OpenAI Rolls Out Text Watermarking for ChatGPT and Codex in the EU Only

OpenAI is introducing an invisible, machine-readable marker in the text its models produce, but the rollout is limited to the European Union for now. The company says the watermarking system, called textGrain, will reach eligible ChatGPT and Codex users on all plans in the EU over the coming weeks, as reported by The Verge.

The regional approach is deliberate. OpenAI says it is not making text watermarking a global default at launch, describing the EU-only rollout as a way to learn from real-world use and feedback before broadening it.

What the watermark does — and does not do

TextGrain embeds a signal in model output that is intended to be detectable by machine, not visible to readers. OpenAI says its method matched or exceeded other approaches, including Google DeepMind's SynthID for text, which also underpins watermarking that Anthropic announced in August. Like OpenAI, Anthropic acted to meet the EU AI Act's requirements. The Verge notes that not everyone welcomed Anthropic's addition.

OpenAI published AI benchmark scores indicating similar performance between watermarked and unwatermarked text. But the company also attached limits to what the technology can establish. It says textGrain does not guarantee reliable detection. Watermarks, according to OpenAI, do not verify accuracy, determine ownership of text, measure how much a human contributed, or prove human authorship. Those caveats matter because a watermark is often described as a provenance signal; OpenAI is framing it more narrowly.

Access for API customers and researchers

Outside the EU, the rollout takes an opt-in form. Starting now, API customers worldwide can choose to receive watermarked text outputs for select models. OpenAI says this lets customers decide how watermarking fits their own transparency obligations and the experiences they provide to users. The company adds that it is working with cloud partners to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks.

A separate track covers detection. Approved researchers and expert organizations can apply for access to a detector starting today. OpenAI says access will initially be granted case by case, in line with a Code of Practice, to support evaluation and improvement of text provenance. The detector reports whether it finds an OpenAI watermark, without identifying the user or exposing prompts or conversations. Because of the risk of missed watermarks and false positives, OpenAI is not releasing the detector publicly at launch.

What remains uncertain

The source material leaves several practical questions open. It does not specify which models support opt-in watermarking for API customers, how long the EU-only phase will last, or what conditions would trigger a wider default. It also does not detail how detection accuracy will be measured or reported, beyond OpenAI's acknowledgment that missed watermarks and false positives are real concerns.

For EU users, the change is largely automatic: text output from ChatGPT and Codex will carry the marker whether or not they seek it. For API customers elsewhere, the decision is theirs. For researchers, access depends on approval. OpenAI's own framing suggests the technology should be understood as a limited provenance signal rather than proof of authorship or accuracy.

Attribution: This report is based on The Verge's coverage of OpenAI's text watermarking announcement. Confirmed details come from OpenAI's stated rollout plans and caveats as reported; interpretation about user impact is the editorial desk's, not OpenAI's.