OpenAI is introducing an invisible watermarking system for text generated by ChatGPT and Codex in the European Union to comply with new regulatory requirements, while maintaining an opt-in model for API users globally. The move distinguishes OpenAI’s approach from competitor Anthropic, which mandates watermarking for its Claude models regardless of access method.
What Happened
The EU AI Act reportedly requires AI providers to label generated text in a machine-readable format. To meet this standard, OpenAI has developed "textGrain," a technology that embeds an invisible statistical signal into the model's word choices. The company stated that watermarking will be activated for ChatGPT and Codex users in the EU over the coming weeks. For API customers worldwide, including those using cloud partners like Microsoft Azure, watermarking will remain an opt-in feature.
According to OpenAI, textGrain performed comparably to or better than other methods in internal tests, including Google’s open-source SynthID technology. The company stated that watermarking does not hurt output quality, citing tests on its frontier model Astra across benchmarks like GPQA Diamond and DeepSWE, where it reported no significant performance differences with watermarking on or off. The company plans to release textGrain as open source to allow others to build upon it.
However, detection reliability varies significantly based on text length and subject matter. With the detector set to a 1 percent false-positive rate, it identified watermarks in approximately 95 percent of 400-token passages about psychology, dropping to about 80 percent at 200 tokens. For mathematical content, where the model has less lexical freedom, detection rates were "substantially lower," with a 400-token math passage detected at only about 60 percent. Editing also impacts detectability; OpenAI reported that replacing just 10 percent of words with synonyms in a 400-token passage reduced detection rates from about 92 percent to 66 percent. Replacing 25 percent of the words dropped detection to 17 percent.
Why It Matters
OpenAI’s strategy highlights a diverging regulatory and technical approach among major AI labs. By making API watermarking optional globally, OpenAI offers flexibility to developers and enterprises that may prioritize seamless integration or privacy over mandatory provenance tagging. This contrasts with Anthropic’s global mandatory watermarking for Claude. Anthropic has not published detection rates for its system but claims it holds up well against edits.
Access to the textGrain detector will initially be restricted to selected researchers and specialist organizations who apply through a specific form, with approvals granted case-by-case under the EU’s Code of Practice. The detector will report only whether a watermark was detected, without identifying users or prompts. OpenAI cited the detector’s potential to flag unmarked text or miss watermarks as the reason for this limitation, stating it plans to expand access "when we believe results can be interpreted responsibly." Meanwhile, existing verification tools for images and audio, such as openai.com/verify, remain publicly available.
The Bottom Line
OpenAI is complying with EU regulations via its new textGrain watermarking system for ChatGPT and Codex in Europe, while keeping API watermarking optional for global users. The company asserts the technology maintains model quality on benchmarks like GPQA Diamond, though detection rates vary by content and editing, and the detector provides limited information by design.