OpenAI is embedding invisible statistical signals into ChatGPT output to meet European Union requirements for labelling AI-generated text.
Unlike Anthropic, which applies watermarking globally, OpenAI will make this feature optional for API customers outside the EU. The company plans to activate textGrain for ChatGPT and Codex users in Europe over the coming weeks. API users worldwide will have the choice to enable or disable the signal.
The technology works by altering word choices to create a detectable pattern. It functions similarly to SynthID, the open-source system used by Google and adopted by Anthropic for its Claude model.
Detection rates drop with edits
OpenAI claims textGrain performed as well as or better than other methods in internal testing, including Google’s SynthID for text. The company intends to release the technology as open source.
Performance depends heavily on length and topic. When the detector is set to a one percent false-positive rate, it identified watermarks in roughly 95 percent of 400-token passages on psychology. That figure fell to about 80 percent at 200 tokens.
Detection was substantially lower for math content because the model has less freedom to choose words. Longer passages might strengthen the statistical signal, though results remain dependent on the subject matter. OpenAI provides no data to test this assumption.
Editing the text makes the watermark harder to find. Replacing 10 percent of words with synonyms cuts detection rates for 400-token passages from about 92 percent to 66 percent. Swapping a quarter of the words drops detection to 17 percent, making the watermark easy to evade.
This weakness could benefit OpenAI. Users who do not want their ChatGPT usage detected might switch to open-weight models if the watermarks become harder to remove.
A technical report explains how textGrain works in depth. Anthropic has not published detection rates for Claude’s watermark, though it says its system holds up well against edits.
OpenAI says watermarking does not hurt output quality, citing tests of its frontier model Astra. The company reports no significant performance differences with watermarking on or off across eight benchmarks, including GPQA Diamond, BrowseComp, and DeepSWE. Those results do not establish whether watermarking affects writing quality, a concern critics have also raised about Claude.
OpenAI states that a detected watermark reveals nothing about the level of human creativity or editing. It does not establish ownership, assign responsibility, identify a user, or prove accuracy. Failing to detect a watermark does not prove a human wrote the text either. The passage could be too short, edited, translated, or come from a model the detector does not support.
Detector access limited initially
Only selected researchers and specialist organisations will initially receive access to the textGrain detector. They must apply via a form. OpenAI will grant access on a case-by-case basis under the EU’s Code of Practice. Anthropic takes a similar approach with its detection API.
The tool reports only whether it detected an OpenAI watermark. It will not identify users or reveal prompts or conversations.
OpenAI says it is restricting access because the detector can flag unmarked text or miss watermarks. The company plans to expand access when it believes results can be interpreted responsibly. It has not said when that might happen.
Existing verification tools for images and audio remain publicly available, including openai.com/verify and the Content Provenance API.
What it means
Creators in the EU will see a label on their AI-generated text without needing to change settings. Users outside Europe retain full control over whether their output carries the signal. The ability to edit text and break the watermark means the technology is not a perfect shield against human modification.
