments to the words a model selects. There is no visible badge or character for readers to notice: identification depends on a detector recognising the pattern within the passage. According to TechCrunch, the company said copying and pasting the text preserves the watermark because it is carried by the wording itself.

That makes the mechanism different from a label attached to a document or displayed alongside an answer. For businesses moving generated text between applications, the identifying signal can accompany the content without relying on the original interface. Its survival through copying does not, however, establish that it will withstand subsequent changes to the wording.

EU transparency rules shape the regional approach

The EU AI Act’s transparency provisions took effect on 2 August, according to TechCrunch. They require AI-generated material to carry markings that other systems can recognise. OpenAI’s announcement places text generation within that compliance framework, alongside the broader industry effort to make synthetic content identifiable.

Anthropic, Google, Meta, Microsoft and OpenAI have committed to the EU’s code of practice covering AI-generated content, TechCrunch reported. Their participation provides a shared policy backdrop, although companies are choosing different geographical boundaries for deployment. Anthropic announced watermarking for Claude two months before OpenAI and is implementing it worldwide.

OpenAI had previously developed a text-watermarking system without releasing it. TechCrunch cited The Wall Street Journal’s 2024 reporting that concern about customers moving to competitors without watermarks was among the reasons for holding it back. The earlier decision shows that deployment has involved commercial considerations as well as technical readiness.

The latest announcement includes a technical paper describing OpenAI’s method, textGrain, written with researchers from the University of Pennsylvania and Yale. TechCrunch reported that the method uses a secret key to organise possible next-word choices. Across hundreds of small interventions, those choices build a signal that can be checked using the resulting text and the key, without needing the original conversation.

OpenAI reports limited performance impact but detection weaknesses

OpenAI said its testing found no significant effect on model performance when watermarking was enabled, according to TechCrunch. The company also said the signal does not reveal the user’s identity. Those positions address two distinct concerns for business users: whether the feature changes output quality and whether it exposes who generated a passage.

The company’s detection results nevertheless show a vulnerability to editing. In one test reported by TechCrunch, substituting synonyms for 10% of the words reduced detection from approximately 92% to 66%. OpenAI also identified brief passages, mathematical responses and translated material as more difficult cases, limiting the circumstances in which a detector can provide a dependable result.

OpenAI cautioned against treating an absent watermark as evidence that a person wrote the text. As TechCrunch reported, content may evade detection after substantial editing, contain too little material for analysis or originate from another AI provider. A negative result therefore cannot settle an authorship question.

A positive result has a different limitation: OpenAI said it can indicate its system’s involvement without measuring the human contribution to the final work. That distinction connects with the reaction to Anthropic’s announcement. TechCrunch reported objections from some Claude users who regarded their own instructions, contextual input and decisions as the substantive contribution, with the model serving as an instrument rather than the author.

Rollout and restricted detector access come next

OpenAI plans to extend watermarking to eligible ChatGPT and Codex users across every subscription plan in the EU over the coming weeks, TechCrunch reported. The announcement gives a rollout window rather than a single completion date, so eligibility and availability will matter for organisations managing access across teams.

Developers using OpenAI’s API have a separate route. From the announcement date, they can enable watermarking for selected models anywhere in the world, but the setting remains disabled unless they choose it. For a UK developer, that optional API feature is distinct from the EU-only ChatGPT and Codex rollout; OpenAI has not announced a worldwide default at launch.

Access to the detector will initially be restricted to approved researchers and specialist organisations, according to TechCrunch. OpenAI said those participants would help assess reliability and appropriate uses. Businesses should therefore distinguish between receiving watermarked output and having access to the means of checking it: the announcement does not provide general detector access.

For companies expanding across European markets, the practical task is to distinguish the product, model and location involved before assuming how watermarking applies. The detection findings also create a concrete constraint for content-review processes: the signal cannot serve as a complete record of authorship or human oversight. Those limits are relevant to the broader AI risks in market-entry planning previously covered by Market Entry Wire.