OpenAI to add text watermarks for ChatGPT and Codex users in the EU

OpenAI plans to add invisible watermarks to eligible ChatGPT and Codex text outputs for users in the European Union, while offering the feature as an opt-in setting to API customers worldwide.

OpenAI announces in an official blog post that the rollout is intended to help meet transparency requirements under the EU AI Act. The law requires providers of generative AI to make AI-generated text identifiable in a machine-readable form.

The company’s system, called textGrain, changes word choices in ways designed to leave an invisible statistical pattern. A detector can then assess whether a passage contains an OpenAI watermark. OpenAI says the mark does not identify a user, account, prompt, or conversation.

The company will initially limit access to its detector to approved researchers and expert organizations. OpenAI cites the technology’s risk of false positives, where a detector finds a watermark that is not there, and false negatives, where it misses one.

Editing remains a major limitation

OpenAI’s tests show that detection works better with longer and less constrained text. At a 1 percent target false positive rate, the detector found watermarks in roughly 80 percent of 200-token passages and about 95 percent of 400-token passages in subjects such as psychology. Results were weaker for mathematics, where models have fewer possible word choices.

Even modest revisions can significantly weaken the signal. Replacing 10 percent of words with synonyms reduced detection in one 400-token test from about 92 percent to 66 percent. Replacing 25 percent reduced it to 17 percent.

OpenAI says a detected watermark does not prove authorship, ownership, accuracy, legality, or the extent of human involvement. Likewise, no detected watermark does not prove that a human wrote a text. The company says it intends to publish more technical details and eventually open-source the technology.