To comply with the EU AI Act, OpenAI is rolling out an invisible text watermarking technology called textGrain for ChatGPT and Codex users in the European Union. While Anthropic enforces global watermarking, OpenAI is making its watermark optional for API developers worldwide. The technology subtly alters word choices without affecting model performance, but OpenAI warns that simple text editing or translation can easily evade detection, leading the company to restrict detector access to approved researchers.
Compliance with European Union regulations
- ▪OpenAI is introducing text watermarking in the European Union to comply with the EU AI Act's transparency rules, which took effect on August 2, 2026, and require AI providers to label generated text
- ▪AI companies including OpenAI, Anthropic, Google, Meta, and Microsoft have committed to following the European Union's code of practice on AI-generated content
Limitations of textGrain detection
- ▪OpenAI's textGrain watermark detection rates are lower for short passages, translated text, and math content, dropping to 60 percent for math passages compared to 94 percent for psychology passages at 400 tokens
- ▪OpenAI's tests showed that replacing 10 percent of words in a 400-token passage with synonyms dropped detection of its textGrain watermark from 92 percent to 66 percent, while replacing 25 percent dropped it below 20 percent
Policies on API text watermarking
- ▪Unlike OpenAI's optional API watermarking, Anthropic enforces mandatory global watermarking for its Claude models regardless of how users access them
- ▪Starting October 5, 2026, OpenAI is making its textGrain text watermarking opt-in and off by default for API developers worldwide, including through cloud partners like Microsoft Azure
Access and privacy of the detector
- ▪OpenAI's textGrain detector tool will only report whether it detects an OpenAI watermark and will not identify users or reveal their prompts or conversations
- ▪OpenAI is initially restricting access to the textGrain detector to approved researchers and expert organizations on a case-by-case basis under the European Union's Code of Practice
Regional rollout strategy
- ▪OpenAI is rolling out an invisible text watermarking technology called textGrain to eligible ChatGPT and Codex users across all plans in the European Union over the coming weeks from October 5, 2026
- ▪OpenAI stated that it is not making text watermarking a global default at launch to learn from real-world use and feedback
Development of textGrain
- ▪The textGrain technology, co-developed by OpenAI with researchers from the University of Pennsylvania and Yale, embeds an invisible statistical signal by subtly shaping OpenAI's model's word choices
- ▪OpenAI plans to release its textGrain watermarking technology as open source to allow other developers to build on it
Impact on writing quality
- ▪OpenAI's internal tests of its Astra model across eight benchmarks, including GPQA Diamond, BrowseComp, and DeepSWE, showed no significant performance differences with textGrain watermarking turned on or off
- ▪Critics have questioned whether text watermarking affects writing quality, a concern that has also been raised regarding Anthropic's Claude watermark
Limitations of watermark detection
- ▪OpenAI cautioned that the absence of a detected textGrain watermark does not prove human authorship of a text passage
- ▪OpenAI stated that a detected watermark from its textGrain technology does not verify text accuracy, establish ownership, identify a user, or measure how much human creativity or editing went into the text
User resistance to watermarking
- ▪OpenAI previously delayed releasing its textGrain text watermarking technology due to concerns that users would switch to rival platforms that do not watermark text
- ▪Anthropic's decision to watermark Claude drew backlash from users who argued they supplied the instructions and decisions while Claude was just a tool
Debatable claims
- ▪Watermarking AI-generated text unfairly diminishes the role of the human prompter
- ▪Text watermarking is too easily evaded to be an effective AI detection tool
- ▪Subtly altering word choices for watermarking degrades the quality of AI writing
- ▪AI companies should enforce mandatory global text watermarking
Story comments
Loading comments…