OpenAI Adds Invisible Text Watermarks in Europe to Comply With EU AI Act

OpenAI will begin embedding invisible text watermarks across eligible ChatGPT and Codex outputs in the European Union over the coming weeks to comply with transparency mandates under the EU AI Act, Tech Startups reported (https://techstartups.com/2026/10/05/chatgpt-text-is-getting-an-invisible-watermark-in-europe-to-comply-with-the-eu-ai-act-heres-how-it-works/). The feature remains off by default in other jurisdictions, though API customers globally can opt in for select models immediately.

Branded as textGrain, the system modifies the statistical probability distribution of word selections during text generation rather than injecting visible tags, metadata, or hidden characters. A proprietary detector then parses the resulting text passage to estimate whether it was generated by an OpenAI model. OpenAI stated in a post on X that it is expanding its provenance framework from audio and images to text to satisfy European regulatory obligations, while acknowledging the inherent constraints of current text-watermarking approaches.

OpenAI disclosed that the technology experiences significant degradation under common real-world conditions such as shortening, paraphrasing, translation, and editing. In vendor-run tests on 400-token passages, the detector correctly identified watermarked text roughly 92 percent of the time. Replacing 10 percent of the words with synonyms reduced detection accuracy to 66 percent, and a 25 percent replacement dropped it to 17 percent.

Length and subject matter also constrain detection reliability. Setting a target false-positive rate of 1 percent, OpenAI detected watermarks in approximately 80 percent of 200-token psychology passages and 95 percent of 400-token passages. Performance dropped considerably on mathematics text, where constrained phrasing offers fewer statistical alternatives during generation.

Because of risks associated with both false positives and false negatives, OpenAI is restricting initial detector access to approved researchers and specialist organizations, with no public verification tool available at launch. Internal benchmark testing on Astra, OpenAI's latest frontier model, indicated textGrain caused no meaningful performance drops, with benchmark scores shifting marginally in both directions. OpenAI stated it plans to publish additional technical specifications and eventually release the watermarking system as open source.