OpenAI will embed an invisible watermark in ChatGPT and Codex text in the European Union, though its own tests show swapping 10% of words cuts detection to 66%.
Key Points:
- The watermark reaches eligible ChatGPT and Codex users on all plans in the EU over the coming weeks and stays opt-in for API customers.
- Replacing 25% of words with synonyms cut detection to 17% in the company's own evaluation.
- Detector access is limited to approved researchers and expert organizations at launch.
OpenAI Watermark Rollout
The company announced the plan Monday in a blog post, saying the marking will reach eligible ChatGPT and Codex users on all plans in the EU over the coming weeks. It will not become a global default at launch.
Developers using the company's API anywhere in the world can turn it on for select models starting now, though it stays off by default. The move answers the EU AI Act, whose marking obligations took effect Aug. 2 and require providers to make AI-generated text identifiable by machines. Companies already on the market have until Dec. 2 to comply.
Also Read: Binance AI Agent Builds Strategies From Plain Words, Charges 19.99 USDC To Trade
TextGrain Detection Limits
The method, called textGrain, adds no visible symbol but subtly shapes the model's word choices, leaving a statistical pattern that a detector can pick up and that travels with copied text. OpenAI said it matched or beat other approaches it tested, including Google's SynthID for text, and that it plans to release the technology as open source.
Editing weakens that signal.
In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words cut it to 17%. Short passages, math answers and translated text are also harder to flag, with the detector catching about 80% of 200-token passages on topics such as psychology at a 1% false positive rate.
ChatGPT Detector Access
OpenAI is limiting its detector to approved researchers and expert organizations for now, citing the risk of missed watermarks and false positives. It also cautioned that a missing watermark "does not prove human authorship," and said the mark does not identify the user, the account or the prompt behind a passage.
The step follows Anthropic, which said in August that it would watermark text generated by Claude worldwide, a decision that drew complaints from some users. OpenAI had built a text watermark before but held it back, partly over concerns that users would switch to rivals, according to a 2024 news report. Both companies, along with Google, Meta and Microsoft, have committed to the EU's code of practice on AI-generated content.
Read Next: Giorgia Meloni Wants A Trademark On Her Voice, Will It Stop AI Deepfakes?
