Announced October 5, 2026, OpenAI will add an invisible statistical watermark — called textGrain — to text generated by ChatGPT and Codex for EU users, and offer it as an opt-in setting to API developers worldwide, to satisfy the EU AI Act’s transparency rules.
2026-10-05 · by Moe Ameen
OpenAI announced on October 5, 2026 that it will begin adding an invisible watermark to text generated by ChatGPT and Codex, starting with users in the European Union. The company published a technical report for the method, which it calls textGrain, alongside the announcement. The rollout reaches eligible ChatGPT and Codex users on all plans over the coming weeks, but only in the EU. Separately, developers using OpenAI’s API anywhere in the world can turn the watermark on for select models starting the same day — it is off by default.
The mark is not a visible symbol. textGrain works by subtly shaping the model’s word choices as it writes, using a secret key to sort next-word predictions across many small nudges, leaving a statistical pattern a reader cannot see but a detector can pick up. Because the signal lives in the words themselves rather than in file metadata, it travels with the text when it is copied and pasted. The technical report was co-written with researchers from the University of Pennsylvania and Yale, and OpenAI says it plans to release textGrain as open source.
The driver is regulation. The EU AI Act’s transparency rules, which took effect on August 2, 2026 under Article 50, require providers of generative AI to mark machine-generated content in a way other systems can identify. OpenAI opened applications for access to its text-watermark detector, which will initially be limited to approved researchers and expert organizations while the company evaluates and improves the technology.
OpenAI was candid about the limits. Editing erodes the signal: in its own testing, replacing about 10% of the words with synonyms dropped detection from roughly 92% to 66%, and replacing 25% dropped it to 17%. Short passages, math answers, and translated text are harder to detect, and the detector produces both false positives and false negatives, especially on shorter texts or specialized domains like mathematics. The company stresses that the watermark does not measure how much a human contributed, establish ownership, identify a user, or verify accuracy — and that the absence of a detected mark does not prove a human wrote something. Treat the specific figures here as launch-window numbers and confirm details against OpenAI’s own documentation.
There is a fast, high-intent play while this is fresh. "OpenAI is watermarking ChatGPT text — what does it actually mean?" is a question people are searching this week, and most quick recaps blur the detail that matters: the mark is EU-first, it degrades under editing, and the detector is not open to the public. Drop your own clear take into [Kompozy](/) as a source and the engine turns one explainer into a week of coverage in a single pass — a [Blog Article](/glossary/output-buckets) for searchers, a [Carousel](/glossary/hyperframes) that lays out what changes for anyone drafting with AI, a captioned [Persona Short](/glossary/persona-shorts) that explains textGrain to camera, and native text posts — then [Autopilot](/glossary/autopilot) schedules and publishes them across the eight social platforms plus blog and email before the news cools.
The longer game matters more than the recap. As text watermarking spreads from ChatGPT to the rest of the field, the differentiator stops being whether AI touched your copy and becomes whether that copy is distinctive and gets in front of people. That is the job Kompozy is built for: the [Persona Brief](/glossary/persona-brief) governs voice and banned phrases so output reads like you rather than median-prompt AI, and the per-post review gate is where you set a consistent AI-disclosure line once instead of policing it tool by tool. One honest note, because it is the point rather than a loophole: Kompozy generates copy with OpenAI alongside Claude, so text produced through a watermark-enabled model can carry the same signal — the aim is a clean, disclosed body of on-brand work, not a way to strip the mark. For the taxonomy behind all of this, see the glossary on [AI text watermarking](/glossary/ai-text-watermarking).
It is an invisible statistical mark, called textGrain, that OpenAI announced on October 5, 2026. It subtly shapes the model’s word choices using a secret key so that ChatGPT and Codex output carries a pattern a reader cannot see but a detector can identify. Because the signal is in the words, it survives copy-paste. It is rolling out to EU users first and is available as an opt-in API setting worldwide.
For ChatGPT and Codex, yes — at launch the consumer watermark is being added only for eligible users in the European Union, on all plans, to meet the EU AI Act’s transparency rules. Developers using OpenAI’s API can turn it on for select models anywhere in the world, but it is off by default. OpenAI has not committed to a specific global consumer rollout date.
Editing weakens it. In OpenAI’s own testing, replacing about 10% of the words with synonyms dropped detection from roughly 92% to 66%, and replacing 25% dropped it to 17%; short passages, math answers, and translated text are also harder to detect. That fragility cuts both ways — because a heavy rewrite can erase the signal, an absent mark does not prove a human wrote the text.
No. OpenAI is explicit that the watermark does not measure human contribution, establish ownership, identify a user, or verify accuracy, and that a missing mark is not evidence of human authorship. The detector also produces false positives and false negatives, especially on short or specialized text, so it should never be the sole basis for a moderation or academic-integrity decision.