OpenAI launches textGrain text watermarking, starting in the EU
On October 5, OpenAI announced an expansion of its provenance efforts to text for the first time, with a scheme called textGrain, to comply with Article 50 of the EU AI Act. API customers worldwide can opt in for select models immediately, and within weeks eligible text from ChatGPT and Codex in the EU will automatically carry an invisible watermark. This is the company's first systematic public roadmap for text watermarking.
Confirmed
- Effective immediately, global API customers can opt in to text watermarking for select models; within weeks, eligible ChatGPT and Codex outputs in the EU will carry invisible watermarks automatically
- The watermark works by applying a keyed statistical bias during token sampling—sometimes choosing near-optimal words instead of the best one—embedding an invisible statistical signal detectable by a detector; it adds no visible marks or special characters and survives copying
- OpenAI admits limitations: detection is hard on short text, and rewriting or translation can fully remove the watermark, so the detector is currently limited to approved researchers while the technology is evaluated and improved; OpenAI treats text watermarking as an ongoing research area
- Details relayed by johnseach indicate roughly 95% detection at 400 tokens, but the watermark breaks under paraphrasing; ImaginaryDinner2710 notes an open-source rewrite-based removal tool already exists
- OpenAI stresses the watermark answers only one question—whether text was likely generated by an OpenAI model—and claims no impact on model performance or user identity; Angaisb confirmed in an explainer video that reply quality is unaffected
- The scheme launches in the EU first rather than as a global default, with OpenAI saying the regional approach leaves room to learn and adjust
- Andrew Curran (relayed by generativist) notes that unlike Anthropic and Google, OpenAI is enabling this only in the EU for now; ImaginaryDinner2710 notes Anthropic already watermarks some Claude models globally, making OpenAI a follower
- This extends OpenAI's existing image and audio provenance tools
Unconfirmed
- Conflicting accounts on detector access: MartinSignoux relayed that OpenAI will open-source the detector, while OpenAI's own notes say it is limited to vetted researchers
- ns123abc speculates the EU-only phasing is a prelude to global rollout, following Anthropic's pattern; OpenAI has not confirmed any timeline beyond the EU
- The textGrain paper has not been published; technical details remain subject to the eventual paper
Why it matters
- This is the first large-scale deployment of verifiable text provenance by a major AI vendor, showing that the EU AI Act's transparency requirements now have real teeth
- The invisible watermark is claimed not to hurt output quality, but ResearchCrafty1804 argues the injected statistical signal implies a KLD>0 bias that could degrade quality cumulatively; VraserX calls the EU mandate a bad idea, arguing detectability necessarily costs generation quality; Tansu Yegen contends text watermarks degrade with editing and long-term solutions may lie in provenance certificates instead
- The rewrite-to-remove weakness and the appearance of open-source removal tools highlight how easily the mechanism can be evaded; the phased rollout suggests OpenAI is still watching for false positives and abuse, with expansion beyond the EU dependent on pilot results
2026-10-05 ~ 2026-10-07 · 25 related posts
Primary sources
- [source] OpenAI rolls out textGrain text watermarking for EU AI Act, going open source — btibor91 · 2026-10-05
- OpenAI explains how it will watermark ChatGPT text to comply with EU provenance rules — rhiever · 2026-10-06
- [source] OpenAI to watermark ChatGPT text outputs to comply with EU AI Act — OpenAI · 2026-10-06
- [source] How OpenAI's text watermark works: invisible statistical signals, no performance hit — OpenAI · 2026-10-06
- OpenAI admits text watermarks are fragile — detector restricted to approved researchers — OpenAI · 2026-10-06
- OpenAI launches textGrain watermark for AI text in the EU, with no quality impact — Angaisb_ · 2026-10-06
- OpenAI's EU-only text watermarking may be a prelude to global rollout, blogger argues — ns123abc · 2026-10-06
- OpenAI adopts textGrain text watermarking from Penn/Yale researchers — weijie444 · 2026-10-06
- OpenAI rolls out text watermarking in ChatGPT and Codex, EU-only for now unlike Anthropic and Google — generativist · 2026-10-06
- OpenAI's textGrain watermark: EU-only rollout, no public detector, and synonym swaps drop detection to 17% — Imaginary_Dinner2710 · 2026-10-06
- Does OpenAI's Watermarking Degrade Model Quality? Reddit Debates the KLD Cost — ResearchCrafty1804 · 2026-10-06
- OpenAI launches invisible text watermarking for EU, unveils textGrain technique — 机器之心 · 2026-10-06
- EU's push to watermark AI writing is a bad idea that will hurt output quality, argues dev — VraserX · 2026-10-06
- OpenAI follows Anthropic with text watermarks; dev open-sources rewrite-based remover — Imaginary_Dinner2710 · 2026-10-06
- Text watermarking is fragile by design; provenance may be the real fix, argues Tansu Yegen — TansuYegen · 2026-10-06
10 near-duplicate retellings: rhiever · ns123abc · XFreeze · Polymarket · MartinSignoux · johnseach · rohanpaul_ai · 新智元 · alex_verem · paulnovosad