Artificial intelligence research company OpenAI has announced that it will begin embedding invisible watermarks into text produced by ChatGPT and Codex within the European Union. The measure is designed to align with the regulatory mandates of the EU AI Act, as outlined in a company blog post. The landmark legislation introduced transparency obligations on August 2, requiring developers of artificial intelligence systems to tag machine-generated text in a manner that external detection tools can reliably verify.
Deployment Timeline and Developer Settings
According to OpenAI, the watermarking feature will roll out over the coming weeks to eligible ChatGPT and Codex subscribers across all subscription tiers, restricted exclusively to the European Union. The system will not function as a global default across standard consumer interfaces at launch. However, developers utilizing the company's API globally can manually enable the functionality for select models, where the feature remains disabled by default.
The Mechanics of textGrain
Rather than embedding visible characters or visual indicators, the watermarking mechanism subtly alters word choices during generation. This establishes a hidden statistical signature that remains undetectable to human readers while remaining recognizable to specialized detection algorithms. Because the watermark is woven directly into the syntax, it persists even when text is copied and pasted across different applications. OpenAI confirmed that the mechanism does not collect or track user identities, and benchmark evaluations revealed no meaningful impact on model performance.
Alongside the feature release, the company issued an accompanying technical report detailing the underlying method, termed textGrain. Co-authored with academic researchers from the University of Pennsylvania and Yale, the paper explains how a cryptographic secret key guides next-word probability distributions. As hundreds of these calibrated adjustments accumulate throughout a passage, a verification tool can detect machine generation using only the target passage and the key.
Vulnerability to Human Editing
Internal evaluations show that subsequent human editing can dilute the watermark. When testing alterations where 10 percent of generated words were substituted with synonyms, the verification rate fell from roughly 92 percent down to 66 percent. Detection also degrades substantially across concise sentences, mathematical formulas, and cross-language translations.
“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses.”
The company explicitly warned that an absence of watermarks does not confirm human authorship, as text may simply be brief, heavily modified, or produced by competing generative models. The watermark only verifies whether an OpenAI model generated or processed parts of a document, without measuring human effort or original thought.
Competitive Context and Transparency Codes
The decision follows a global rollout by Anthropic, which recently committed to watermarking outputs generated by Claude. Anthropic's move triggered resistance among subscribers who argued that human users supply the logic, context, and editorial direction while the model merely functions as an instrument. OpenAI previously built a text watermarking prototype but deferred deployment over concerns that strict watermarks would drive customers toward competing models. Both companies, alongside Google, Meta, and Microsoft, have formally committed to the European Union's code of practice concerning transparency for synthetic content.
