OpenAI is adding an invisible, machine-readable watermark to text and code generated by ChatGPT and Codex, but only for users in the European Union at first. The company says the rollout, using technology it calls textGrain, will happen over the coming weeks and will cover eligible users on all plans in the EU. OpenAI is not making watermarking a global default at launch, according to a report by The Verge.
The move is designed to meet Article 50 of the EU AI Act, which requires providers of generative AI systems to make text outputs identifiable in a machine-readable way. Existing providers have until December 2 to comply. Anthropic made a similar move earlier this year with a watermark based on Google DeepMind's SynthID for text.
What textGrain watermarking does
TextGrain works by adding an invisible statistical signal to the model's word choices, according to Engadget. A compatible detector can then read that signal and report whether an OpenAI watermark is present. OpenAI claims the system matched or exceeded other approaches, including Google DeepMind's SynthID for text, in internal benchmarks.
But OpenAI cautions that textGrain does not guarantee reliable detection. Watermarks do not verify accuracy, determine ownership of the text, measure how much a human contributed, or prove human authorship, the company says. Detection is also harder for short passages, mathematical content and text that has been edited.
textGrain detection rates and limits
OpenAI has published detection figures that show how fragile watermarks can be. Tests with a target false-positive rate of 1 percent identified watermarks in about 95 percent of 400-token passages about psychology. At 200 tokens, detection fell to about 80 percent. Rates were substantially lower for math content, where the model has less freedom to choose words.
Editing weakens the signal further. Replacing 10 percent of words with synonyms cut detection from about 92 percent to 66 percent for 400-token passages. Replacing a quarter of the words dropped it to 17 percent, making the watermark easy to evade. OpenAI says watermarking does not hurt output quality, citing no significant differences across eight benchmarks including GPQA Diamond, BrowseComp and DeepSWE.
