OpenAI Will Add Invisible Watermarks to ChatGPT Text in the EU

OpenAI will begin adding invisible watermarks to text generated by ChatGPT and Codex in the European Union, giving authorised systems a new way to identify content produced by its AI models.

The company said the feature will roll out over the coming weeks to eligible ChatGPT and Codex users across all plans in the EU. Developers using OpenAI’s API can also enable the watermark for supported models worldwide, although it will remain off by default outside the EU.

The watermark changes patterns in the text itself

OpenAI’s system, called textGrain, does not add a visible label, symbol or metadata tag. Instead, it subtly influences the words a model selects while generating a response, creating a statistical pattern that specialised detectors can identify.

Because the signal is embedded in the language itself, it can remain present when someone copies and pastes the text. OpenAI says the watermark does not identify the individual user and that its testing found no meaningful reduction in model performance when the system was enabled.

A technical paper on textGrain, produced with researchers from the University of Pennsylvania and Yale, describes how a secret key can influence the model’s next-word choices. Repeated across a sufficiently long passage, those small changes create a pattern that can later be tested using the corresponding detector.

Editing can weaken the watermark

The technology has important limitations. OpenAI found that replacing about 10% of words with synonyms reduced detection performance in one test from roughly 92% to 66%. Short passages, mathematical answers, and translated text can also be harder to identify reliably.

Because of those limitations, OpenAI is initially restricting detector access to approved researchers and expert organisations. The company also cautions that failing to detect its watermark does not prove a human wrote the text, since the signal may have been weakened through editing or the content may have been generated by another AI system.

OpenAI previously developed text-watermarking technology but did not release it broadly. The Wall Street Journal reported in 2024 that the company had considered concerns that users could move to competing AI services that did not watermark their output.

The EU AI Act is driving the rollout

The European rollout comes as new transparency obligations under the EU AI Act take effect. European Commission guidance on AI transparency requires providers of generative AI systems to make certain AI-generated content detectable in machine-readable form.

OpenAI is among the technology companies participating in the EU’s voluntary framework for implementing those requirements. The European Commission has said companies including OpenAI, Google, Meta and Microsoft are supporting the code of practice covering transparency for AI-generated content.

The watermark does not show how much of a passage was written, edited or directed by a human. Instead, it is intended to provide another provenance signal indicating that an OpenAI system generated or processed at least part of the text.