OpenAI starts textGrain text watermarking in ChatGPT and Codex for EU users

OpenAI has started rolling out an invisible, machine-readable watermark for text generated by ChatGPT and Codex, beginning with users in the European Union. According to The Verge, the system, known as textGrain, will reach eligible users on all plans across the EU over the coming weeks, while OpenAI has deliberately stopped short of making watermarking a global default at launch.

An EU-first rollout tied to the AI Act

OpenAI frames the regional approach as a way to learn from real-world use and feedback before deciding what comes next. The timing is tied to transparency obligations under the EU's AI Act. Anthropic made a similar move in August, announcing watermarking based on Google DeepMind's SynthID for the same regulatory reasons, The Verge notes.

API customers are treated differently: starting immediately, they can opt in to watermarked outputs for select models anywhere in the world. OpenAI says it is also working with cloud partners so that watermarking becomes available for its model outputs accessed through partner services in the coming weeks.

Detection stays limited to vetted researchers

The tool for detecting the watermark will not be public at launch. Instead, approved researchers and expert organizations can apply for access starting now, with grants handled case by case under the EU's Code of Practice. According to OpenAI, the detector reports only whether it finds an OpenAI watermark, and does not identify the user or reveal prompts or conversations.

The company points to the risk of missed watermarks and false positives as the reason for the restriction, positioning early access as a way to support evaluation and improvement of text provenance rather than consumer-grade verification.

What the watermark does and does not prove

OpenAI claims textGrain "matched or exceeded" competing approaches, including SynthID for text, and published benchmark results suggesting watermarked and unwatermarked text perform similarly. But the company is blunt about the limits. TextGrain "does not guarantee reliable detection," OpenAI notes, and text watermarks more broadly cannot verify accuracy, determine who owns a piece of text, measure how much a human contributed, or prove human authorship.

The Verge also reports that the announcement was not universally welcomed, although the mechanics of the feature, an invisible signal embedded in model output rather than a user-facing badge, mean most EU users will notice no change in how ChatGPT or Codex behaves.

Why it matters

This is the first time OpenAI has shipped provenance signals directly into its flagship consumer products, and it makes the EU the proving ground for a capability that could later spread worldwide. With Anthropic and Google pursuing comparable text watermarking under the same regulation, the AI Act is effectively standardizing provenance infrastructure across the major model providers, a shift away from voluntary and inconsistent third-party detectors toward embedded, machine-readable signals.

The trade-offs are just as consequential. By keeping detection in the hands of vetted researchers, OpenAI avoids the false-positive problems that have long plagued public AI detectors, but it also limits the immediate practical value for educators, publishers and platforms. And as OpenAI itself concedes, a watermark can only ever indicate that a model was involved; it cannot settle questions of accuracy, ownership or human contribution. How regulators, researchers and users weigh that partial guarantee will determine whether watermarking becomes trusted infrastructure or merely a compliance checkbox.