OpenAI will invisibly mark texts generated by ChatGPT in the European Union

OpenAI will introduce, in the coming weeks, an invisible watermark for texts generated by ChatGPT and Codex in the European Union, as part of measures to comply with the requirements of the AI Act. The marker will be available to eligible users, regardless of their subscription type.

The technology, called textGrain, inserts a statistical signal into the way the model selects words. It can be detected automatically and indicates whether a text contains an OpenAI watermark, without revealing the user’s identity, the prompt, or the conversation from which it originated.

According to the company’s assessments, the watermark was identified in approximately 80% of texts containing 200 tokens and in 95% of those containing 400 tokens. However, effectiveness declines after editing: replacing 10% of the words with synonyms reduced the detection rate from 92% to 66%, while replacing 25% of the words brought the rate down to 17%. Short, restrictive, or translated texts are also more difficult to verify.

OpenAI warns that the system may produce false positives or false negatives. The watermark cannot determine how much a person wrote, who is responsible for the text, whether the information is accurate, or whether material without a marker was written by a human.

The detector will not be publicly available at launch. Access will initially be offered to approved researchers and specialized organizations. Globally, certain API customers will be able to opt in to the feature.