OpenAI's textGrain watermark: EU-only rollout, no public detector, and synonym swaps drop detection to 17%
Imaginary_Dinner2710 · reddit · 2026-10-06
OpenAI published its ChatGPT text-watermarking plan (textGrain) on October 5, and a Reddit user summarized what it says and omits:
- How it works: no hidden characters; the model's word choices are subtly nudged, and a key-holding detector spots the statistical pattern. Copy-paste or Unicode cleaning won't remove it.
- Rollout: coming weeks for EU ChatGPT and Codex users on every plan; other regions excluded at launch; API stays off unless developers opt in.
- No self-checking: the detector is open only to approved researchers; OpenAI's public checker covers images and audio only. Commercial "AI detectors" are style-based and hold no key.
- Official numbers: 95% detection on 400-token answers at a 1% false-positive rate (lower for math); replacing a quarter of words with synonyms cut detection from 92% to 17% in OpenAI's own ideal-condition test.
- Context: Claude has done the same worldwide since August; Gemini's app since 2024.
Related event: OpenAI Rolls Out textGrain Text Watermarking, Starting in the EU(15 posts)→
More from Models
- Will OpenAI Merge Chat and Codex Quotas? Users Fear Tighter Basic Chat Limits — Iwantthegreatest · 2026-10-06
- 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