OpenAI will automatically watermark ChatGPT-generated text in the EU starting in compliance with the EU AI Act, while the feature will be opt-in elsewhere. The proprietary textGrain watermark uses subtle word-choice patterns detectable at roughly 92% accuracy, but its reliability drops sharply when output text is modified.
Background
The EU AI Act, which took effect in August 2026, requires AI providers to mark generated content in a machine-readable way. Watermarking LLM output remains an unsolved technical challenge with no universally reliable solution.
- Source
- Ars Technica
- Published
- Oct 7, 2026 at 04:50 AM
- Score
- 7.0 / 10