OpenAI Adds Invisible Watermarks to ChatGPT Text in the EU

OpenAI will add an invisible, machine-readable watermark to text generated by ChatGPT and Codex in the European Union, aligning the company with the EU AI Act's transparency rules. The method, called textGrain, is live for API developers worldwide who opt in and reaches EU ChatGPT and Codex users on all plans over the coming weeks, according to TechCrunch.

How textGrain works

The watermark is not a visible symbol. It works by subtly shaping the model's word choices, leaving a statistical pattern readers cannot see but a detector can pick up. Because the signal lives in the words, it survives copy and paste. OpenAI says it does not identify the user and that it saw no meaningful change in model performance with the feature switched on.

OpenAI published a technical report with researchers from the University of Pennsylvania and Yale, describing how a secret key sorts next-word predictions into a signature a detector can spot from the text alone.

Compliance, and the limits of provenance

The EU AI Act's transparency obligations took effect on August 2 and require companies to mark AI-generated content in a way other systems can identify. Anthropic began watermarking Claude text worldwide two months earlier, drawing backlash from users who argued they supplied the instructions and context. OpenAI had built a text watermark before but held off, partly over fears users would switch to rivals that did not watermark, as earlier reporting documented.

The limits are real: replacing 10% of words with synonyms dropped detection from roughly 92% to 66% in OpenAI's own tests. Short passages, math answers and translated text are harder to detect. OpenAI is limiting initial detector access to approved researchers and expert organizations, and cautions that a missing watermark does not prove human authorship.

Why it matters for builders

If you ship AI features into the EU, your pipeline's output may now carry a provenance signal. The API switch is off by default, so opt-in is a deliberate choice.

Detector access is the sharper constraint: because OpenAI restricts it to vetted organizations, teams cannot independently verify what their own outputs carry. Treat the watermark as a provenance hint, not proof. It indicates that an OpenAI system generated or processed part of a passage, not how much human judgment went into it.

For agent and automation builders, the practical lesson is to tag provenance at the pipeline level rather than relying on the model's watermark alone. Compliance, logging and audit trails stay your responsibility.