OpenAI says it will add an invisible statistical watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks. The plan is not a completed rollout, and the detection results the company has published have not been independently tested in the reporting reviewed.
The system, which OpenAI calls textGrain, alters statistical patterns in a model’s word choices. A detector then looks for the resulting signal in a passage. OpenAI has not specified which models or kinds of output will qualify for the EU rollout.
What has OpenAI announced for the EU?
In an announcement published Monday, Oct. 5, OpenAI said eligible ChatGPT and Codex text output in the EU would receive watermarks over the coming weeks. The company has not said the rollout is complete, or defined eligibility by model, plan, output type or user circumstance.
OpenAI described two related steps with different reach. Customers using its API worldwide can opt in to text watermarking for select models, and the feature will be off by default. The company said API opt-in would begin Monday, the same day it would start accepting applications for detector access.
The detector will not initially be open to the public. OpenAI says approved researchers and expert organizations will be able to apply, and that it is working with a limited group of researchers to study the method’s limits. The announcement does not report independent findings from those researchers.
The EU’s AI Act requires providers of covered generative AI systems to mark certain synthetic content in a machine-readable way and make it detectable, subject to technical feasibility and exceptions. European Commission guidance says those obligations began applying Aug. 2, 2026. Systems placed on the market before that date have until Dec. 2, 2026, to meet the marking and detection requirement.
How well does textGrain detect generated text?
OpenAI reports that textGrain detected watermarks in about 80% of 200-token psychology passages and about 95% of 400-token psychology passages, with a target false-positive rate of 1%. Those are results from the company’s evaluations, not a general accuracy rate for everyday text.
The reported results varied with the material. OpenAI says detection was weaker for shorter or constrained passages and substantially lower in its mathematics tests than in its psychology tests. The company has not established how the method performs across ordinary user text, languages, genres or mixed human-and-model writing.
Editing also reduced detection in OpenAI’s evaluation of 400-token passages. The company reported that replacing 10% of words with synonyms lowered detection from about 92% to 66%; replacing 25% lowered it to 17. These figures describe those tests, not a guarantee of how the detector will respond to every edited passage.
OpenAI says its results compared favorably with approaches including SynthID for text, but that is the company’s own comparison, not an independent head-to-head assessment. OpenAI also cautions that laboratory performance does not ensure reliable results in everyday use. Its technical report says the theoretical false-positive calculation depends on assumptions and that the deployed method needs empirical calibration checks.
An Engadget report describes the rollout as a planned measure but does not provide an independent test of textGrain. The sources reviewed do not establish that outside researchers have reproduced OpenAI’s detection rates or tested them on typical EU user output.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Does the plan include Codex source code?
OpenAI names eligible Codex text output, but it has not explained whether that includes source code or which Codex responses would qualify. The mention of Codex alone does not establish that all code it generates will carry a watermark.
That distinction matters for the legal requirement. European Commission guidance lists source code among the output types outside Article 50(2)’s marking obligation. OpenAI has not said whether it intends to watermark source code voluntarily, or how textGrain would perform on code or passages combining code and prose.
The announcement therefore leaves two questions separate: what output OpenAI will choose to watermark, and what output the EU rule requires to be marked. The company has not yet provided enough detail to settle the first question for Codex.
What does a text watermark say about provenance?
OpenAI presents provenance as a collection of signals and verification methods, rather than a single test that establishes who created content. For supported image outputs, the company says it uses Content Credentials and watermarks; its broader approach to images and audio also includes verification tools.
A positive text detection, OpenAI says, indicates a provenance signal associated with its systems. It does not identify the user, measure a person’s contribution, establish authorship or ownership, verify accuracy, or determine legal responsibility. A detector that finds no signal does not prove a person wrote the passage.
OpenAI says detector access will be restricted initially because the system can miss watermarked text or flag other text incorrectly. The company has also said it plans to make the technology open source, but the announcement does not give a date for that release.
The key test is whether independent evaluators can reproduce the company’s results under varied conditions, including short, edited, translated and mixed-source passages. Until then, the published percentages describe OpenAI’s own tests, while the exact scope of the EU rollout—and whether Codex source code is included—remains unsettled.