OpenAI Begins Text Watermarking for ChatGPT in EU

OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act, the company announced on Monday, Oct. 5, 2026. According to a report by TechCrunch written by reporter Aditya Mehta, the transparency rules of the EU AI Act took effect on August 2 and require artificial intelligence companies to mark AI-generated content so that other systems can identify it.

Rollout Schedule And Regional Scope

OpenAI stated that the watermark will roll out over the coming weeks to eligible ChatGPT and Codex users across all subscription plans, though the deployment is restricted to the EU. Meanwhile, developers using OpenAI application programming interfaces anywhere in the world can enable the feature for select models, though it remains turned off by default. The company elected not to make text watermarking a global default at launch.

The technical deployment arrives as platform operators monitor evolving regulatory frameworks, paralleling how Chinese humanoid startup JoyIn challenges OpenAI over model similarities following new release across competitive international markets. The underlying system, designated textGrain, relies on secret keys to sort next-word predictions during text generation. Accumulated across hundreds of word choices, this pattern permits detectors to spot AI-generated content using only the text and the key.

Technical Constraints And Editing Vulnerabilities

OpenAI published a technical report detailing the method alongside its announcement, co-written with researchers from the University of Pennsylvania and Yale. The invisible mechanism travels with copied and pasted text because it resides within the word choices themselves, without identifying individual users. OpenAI reported seeing no meaningful change in model performance when the feature is switched on.

Testing revealed significant vulnerabilities to tampering and editing. In one company test, replacing 10% of the words with synonyms dropped detection reliability from about 92% to 66%. Furthermore, short passages, math answers, and translated text present distinct detection hurdles. These constraints influenced initial distribution plans, aligning with strategies seen when companies expand software architectures, much like how OpenAI expands ChatGPT plug-ins with app-like interfaces and automations for specific operational tiers.

Attribution Limits And Competitive Context

Due to the aforementioned reliability limitations, OpenAI restricted initial detector access to approved researchers and expert organizations tasked with evaluating responsible uses. The company cautioned that a missing watermark does not prove human authorship, noting that text might simply be too short, heavily edited, or sourced from another platform. Watermarks can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.

The announcement follows a decision by Anthropic two months prior to watermark text generated by Claude on a worldwide basis, a policy that drew user backlash regarding instructions and tool usage. OpenAI previously built a text watermark but delayed its release over market share concerns, according to a 2024 report by The Wall Street Journal. Anthropic, Google, Meta, Microsoft, and OpenAI are among the technology developers committed to following the European Union code of practice on AI-generated content.

When you purchase through links in our articles, we may earn a small commission, though this does not affect editorial independence. Aditya Mehta is a reporter at TechCrunch covering artificial intelligence, supported by the Tarbell Center for AI Journalism and attended UC Berkeley. Readers can contact Aditya by emailing aditya.mehta@techcrunch.com or via encrypted message at adymehta.74 on Signal.