OpenAI rolls out textGrain invisible watermarking for ChatGPT and Codex in the EU

OpenAI announced on October 5, 2026 that it will embed invisible watermarks in text produced by ChatGPT and Codex for eligible users inside the European Union within the coming weeks, while opening an opt in watermarking option to API customers worldwide. The method behind the rollout is called textGrain, and OpenAI describes it as a statistical signal woven into the words a model chooses rather than a visible label attached to them.

The timing is not accidental. Article 50 of the EU AI Act places transparency duties on providers of generative AI, and those obligations started to apply on August 2, 2026. They require companies to mark AI generated content in a way that other systems can recognize. AI Affairs reported on October 5 that a provider that breaches those transparency duties can face a fine of up to 15 million euros or 3 percent of global annual revenue.

Watermarks of this kind do not appear as symbols or notices in an interface. They work by subtly influencing which words a model selects, leaving a pattern that a reader cannot perceive but a detector can identify. Because the pattern lives in the text itself, it survives copying and pasting and travels wherever the output is reproduced.

OpenAI published a technical report the same day under the title textGrain, subtitled Entropy Calibrated Watermarking for Language Model Text, co authored with researchers from the University of Pennsylvania and Yale University. The company says it intends to release the method as open source. OpenAI also began accepting applications for its text watermark detector, initially limited to approved researchers and expert organizations, granted case by case under the EU Code of Practice.

Key Facts

TechCrunch reported on October 5 that OpenAI announced the plan in a blog post that day, describing watermarks that will reach every tier of eligible ChatGPT and Codex users in the European Union over the next few weeks. Global API developers can switch the feature on for selected models immediately, but it stays off by default, and OpenAI said it will not make watermarking a global default at launch.

According to OpenAI, the detector only reports whether an OpenAI watermark is present. It does not identify users, and it does not reveal prompts or conversations. The company also states that the watermark does not measure human contribution, does not settle ownership or liability, does not verify accuracy, and that the absence of a detected watermark is not proof that a person wrote the text.

Numbers published with the launch sketch both the reach and the limits of the approach. At a 1 percent target false positive rate, detection reached about 80 percent for 200 token passages and about 95 percent for 400 token passages on psychology related content. Detection fell sharply in domains such as mathematics, where word choice is constrained. Editing weakens the signal as well: in a 400 token passage test, replacing 10 percent of words with synonyms cut detection from about 92 percent to 66 percent, and replacing 25 percent dropped it to 17 percent.

On quality, 9to5Mac reported on October 5 that watermarked and unwatermarked text scored close to each other across several benchmarks. OpenAI cited an Artificial Analysis Intelligence Index of 49.76 with watermarking against 49.57 without, Terminal Bench Science 0.1 at 60.00 percent against 56.90 percent, and GPQA Diamond at 93.94 percent against 94.44 percent. OpenAI ran those evaluations on Astra, which it calls its newest frontier model.

Unite.AI reported on October 5 that the transparency code of practice tied to the AI Act was drafted by independent experts in a multi stakeholder process convened by the AI Office, with the final text published on June 10, 2026. Signing is voluntary, but Article 50 remains a legal obligation, and by the end of July 2026 roughly 190 companies and organizations had signed. Anthropic, Google, Meta, Microsoft and OpenAI have all committed to follow the EU code on AI generated content.

Analysis

What this really means is that provenance labeling has moved from a promise made at conferences into a compliance deadline with money attached. OpenAI has chosen a phased, geography first design: the EU product surface gets watermarks automatically, while the global API surface gets an opt in switch that stays off. 9to5Mac reported on October 5 that the API setting is not enabled by default, and that ChatGPT and Codex watermarking initially covers only eligible users in the European Union.

The design also reflects an honest reading of the science. OpenAI itself calls text watermarking and detection an early technology with significant limitations, and the published figures support that framing. Detection holds up on long passages of flexible prose and weakens on short passages, on mathematical answers and on translated text, and it degrades quickly under ordinary editing. A watermark that a writer can weaken by swapping one word in ten is a provenance signal, not a provenance guarantee.

The bigger picture here is that the enforcement mechanism and the verification mechanism are misaligned. Watermarking will be applied to EU outputs automatically, yet access to the detector is gated to approved researchers and expert organizations, granted case by case. AI Affairs reported on October 5 that researchers, platform teams and news organizations therefore cannot routinely verify outputs on their own. Publishers that want to know whether a document came from an OpenAI model must apply for permission rather than run a check.

The competitive context matters too. TechCrunch reported on October 5 that Anthropic announced watermarking for Claude text two months earlier, applied globally, and that the move drew a backlash from some users. The OpenAI approach is narrower by comparison. Unite.AI reported on October 5 that OpenAI is working with cloud partners to watermark model outputs inside their services, and OpenAI says textGrain matched or beat every method it tested, including Google SynthID for text.

Why It Matters

For anyone who publishes, moderates or fact checks online text, this is an early large scale test of whether machine readable provenance can work in the wild. The benefit is real when it works: a detector that flags an OpenAI watermark gives a platform one more signal in a world where synthetic text is cheap and abundant. That is the outcome Article 50 was written to produce.

The risk sits on the other side of the same coin. A missing watermark proves nothing, and OpenAI says as much, so it cannot be used to accuse a writer of faking human authorship. A false positive is worse, because it could be used to dismiss legitimate writing as machine generated. With detection at roughly 80 percent on 200 token passages before any editing, and with the detector limited to a small set of approved institutions, the public has little independent means to measure either error rate.

The timeline reaches beyond Europe as well. Article 50 became applicable on August 2, 2026, and the OpenAI rollout follows within about two months. Other providers carry the same legal duty, and the code of practice already had about 190 signatories by the end of July 2026. What begins as a regional compliance step can harden into a global default if the tooling proves cheap and the quality cost stays small, and the benchmark numbers suggest the quality cost is small today.

Next Up

Three things are worth watching. First, whether the EU rollout genuinely arrives across every tier in the coming weeks, and whether OpenAI extends the setting beyond Europe or keeps it regional. Second, whether detector access widens beyond approved researchers and expert organizations, because outside verification is the only way the 92 percent to 66 percent editing figure and the 80 percent and 95 percent detection figures get tested by people who did not produce them.

Third, whether the promised open source release of textGrain materializes, and whether providers converge on one watermark standard or ship competing ones that platforms must each integrate. OpenAI already offers verification tools for images and audio through its verification site and its Content Provenance API, so the company is assembling a multi modal provenance stack rather than a single text feature. The coming months will show how much of that stack regulators, publishers and ordinary readers can actually inspect.