OpenAI to Watermark ChatGPT and Codex Text in the EU

OpenAI announced text watermarking on October 5. Eligible ChatGPT and Codex outputs across EU plans will receive the signal over the coming weeks. Selected API models offer global opt-in access, disabled by default.

The system, textGrain, changes how a model selects tokens, the words or word fragments forming a response. OpenAI's technical report describes a secret key and preceding text guiding those selections. The result carries a statistical pattern rather than a visible label.

The method groups possible tokens into blocks, then selects a block and a token within the block. An entropy budget limits the sampling randomness removed to produce the signal. The detector reconstructs the pattern from the supplied text and secret key, without needing the generating model or its entropy budget.

Editing weakens detection. In OpenAI's tests, replacing 10% of words with synonyms reduced detection from roughly 92% to 66%. Replacing 25% lowered detection to 17%. Short passages and mathematics also proved harder to detect.

OpenAI already uses separate signals for images and audio. Its help center describes C2PA metadata recording file provenance and SynthID embedding a signal within supported media. Those checks establish neither accuracy nor ownership.

The text detector initially opens to approved researchers and expert organizations. Results reveal neither user identity nor human contribution. A missing watermark does not establish human authorship.

As we informed before, Anthropic will also watermark Claude-generated text.