Access to the tool that identifies artificially generated content will be limited to vetted researchers and expert organizations. OpenAI is restricting this detector because its new invisible markers are often undetectable in short passages, and rearranging words or translating the text can remove the signal entirely. In internal tests targeting a 1% false positive rate, the system captured 95% of 400 token psychology passages but only 80% of texts with 200 tokens.
The textGrain system will be deployed for ChatGPT and Codex users in the European Union over the coming weeks to comply with Article 50 of the EU AI Act, with a global opt-in available for API customers today. While the company reports no meaningful quality loss in outputs, the watermarks are vulnerable to simple evasion. Swapping 10% of words with synonyms reduces detection from 92% to 66%, and substituting 25% of the words drops the detection rate to 17%.
