OpenAI said it will begin invisibly watermarking text generated by ChatGPT and its Codex coding assistant for users in the European Union, rolling the feature out over the coming weeks to people on all subscription tiers. The company is also making the underlying watermarking capability available immediately to developers building on its API worldwide, though it will remain switched off by default outside the EU, OpenAI told TechCrunch.
The move is a direct response to the European Union's AI Act, whose transparency provisions took effect on August 2 and require companies offering general-purpose AI systems to mark synthetic text, images, audio and video outputs in a way that other systems can detect as machine-generated. "These rules require AI companies to mark AI-generated content in a way other systems can identify," OpenAI said in explaining the change.
How the Watermark Works
Rather than stamping text with a visible symbol, OpenAI's approach works by subtly steering which words and phrasings the model chooses as it generates a response, creating a statistical pattern that is invisible to a human reader but detectable by a matching tool. OpenAI says the watermark travels with the text even after it is copied and pasted elsewhere, and that enabling it produces no meaningful change in the quality of ChatGPT's answers.
The company's own testing found real limits to the approach. Replacing just 10 percent of a passage's words with synonyms dropped detection accuracy from 92 percent to 66 percent, and short responses, math answers and machine-translated text proved considerably harder to flag reliably. OpenAI was blunt about what that means in practice: "a missing watermark does not prove human authorship."
OpenAI has previously said it built text-watermarking technology years ago but held it back over concerns that it could be circumvented by motivated users and might disproportionately flag writing from non-native English speakers, whose phrasing patterns can resemble AI-generated text. The company described its rollout now as a "deliberate approach" that still carries those risks, even as it complies with the EU mandate.
Following Anthropic's Lead
OpenAI's move follows Anthropic, which announced roughly two months earlier that it would embed similar invisible watermarks into text produced by its Claude chatbot — and, unlike OpenAI's EU-only rollout, apply it globally by default. That decision drew pushback from some Claude users who objected to having their work-related AI use made detectable without an easy opt-out. Google, Meta, Microsoft and OpenAI have all signed onto the EU's voluntary code of practice on transparency for AI-generated content, alongside roughly 190 other organizations.
Whether OpenAI extends default watermarking beyond the EU remains an open question the company has not committed to answering. For now, the rollout gives European regulators their clearest test yet of how rigorously AI developers will comply with a rule that, by design, depends on tools most ordinary users will never see.
