OpenAI rolls out invisible watermark for ChatGPT text in the EU
OpenAI has started marking AI-generated text with a hidden signal. The new textGrain watermark now appears in content from ChatGPT and Codex, but only for people in the European Union. This step answers new legal demands under the EU AI Act. The watermark is invisible to readers. Only machines can spot it.
TextGrain does not use visible marks or metadata. Instead, it tweaks word patterns in the output. Readers cannot see these changes. OpenAI's own tools can detect them. The company says the watermark survives copy-paste. Basic edits will not erase it. OpenAI explains in its technical report that textGrain works by adding a statistical signal. It changes word choices in small ways, not by adding symbols or hidden tags.
The textGrain watermarking method was developed in collaboration with researchers from the University of Pennsylvania and Yale, and relies on a secret key to rank possible next words and accumulate a statistical pattern over longer texts.
How textGrain works and what users should know
OpenAI says watermarking does not slow down ChatGPT or Codex. Benchmarks like the 'Artificial Analysis Intelligence Index' and 'AutomationBench' show no drop in performance. The watermark does not reveal who wrote the prompt or the user's identity. It only marks that the text came from OpenAI's models.
Detection is not perfect. The watermark shows up best in long, open-ended text. OpenAI reports an 80 percent detection rate in 200-token psychology samples. That jumps to 95 percent for 400-token samples. In math or other rigid fields, detection drops. Editing the text weakens the signal. Swapping 10 percent of words for synonyms cuts detection to 66 percent. At 25 percent, it falls to 17 percent. Translation makes detection even harder. These numbers come from the OpenAI technical documentation.
Regulatory push and a limited launch
The EU AI Act now forces generative AI providers to make their outputs machine-readable as AI-generated. OpenAI's launch of textGrain in the EU is a direct answer to this rule. There are no plans for a worldwide rollout yet. API customers outside the EU can use the watermark on select models, but only if they choose. OpenAI plans to share its detection tool with researchers and certain organizations. The company warns that detection will not always work in real-world use.
Transparency obligations for generative AI under Article 50 of the EU AI Act, including machine-detectable labeling of generated content, are set to take effect from August 2, 2026.
OpenAI is not the only one using this approach. Google DeepMind's SynthID and Anthropic's Claude watermarking system use similar tricks to mark AI text. The industry is moving toward technical ways to track content origins. But these systems have limits. Editing, translation, and short texts make detection much harder.
TextGrain does not check if the content is true or if a human helped write it. It only adds a technical marker for transparency. OpenAI is rolling out the watermark in the EU first. The company wants to see how it works in practice before going wider. For now, textGrain is a tool for legal compliance. It is not a cure-all. Its limits matter as much as its strengths. The next big test for the industry will be finding a balance between transparency, privacy, and the technical limits of watermarking as AI keeps changing.
