OpenAI’s textGrain Watermark: What EU Deployers Must Know Right Now
In its October 5, 2026 announcement, OpenAI described textGrain as an invisible statistical watermark based on model word choice. The company said eligible ChatGPT and Codex output in the EU would receive it through a staged rollout, and that selected API customers worldwide could opt in for eligible models, with watermarking off by default. These are dated product statements; verify current availability on publication day.
For deployers, the legal distinction matters. Under EU AI Act Article 50, providers of in-scope generative systems have a machine-readable-marking duty for synthetic content, subject to the provision’s scope and exceptions. Deployer disclosure duties are separate and apply in specified cases—not to every AI-generated text.
What textGrain actually is
textGrain embeds a statistical signal in model word choice, which a detector can identify. In its October 5 announcement, OpenAI said initial detector access would be by application for approved researchers and expert organisations. OpenAI says the detector reports whether an OpenAI watermark was detected; it does not identify users, reveal prompts or expose conversations. See the announcement and the technical report.
Do not assume every output is marked. OpenAI’s announcement says watermarking is off by default for eligible API models and that selected API customers may opt in. Product settings and rollout can change, so check the model and configuration used in each workflow. Provider comparisons should be model-specific and current; a statement about one product or planned rollout does not establish a blanket policy across a provider’s access routes.
The limits are the story
Here is what deployers must understand. OpenAI’s own testing shows that detection performance depends on the text and the test conditions.
At a target 1 percent false-positive rate, OpenAI reports detection in about 95 percent of 400-token psychology passages and about 80 percent of 200-token psychology passages. Its report shows substantially lower detection for mathematical passages; results vary with text type and length. See the technical report for the test conditions.
Editing can weaken the signal. In OpenAI’s reported 400-token synonym-substitution test, detection fell from 92 percent to 66 percent after replacing 10 percent of words, and to 17 percent after replacing 25 percent. That shows reduced detection under those conditions; it does not mean every light paraphrase defeats the watermark.
And detection is only half the story. OpenAI says a detected watermark does not establish human contribution, authorship or ownership, responsibility, or accuracy. A missing watermark does not prove human authorship either: the text could be too short, edited, translated, or generated by a model the detector does not cover. See OpenAI’s explanation of the limits.
What this means for your AI Act compliance
If you are building deployer controls, three takeaways.
First, watermarks are a signal, not proof of authorship or compliance. Do not build a provenance process on watermark detection alone. It is one technical input, not a substitute for the duties that apply to your workflow.
Second, distinguish provider from deployer obligations. Article 50’s machine-readable-marking duty is for providers of in-scope systems. The deployer disclosure duty for AI-generated or manipulated text applies when it is published to inform the public on a matter of public interest; that text duty does not apply where the content has undergone human review or editorial control and a person holds editorial responsibility. Deepfake disclosure is a separate deployer obligation. Keep records as part of governance, but do not treat recordkeeping alone as proof of compliance or as a replacement for required disclosure. See Article 50.
Third, check the opt-in question. OpenAI says selected API customers may opt in to watermarking for eligible models, and that watermarking is off by default. If your governance model assumes generated content is labeled, verify the actual settings and rollout for every model and workflow rather than assuming the signal is present.
The bigger debate
A watermark cannot settle who wrote a text, what human contribution exists, who is responsible, or whether the content is accurate. Treat it as a limited provenance signal and ground policy in the legal duties and documented processes that apply to your use case.
For deployers, the practical work is documentation, the specific disclosures required for a given workflow, and defensible processes. The distinction between provider marking and deployer disclosure matters more than assuming one watermark solves both.
The bottom line
textGrain may help identify some eligible OpenAI-generated text, but OpenAI’s published tests show that detection varies by text type, length, and editing. It is not standalone evidence of authorship or compliance. Map duties by role, content, and workflow, and recheck product availability before relying on the control.
About Author
