[Digital Today reporter Chi-gyu Hwang (황치규)] OpenAI will apply an invisible watermark to text generated by ChatGPT and Codex in the European Union. TechCrunch reported on Oct. 5 local time that the move is meant to comply with EU AI Act transparency rules that took effect on Aug. 2.

The watermark will be rolled out over the next few weeks to ChatGPT and Codex users in the EU across all plans. Developers using the OpenAI API can choose to switch on the feature for some models from Oct. 5 regardless of location, but it is off by default. OpenAI decided not to make text watermarking a default setting worldwide at launch.

The watermark does not work by attaching a visible mark. It slightly adjusts the model's word choices so the text appears unchanged to human readers but leaves a pattern that detectors can identify. The pattern remains in the text itself, so it persists even when copied and pasted.

OpenAI said the approach does not identify users and that there was no meaningful change in model performance even after the feature was turned on.

OpenAI also released a technical report on its 'TextGrain' technology. Co-authored by researchers at the University of Pennsylvania and Yale University, the report describes an example of sorting the next-word prediction order with a secret key. The company said that as these small adjustments accumulate hundreds of times, a detector can distinguish AI-generated content using only the text and the key.

Detection performance can drop with editing. In tests, replacing 10 percent of words with synonyms lowered the detection rate to about 66 percent from about 92 percent. It also said short sentences, math answers and translated text are harder to detect. For this reason, access to the detector will initially be provided only to approved researchers and professional institutions.

OpenAI said the absence of a watermark does not prove the writing was produced by a human. It said the text could be too short or heavily edited, or it could have been generated by another company's AI. It also said the watermark can show whether the OpenAI system generated or processed part of a sentence, but cannot show how much human judgment, editing and creativity were reflected.

The announcement follows Anthropic's statement two months ago that it would apply watermarking globally to text generated by Claude.