OpenAI published its approach to new EU rules on the provenance of AI-generated text and immediately rolled it out in its products. An invisible watermark based on the textGrain technology adds a statistical shift in word selection to generation, which is then detected by a separate detector. Starting today, OpenAI API clients worldwide can opt in to enable labeling for specific models, and in the coming weeks the watermark will appear in the output of ChatGPT and Codex, for now only in Europe. The main nuance is that the detector is initially open only to approved researchers and expert organizations, so ordinary users cannot yet see or independently verify the label.

What happened

The company announced its approach to meeting EU requirements for machine-readable labeling of generative text and simultaneously enabled its first elements. An opt-in labeling flag has been working in the OpenAI API since today: it covers specific models and is disabled by default, and cloud partners will receive the same capability in the coming weeks. A watermark in the output of ChatGPT and Codex will also appear in the coming weeks, but will only affect users in the EU on all plans. Along with the announcement, OpenAI released a technical report on textGrain titled Entropy-Calibrated Watermarking for Language Model Text and promised to open-source the implementation over time. In parallel, open verification of images and audio is already available at openai.com/verify, and the Content Provenance API is available for developers, so the new text labeling completes the company's existing provenance suite.

Context

The event should be understood through the regulatory framework: requirements for labeling AI-generated content are enshrined in the European Commission's Code of Practice, which is based on the EU AI Act, and OpenAI became the first major provider to translate these requirements into a product setting rather than a declaration of plans. Technically, this is not about a visible mark, but about a statistical trace: during generation, the model deviates from the usual distribution of word selection, and the detector searches for this shift across the text as a whole. Judging by the name of the technical report, the method is calibrated by entropy, meaning the strength of the watermark is aligned with the degree of randomness of a specific generation — this indicates a thoughtful engineering approach rather than a decorative mark. It is also notable how the company presents the feature: it publishes a technical report, promises open-source code, and directly names the conditions under which detection weakens.

Why this matters for the industry

For the industry, this is a rare case where a regulatory function is obtained almost for free: in the API, labeling is enabled by a configuration flag, without rebuilding the inference pipeline, so product teams can move the option into their product settings right now, and it is beneficial for engineers to design pipelines in advance so that enabling the watermark is a simple toggle. In fact, a new segment of provenance compliance has emerged: EU clients will begin to formulate requirements for labeling generative text in contracts and tenders, and a major provider already offers a ready-made answer. For startups, however, this is an early bet rather than a ready market: detection is technically fragile to rephrasing, and the verification tool is not yet available to businesses, so an independent market for text verification has not emerged. The value for the industry will grow as promises are fulfilled: open-source textGrain code and real researcher access to the detector will provide independent assessments around which libraries, provenance metadata in CMS, and compliance processes can be built.

Why this matters for users

For readers, there are almost no visible changes right now: the label is invisible, and it is impossible to verify it independently because the detector is open only to approved researchers and expert organizations. If you use ChatGPT or Codex in the EU, in the coming weeks your responses will receive an invisible watermark on all plans; for users outside Europe, nothing changes. From this follows an important rule of interpretation: the absence of a watermark does not mean that the text was written by a human — the watermark may simply not be enabled, and the label itself is not designed for visual verification. What can already be used today is the open tool for verifying images and audio at openai.com/verify and the Content Provenance API, if you work with media rather than text.

What is still unknown / limitations

The stated detection accuracy depends on text length: approximately 80% at 200 tokens and 95% at 400 tokens with a false positive rate of 1%, and replacing 25% of words with synonyms drops detection to 17%. OpenAI directly names the weak points: short texts, mathematics, and editing reduce recognition quality. All these figures have so far been verified only by the vendor itself: the detector is closed, and the technical report and the promise of open-source code only partially compensate for this, so there is still no independent confirmation, including the false positive rate. The statement about the minimal impact of labeling on the quality of the flagship model Astra on the GPQA Diamond, BrowseComp, and Terminal-Bench 4.0 benchmarks should also be correctly interpreted as an assessment rather than a fact — no published deltas are available in the materials. Finally, the onboarding of cloud partners, open-source code, and expanded access to the detector remain promises, and latency metrics and accuracy for individual languages have not been disclosed.

Sources

Author

Look at AI, editorial team