OpenAI introduces textGrain text watermarking: on by default for ChatGPT and Codex in the EU, opt-in for the API worldwide
On Oct 5, 2026 OpenAI published its approach to the EU AI Act's text-provenance rules (Article 50). Over the following weeks it will add an invisible statistical watermark, textGrain, to eligible ChatGPT and Codex text for users in the EU only, on all plans. API customers worldwide can opt in from the same day (off by default). Its detector goes only to approved researchers and expert organizations, because detection is unreliable on short, edited or low-entropy text such as mathematics.
Key facts
- Method: textGrain 'adds an invisible statistical signal to the model's word choices'; a detector checks for it. OpenAI published a technical report ('textGrain: entropy-calibrated watermarking for language model text') and plans to open-source the technology
- OpenAI says textGrain matched or exceeded the other approaches it tested, including Google's SynthID for text
- Detection at a 1% target false-positive rate: ~80% of 200-token and ~95% of 400-token passages for psychology-type content, 'substantially lower' for mathematics. Replacing 10% of words with synonyms cut detection from ~92% to 66%; replacing 25% cut it to 17%
- No meaningful quality cost on Astra (max) benchmarks, e.g. Artificial Analysis Intelligence Index 49.57 vs 49.76 watermarked; GPQA Diamond 94.44% vs 93.94%; Terminal-Bench 4.0 53.90% vs 56.06%
- Scope: ChatGPT and Codex text in the EU only (no global default at launch); API opt-in for 'select models' worldwide from Oct 5; cloud partners to follow 'in the coming weeks'
- Detector access: applications open to approved researchers and expert organizations, granted case by case under the EU Code of Practice on AI-generated content; the tool reports only whether an OpenAI watermark is detected, not the user or prompt. Image/audio verification (openai.com/verify, Content Provenance API) stays public
- OpenAI lists what a watermark does not show: amount of human contribution, ownership, the user's identity, accuracy; a missing watermark does not prove human authorship
- OpenAI's X announcement had ~548k views within hours (fxtwitter)
- Context: Article 50 obligations applied from Aug 2, 2026, with a grace period to Dec 2, 2026 for watermarking (Digital Omnibus); Anthropic began watermarking text from new Claude models in August 2026 and Google has offered SynthID-Text since 2024
What happened
OpenAI had already watermarked images and audio (C2PA Content Credentials and SynthID) and offered public verification. On Oct 5 it extended provenance to text, describing it as a response to the EU AI Act's requirement that generated text be machine-identifiable. It rolls out in phases: opt-in in the API now, then automatic watermarking of ChatGPT and Codex output in the EU over the following weeks, and restricted detector access. OpenAI was open about the limits: detection drops sharply on short or edited passages, and OpenAI says it will revisit the approach as evidence and standards develop. The Verge, PCMag and PYMNTS covered the move as "watermarks, but only in the EU".
Why it matters
In 2024 OpenAI said it had a text watermarking method but held it back, partly because it could "stigmatize use of AI as a useful writing tool for non-native English speakers". Now the world's most-used chatbot watermarks text in one jurisdiction because the law requires it. That creates a regional split in what ChatGPT outputs carry, and it publishes concrete robustness numbers showing that simple paraphrasing defeats most detection.
Changelog
- 2026-10-05: created (sweep 2026-10-05; OpenAI post read in full via a reader proxy, X post reach via fxtwitter)
Related posts (1)
- OpenAI: expanding content provenance to text; EU ChatGPT/Codex watermarking, API opt-in worldwide original ↗ OpenAI @OpenAI · x · 2026-10-05
OpenAI's announcement of textGrain text watermarking under the EU AI Act; ~548k views within hours.
Related events
- EU AI Act 'Digital Omnibus' in force: high-risk rules delayed to Dec 2027, GPAI enforcement starts Aug 2 ★★★★
- EU AI Act enters into force ★★★★★
- Anthropic releases Claude Opus 5.5 — Fable-5.1-level performance at $4/$20, first model of the Claude 5.5 family ★★★★★
- OpenAI releases GPT-6 Astra, its first GPT-6 model ★★★★★
Sources (8)
- officialOpenAI: Our approach to EU text provenance rules
- paperOpenAI: textGrain technical report (PDF)
- officialOpenAI on X: expanding content provenance to text
- discussionTechmeme discussion
- docsOpenAI Help Center: provenance signals in OpenAI-generated content
- pressCrypto Briefing: OpenAI adds optional text watermark API feature to meet EU AI Act rules
- officialOpenAI (2024): Understanding the source of what we see and hear online (text watermarking held back)
- paperarXiv: Watermarks Without Verification: AI Text Watermarking After the EU AI Act
id: 2026-10-05-openai-textgrain-eu-text-watermarking · updated 2026-10-05 · open in the interactive timeline