OpenAI has begun marking ChatGPT and Codex text with an invisible watermark for EU users, framed as AI Act compliance. The mark is evidence about the provider’s system. Whether a company must disclose AI-generated text on matters of public interest turns on something the mark cannot show: documented human review and a named person with editorial responsibility.
What changed
On 5 October 2026 OpenAI announced textGrain, an invisible, machine-readable watermark for text output from ChatGPT and Codex, presented as compliance with the EU AI Act transparency rules that took effect on 2 August 2026. The announcement was reported by TechCrunch, with The Verge and BleepingComputer also covering it. A technical report on the watermark was co-authored with researchers at the University of Pennsylvania and Yale.
Rollout to eligible users in the EU, across all plans, is set for the coming weeks. For API developers worldwide the watermark is opt-in and off by default, and it has been available since the day of the announcement.
OpenAI set out the limits of the technique in its own terms. Replacing 10 percent of words with synonyms dropped detection from 92 percent to 66 percent. Detection is difficult in short passages, in math answers and in translated text. The company stated that the watermark does not prove human authorship, and that watermarks can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.
Why it matters
The AI Act, Regulation (EU) 2024/1689, draws a clean line between two actors. Article 50(2) places a duty on providers of AI systems that generate synthetic text: outputs must be marked in a machine-readable format and detectable as artificially generated, as far as technically feasible. Article 50(4), second subparagraph, places a separate duty on deployers of an AI system that generates or manipulates text published to inform the public on matters of public interest: that text must be disclosed as artificially generated, unless the content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for its publication. Provider duties and deployer duties are separate duties.
A watermark speaks to the first duty and is silent on the second. It can indicate that a system generated or processed part of a passage. It cannot show how much human judgment, editing or creativity went into that passage, and it cannot prove human authorship. The exemption from deployer disclosure turns on exactly what the watermark cannot show: a documented process of human review or editorial control, and a named person or legal entity holding editorial responsibility.
The watermark also degrades under ordinary editing. Synonym substitution at a modest rate cuts detection sharply. Translation and short passages weaken it further. Editing is the very activity that creates the exemption from disclosure. Once the text has been worked, a degraded mark says little about anything. Neither a present mark nor an absent mark settles the deployer question.
Two operational gaps compound this. In-house text pipelines fed through APIs sit outside whatever the consumer products now do, because the watermark is opt-in and off by default for API developers worldwide. And a company that treats provider marking as its compliance answer has outsourced a deployer duty to a tool built for a provider duty.
What changes for management
The practical task is to build and maintain an editorial-control record for published text. At minimum, the record should capture which published text is AI-assisted, who reviewed it, who holds editorial responsibility for its publication, and what is retained as evidence of the review process. That record is what the Article 50(4) exemption rests on, and it is the only artefact that speaks to human judgment.
Classification comes first. Text published to inform the public on matters of public interest needs to be identified as a category, with a documented basis for the boundary decisions. Within that category, AI-assisted text needs to be flagged at the point of drafting or generation, before editing histories blur the trail.
Review and editorial control need to be defined as a process with named participants. “Human review” and “editorial control” carry weight in the regulation only where a natural or legal person holds editorial responsibility for publication. That person or entity should be named in the record for each piece of published text, and the review steps should be described in enough detail to demonstrate the process existed.
Retention matters as much as capture. The evidence of review, the identity of the responsible person, and the AI-assistance flag need to be kept in a form that can be produced later. Watermark detection results, where they exist, can be recorded as one data point about provenance. They cannot substitute for the editorial record, and their absence does not imply human authorship.
API-based generation needs separate attention. Where text pipelines call models through APIs, the watermark is off by default and opt-in. Any reliance on provider marking in those pipelines has to be checked against the actual configuration, and the deployer record has to stand on its own regardless.
What changes for the board
The board’s oversight target shifts from tooling to records. The question of whether a provider marks its outputs is a procurement and vendor-management matter. The question of whether the company can show human review or editorial control with a named responsible person is a governance matter, because it determines whether disclosure is required for text published on matters of public interest.
Boards should expect management to present the editorial-control record as a governed artefact: its scope, its owners, its retention rules, and its coverage of published text. Spot checks and periodic reporting on coverage are reasonable expectations. Where AI-assisted text is published without a recorded review and a named responsible person, the disclosure duty under Article 50(4) applies, and the board should know how often that arises.
Vendor assurances about marking should be treated as information about the vendor’s system. OpenAI’s own statement is that watermarks can indicate generation or processing but not the amount of human judgment, editing or creativity involved. That is a description of provenance evidence, and provenance evidence does not discharge a deployer duty. Boards should probe any management position that treats provider marking as covering the company’s disclosure exposure.
The exemption pathway deserves board-level clarity. Documented human review or editorial control with a named responsible person is the route out of disclosure for public-interest text. Boards should ask whether that route is being used deliberately, with records to support it, or assumed by default.
Questions boards should now ask
- Which published text falls within “matters of public interest” under Article 50(4), and who owns that classification?
- For AI-assisted text in that category, is there a recorded process of human review or editorial control, and is a natural or legal person named as holding editorial responsibility for publication?
- What is retained as evidence of the review process, for how long, and can it be produced on request?
- Where text is generated through API pipelines, is provider watermarking on or off, and does the disclosure position depend on that setting?
- Does any part of the compliance case rest on the presence or absence of a watermark, given that editing degrades detection and the mark cannot prove human authorship?
- How is AI assistance flagged at the point of generation, before editing obscures the trail, and does the flag survive the editing workflow?
- What reporting does the board receive on coverage of the editorial-control record, and on instances where AI-assisted public-interest text was published without recorded review?
The governance consequence is narrow and firm. Watermarking addresses a provider duty under Article 50(2); disclosure is a deployer duty under Article 50(4), and the exemption turns on documented human review or editorial control with a named responsible person. A record of which published text is AI-assisted, who reviewed it, who holds editorial responsibility and what evidence is kept is the artefact the board can govern. Where that record is thin, disclosure is the default position, and provider marking offers no defence.