OpenAI Ships textGrain Watermark: 95% Detection at 400 Tokens, but Editing Kills It
johnseach · x · 2026-10-06
OpenAI launched textGrain on Oct 5, 2026: an invisible statistical watermark on EU ChatGPT/Codex output, opt-in for API users, default off, detector gated to approved researchers.
- At a 1% false-positive target: 80% detection on 200-token passages, 95% at 400 tokens for flexible prose; low-entropy text like math does far worse.
- Swapping 10% of words drops detection from 92% to 66%; 25% drops it to 17%. Translation, another model, or short answers often erase it.
- Quality on GPT-6 Astra was broadly unchanged.
It's EU AI Act compliance, not a detection breakthrough: a positive is weak evidence of OpenAI origin, while a negative is near-noise — institutions treating "no watermark" as human authorship will err often, favoring those who edit or use rival models.
Related event: OpenAI Rolls Out textGrain Text Watermarking, Starting in the EU(15 posts)→
More from Models
- Microsoft page reportedly confirms GPT-6 series uses Looped Transformers — teortaxesTex · 2026-10-06
- Amazon's ALoDLM: Token-Adaptive Looped Diffusion LMs Beat AR Baselines — amazon · 2026-10-06
- Looped LMs at Fixed Points: 3x Smaller KV Cache, 1.79x Faster Prefill — IFM · 2026-10-06
- Australia open-sources Matilda Jev, a 56.8ms decision model that skips text generation — Med1_Ai · 2026-10-06
- Codex politely pushes back on a user's hallucinated claims — "I'm also a journalist" — MikePFrank · 2026-10-06
- Gemini 4 Argon hallucinates half as often as rival frontier models on AA-Omniscience — zacharynado · 2026-10-06