TL;DR
OpenAI is rolling out textGrain, an invisible statistical watermark for EU ChatGPT and Codex users, meeting AI Act transparency rules while acknowledging the technology still has significant detection gaps.
What happened
- OpenAI launched textGrain, an invisible watermark embedded in text generated by ChatGPT and Codex, targeting all EU plans over the coming weeks.
- The system works by subtly shifting statistical word-selection patterns during generation, invisible to readers but detectable by OpenAI's proprietary tool.
- API customers worldwide get a separate opt-in option to enable watermarked output; the default remains off outside the EU.
- OpenAI is restricting detector access to researchers and expert organizations initially, citing false-positive risks and reliability concerns.
- The move follows Anthropic's worldwide Claude watermarking rollout two months earlier, and aligns with EU AI Act compliance commitments made alongside Google, Meta, and Microsoft.
Why it matters
- Detection accuracy is fragile: the detector catches watermarks in roughly 80% of 200-token passages and 95% of 400-token passages, but those numbers collapse under editing.
- Replacing just 10% of words drops detection from 92% to 66%; replacing 25% pushes it to roughly 17%, meaning any motivated editor can largely defeat the system.
- A positive detection does not identify the user, account, or conversation, and says nothing about accuracy, ownership, or responsibility, limiting its legal utility.
- The EU AI Act is now forcing concrete product decisions at the world's largest AI labs, with OpenAI shipping a feature it had previously delayed out of fear users would defect to less-restricted rivals.
- Open-sourcing textGrain could set an industry baseline, but it also hands adversaries a roadmap for circumvention.
What to watch next
- Whether detection reliability improves enough for regulators to treat textGrain as meaningful compliance evidence, or whether the EU demands stronger provenance standards.
- How API adoption rates develop globally: if enterprises opt in at scale, watermarking becomes a de facto standard; if they ignore the toggle, the policy impact stays narrow.
- Whether Anthropic's Claude watermarking and OpenAI's textGrain converge on a shared technical standard under the EU code of practice, or fragment into incompatible proprietary systems.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.