OpenAI Adds textGrain Watermarks to ChatGPT in the EU
To comply with the EU AI Act, OpenAI activates textGrain watermarks for ChatGPT only in Europe, while keeping the detection tool restricted to researchers.
The statistical signal complies with Article 50 of the EU AI Act, but struggles with paraphrasing and leaves detection locked down.
On October 5, 2026, OpenAI announced textGrain, its proprietary system for embedding statistical watermarks into artificial intelligence text. The rollout directly addresses regulatory pressure from Article 50 of the European Union’s AI Act, which requires developers of generative tools to mark synthetic output in machine-readable formats. Instead of activating the feature worldwide, the company is splitting its user base geographically. The invisible watermark will roll out over the coming weeks exclusively for ChatGPT and Codex users located within the EU.
How textGrain applies its statistical watermark
Past industry efforts to track synthetic writing typically relied on fragile tricks, such as invisible Unicode characters, hidden whitespace, or unnatural sentence punctuation. Those marks were easy to strip using simple text editors or browser extensions. The textGrain system works at a deeper computational level, adjusting the probabilistic math inside the language model itself.
During generation, a model assigns probabilities to choose each successive token, or word fragment. According to OpenAI’s technical documentation, textGrain uses a private cryptographic key to introduce subtle statistical biases into that selection. The resulting phrasing sounds natural and readable on screen. Behind the scenes, however, the text carries an invisible distribution signature that detection software holding the matching key can identify.
Regional split for ChatGPT and the API
The transparency mandates under the EU AI Act took effect on August 2, 2026, with an adaptation window for existing systems closing on December 2. To meet European legal requirements without altering outputs for users elsewhere, OpenAI restricted default deployment to the European bloc.
Both free and paid ChatGPT accounts in EU member states will have textGrain enabled by default as the phased rollout progresses. Users in the United States, Latin America, and other regions will continue generating unmarked text. For enterprise teams and developers accessing models through the OpenAI API, the rules differ globally: textGrain remains off by default everywhere, requiring teams to manually toggle the feature in their project settings.
Practical limits and a closed detector
Despite the technical complexity, OpenAI’s benchmark data reveals practical weaknesses in the watermark. Internal test results reported by TechCrunch show that textGrain withstands basic copy-pasting and light editing. However, replacing just 10 percent of words with synonyms or rewriting sentences caused the detection rate to plunge 26 percentage points, sliding from roughly 92 percent to 66 percent.
OpenAI stated that textGrain matched or outperformed Google DeepMind’s SynthID Text in laboratory tests. Because the creator conducted these evaluations internally, the results reflect familiar vendor benchmarks, and OpenAI acknowledges that everyday use will inevitably yield false positives and false negatives.
Access to the verification tool presents another operational hurdle. While image provenance standards like C2PA allow anyone to inspect files through public portals, OpenAI is keeping the textGrain detector private. Neither educators nor general users can run text checks directly. Access requires an application and case-by-case vetting, reserved for academic researchers and regulatory authorities. European legislation forced OpenAI to stamp its text, but the company holding the key decides who gets to inspect the label.