OpenAI watermark falls to 17% detection after edits, weakening provenance checks
A regulator-driven transparency feature arrives with narrow access and a signal that modest rewriting can erase.
Rews reports that OpenAI’s textGrain will mark ChatGPT and Codex text by default in the EU, with the detector finding about 95% of 400-token passages at a 1% false-positive rate in the company’s tests. Replacing 25% of words with synonyms cut detection to 17%, while detector access is initially restricted to vetted researchers. For publishers and fact-checkers, the rollout creates a compliance signal but not a dependable way to authenticate edited AI-generated text.