OpenAI turns on textGrain watermarking in the EU

OpenAI's move is framed as a response to the EU AI Act, which reportedly requires AI providers to label generated text in a machine-readable format. Rather than rolling out a single global default, OpenAI is splitting the approach: automatic watermarking for ChatGPT and Codex users in the European Union over the coming weeks, but opt-in watermarking for API customers worldwide. That stands in contrast to Anthropic, which applies mandatory watermarking to Claude globally regardless of how users access the models. OpenAI's API watermarking will also reach cloud partners such as Microsoft Azure in the coming weeks.

The technology behind this, textGrain, embeds an invisible statistical signal in the model's word choices, working much like Claude's SynthID watermark, which uses Google's open-source technology. OpenAI says textGrain matched or beat other approaches in internal tests, including Google's SynthID for text, and plans to release it as open source. But detection is far from absolute. With the detector set to a 1 percent false-positive rate, it identified watermarks in about 95 percent of 400-token passages about psychology, falling to about 80 percent at 200 tokens, and substantially lower for math content where the model has less freedom to choose words. Editing erodes the signal further: replacing just 10 percent of words with synonyms cuts detection for 400-token passages from about 92 percent to 66 percent, and replacing a quarter drops it to 17 percent.

OpenAI also says watermarking doesn't hurt output quality, citing tests of its frontier model Astra across eight benchmarks, including GPQA Diamond, BrowseComp, and DeepSWE — though those results don't establish whether writing quality is affected, a question critics have also raised about Claude's watermark. The company notes that a detected watermark reveals nothing about human creativity or editing, doesn't establish ownership or accuracy, and that failing to detect one doesn't prove a human wrote the text. How the rollout lands may hinge on whether the opt-in API approach leads to uneven labeling across markets, and on whether detector access — currently limited to selected researchers and specialist organizations under the EU's Code of Practice — expands as OpenAI says it plans when it believes results can be interpreted responsibly.