OpenAI introduces invisible watermarks for texts in ChatGPT and Codex

The company OpenAI announced the launch of the textGrain hidden text marking system in ChatGPT and Codex for users from the European Union.

The innovation is due to the entry into force of the European law on artificial intelligence, which obliges the creators of generative models to ensure machine recognition of the created content. In the coming weeks, hidden labels will be introduced in responses on all tariff plans in the EU. In addition, API clients around the world have been given the opportunity to voluntarily enable this feature for a number of models, and access to the label detector at the initial stage is only provided to accredited researchers and profile organizations.

The textGrain method does not use hidden characters or spaces, but introduces a statistical pattern at the stage of selecting words and tokens by the neural network. According to OpenAI's internal benchmarks, the procedure has almost no effect on the quality of generation. At the same time, the developers note the vulnerability of the technology: in a text of 400 tokens, replacing 10% of the words reduces the detector's accuracy from 92% to 66%, and replacing 25% of the words - to 17%. The detector also performs worse with short fragments and precise disciplines with limited choice of formulations.

The company emphasized that the textGrain watermark does not identify the user, does not reveal the texts of his requests, does not evaluate the share of human participation in writing and does not serve as confirmation of the factual accuracy of the information.