OpenAI will begin to add invisible watermarks to certain model text outputs from ChatGPT and Codex for users located in the European Union over the coming weeks, the company said Monday.

OpenAI's watermarking system is opt-in for API customers worldwide. The system will invisibly designate the text generated by selected AI models as AI-generated starting Oct. 5, though the feature will remain off by default. OpenAI attributed the rollout of the feature to the EU AI Act’s transparency requirements, an effort by the body to cut down on deepfakes and manipulated text "published with the purpose of informing the public on matters of public interest."

The watermarking system, which the firm calls textGrain, embeds a statistical pattern in the model's specific choices of language, but should not alter the general reliability of the model's outputs, as the watermarking produced no meaningful performance differences in OpenAI's benchmarks, the company said. OpenAI said it plans to open-source the technology, though its detector will be available only to approved researchers and expert organizations at first.

The company’s internal testing shows that editing can easily disrupt detection of the watermark, though. In one test of short, 400-token passages, replacing just one in ten words with a synonym cut successful detection from roughly 92% to 66%. Replacing a quarter of the words with synonyms reduced detection to just 17% in one company test.

Anthropic has already taken a similar watermarking step. As The Latent reported when Claude Fable 5.1 launched, its similarly invisible watermarking system prompted objections from users worried that even minor AI assistance could cause their work to be flagged, or that watermarking would make model outputs inferior. Fable 5.1 is currently the third-highest ranked model per The Latent's benchmark, behind OpenAI's GPT-6 Astra and Anthropic's Opus 5.5.

Short passages and subjects with limited flexibility in wording, such as mathematics, also proved harder to detect, per OpenAI's tests, as there is "less flexibility in word choice."

OpenAI said its watermarks cannot establish "who owns the text, whether its use was lawful, whether disclosure was required, or who is responsible for it." The system may also fail to recognize AI-generated texts from other models, or from earlier OpenAI models, and as such, a lack of a watermark does not establish that a human wrote the passage, per the firm.