OpenAI to watermark ChatGPT text in EU under AI Act rules
OpenAI, a signatory to the EU’s AI-content code alongside Google, Anthropic, Meta, Microsoft and Mistral, will watermark ChatGPT text in the EU; tests show the watermark can weaken sharply after editing
OpenAI plans to add an invisible watermark to ChatGPT and Codex text generated for users in the European Union, in response to AI Act transparency rules. The watermarking feature, called textGrain, will roll out over the coming weeks for all EU ChatGPT and Codex plans. Initially, the watermark will not be a global default. The invisible statistical signal in the model's word choices will be analyzed by a detector to verify the watermark's presence.
OpenAI's detector has shown an 80% watermark identification rate in 200-token passages and a 95% rate in 400-token passages, with a 1% false-positive rate. However, the watermark's effectiveness is significantly reduced when content is edited or translated, or when using unsupported models or models that predate watermarking. OpenAI has stated that the watermarking system is not intended for definitive AI authorship tests, but rather as a provenance signal.
The company plans to make textGrain available as open-source technology and aims to incorporate text watermarking into a broader provenance system that includes technologies like Content Credentials and SynthID.
Written by urgent.news from The Hindu - Sci-Tech's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
