OpenAI adds invisible watermarks to ChatGPT text

OpenAI is beginning to add invisible watermarks to text generated or edited with ChatGPT in EU countries, a move intended to make AI-produced content easier to identify, Swedish public broadcaster SVT reported.

The rollout will take place gradually over the coming weeks. At first, only researchers can apply to test the detection system. In theory, the technology could later be used to check whether, for example, a university assignment was generated with AI.

The watermark works by adding a statistical pattern to word choices, which can then be compared with the text. A match indicates that ChatGPT may have generated or edited it, said Peter Funk, an AI researcher at Mälardalen University. The watermark does not reveal personal information or identify who wrote the text.

OpenAI says the watermark is harder to detect in short texts and can become less detectable if the text is edited. It works better on longer texts, and the company says it produces errors in about one per cent of test cases.

Funk cautioned that the system should not be used on its own as proof that a student used AI. If 1,000 students were assessed, around 10 could be wrongly accused of using AI, even if they wrote their work themselves, he said. He also warned that a text without a detected watermark may still have been generated by AI, creating a false sense of certainty.

The change follows transparency requirements in the EU’s AI Act, which took effect in August. The rules mean people should be able to identify content generated or edited by AI. Similar invisible watermarks are already used for some images.