OpenAI to watermark ChatGPT text in the EU, but API users can opt out worldwide
The Decoder · rss · 2026-10-06
According to The Decoder, OpenAI is adding invisible TextGrain watermarks to ChatGPT outputs in the EU. Unlike Anthropic's approach, OpenAI will let API customers worldwide opt out of watermarking.
Testing shows detection rates as high as 95%, but the rate drops to 17% when just a quarter of the words are replaced, highlighting the watermark's fragility against paraphrasing.
More from Models
- Why LLMs never pick Java: dev points to under-sampling in training data — HanchungLee · 2026-10-06
- Codex Pro user with 62k credits weighs downgrade after constant Mac crashes — RunnerZee · 2026-10-06
- Comparing Tesla FSD's coverage to Waymo's service area is apples-to-oranges — binarybits · 2026-10-06
- GPT-6.1 ultra speed rumored at ~8x faster, around 550-620 TPS — ChrisGPT · 2026-10-06
- One Midjourney image, one prompt: Opus 5.5 builds a playable game and its trailer — justin_hart · 2026-10-06
- Researcher Laments: Labs Build Interesting Models the Public Will Never Get to Try — rickasaurus · 2026-10-06