How-to
OpenAI text watermarking: API opt-in, EU rollout ahead
OpenAI's textGrain system adds an invisible signal to word choices, but detection is limited and easily weakened by editing.
9to5Mac reports that OpenAI has detailed how it will comply with the EU AI Act's requirement to mark AI-generated text. The system, called textGrain, adds an invisible statistical signal to the model's word choices.
The approach is similar to one from Anthropic, which began tweaking word selection based on a hidden key. Both methods are designed to leave a mark that doesn't affect meaning or readability.
Rollout scope
OpenAI's deployment is more conservative than Anthropic's. API customers anywhere can opt in to watermarking on select models starting today. For ChatGPT and Codex, an invisible watermark will be added only for eligible users in the European Union, with a rollout scheduled in the coming weeks. The watermark remains off by default for API users and is initially limited to EU users in the chat and coding tools.
The detection tool that checks for the watermark is not open to the public. OpenAI is accepting applications from approved researchers and expert organizations as it works to evaluate and improve the technology.
Limitations
The company is direct about the system's shortcomings. False positives and false negatives occur, especially in short texts or in domains like mathematics where word choice is constrained. In testing on 400-token passages, replacing 10 percent of words with synonyms reduced detection from about 92 percent to 66 percent. Replacing 25 percent dropped it to 17 percent.
OpenAI stresses that a detected watermark does not confirm accuracy, establish ownership, or identify the user. The absence of a watermark is not proof of human authorship. Watermarked outputs actually scored slightly higher on some benchmarks, but OpenAI says there is no meaningful difference in overall performance.
The company says it plans to revisit its approach as technology and standards evolve.
Bottom line: If you are relying on watermarks to catch AI-generated text, these numbers are a reality check. Editing or translation can easily erase the signal.
Related stories
Phones
OpenAI adds text watermarking to ChatGPT, Codex in EU first
The invisible watermark is rolling out to European Union users first, with opt-in available globally for API customers.
How-to
Meta Muse vs. ChatGPT Dots: what each one costs
Two always-on AI agents promise to work while you are away, and Engadget's comparison shows where they differ on price, permissions and availability.
How-to
AI is a cognitive hot dog, The Verge argues
The Verge says making AI a regular mental meal weakens the systems that keep human thinking active.