OpenAI Launches textGrain Text Watermarking: Over 90% Detection on 400 Tokens
新智元 · wechat · 2026-10-06
OpenAI has unveiled textGrain, a text watermarking system that subtly biases token selection using a secret key, making AI-written text detectable without any visible difference. API customers can enable it today, with ChatGPT and Codex to follow in coming weeks.
Key facts:
- With a 1% false-positive rate, over 90% of 400-token responses can be detected, reportedly matching or exceeding Google's SynthID;
- On GPT-6Astra, watermarking leaves 8 core benchmarks essentially flat, with 5 slightly improving; user dislike rates are unchanged;
- The detector will only be available to vetted researchers and institutions, while textGrain itself is open-sourced.
Technically, textGrain builds on optimal transport (the 1781 Monge 'earth mover' problem): it sets a 'randomness budget' and uses the Sinkhorn algorithm to shift probability toward key-preferred word groups while remaining unbiased on average across keys. The paper proves the statistical correlation gained by watermarking exactly equals the randomness lost. The 20-page report is led by Weijie Su (Wharton professor who joined OpenAI in May) with Yale's Xiaohong Chen among 9 authors spanning statistics, economics, and ML. All three major labs—Google, Anthropic, OpenAI—now embed provenance into AI text.
Related event: OpenAI launches textGrain text watermarking, starting in the EU(25 posts)→
More from Models
- Google silently disables logprobs on Gemini Hydrogen, Neon, Argon and Carbon models — dejanseo · 2026-11-20
- Bug Hunt chart: score vs subscription cost, Gemini 4 Neon xhigh tops at 63 for $1.80 — PawelHuryn · 2026-11-20
- Measured: Claude Max delivers 46.4x API value on Opus 5.5, beating ChatGPT Pro's 10.25x — PawelHuryn · 2026-11-20
- Dev complains GPT keeps skipping steps and ignoring instructions — lxfater · 2026-11-20
- GPT-6.2 Astra and Sol show OpenAI optimizes performance per dollar, not raw capability: analysis — VraserX · 2026-11-20
- Mamba+MoE hybrid reasoning model Olala-v1.0-mini-7A1B trends on Hugging Face — OVHaiLLM · 2026-11-20