OpenAI launches textGrain to watermark EU AI text

If you copy and paste an answer from ChatGPT, it looks like normal text to the naked eye. But underneath the hood, a statistical breadcrumb trail is now following your words around.

OpenAI just launched textGrain, an invisible text watermarking system built to comply with the EU AI Act's machine-readable provenance rules that took effect on August 2. But the company isn't flipping the switch everywhere just yet.

Over the coming weeks, text watermarking will roll out by default only to eligible ChatGPT and Codex users in the EU across all subscription plans. For the rest of the world, API developers can opt in starting today on select models, but it stays off by default.

Hiding signals in plain sight

Unlike an image watermark or a metadata stamp, textGrain doesn't tack on visible logos or disclaimers. It works by subtly nudging the model's word choices during text generation.

Co-developed with researchers from Yale and the University of Pennsylvania, the system uses a secret key to order next-word probability predictions. Stacking hundreds of these microscopic nudges creates a statistical pattern invisible to humans, which OpenAI's detector can pick up using just the raw text and the key. In benchmarks on its latest frontier model, Astra, OpenAI claims performance didn't take a hit.

The real catch is durability. Text is famously fragile, and textGrain breaks down fast under real-world edits. In OpenAI's own testing on 400-token passages, swapping just 10% of the words with synonyms dropped the detection rate from 92% to 66%. Swapping 25% cratered it to 17%.

Shorter passages, translations, and math problems with strict word choices are even harder to verify. Because false positives and misses remain a headache, OpenAI is keeping the detector locked down for now, granting access only to vetted researchers and expert organizations.

Why it matters

OpenAI is basically walking a regulatory tightrope here. Back in 2024, reports revealed the company had built text watermarking tech but shelved it over fears that users would flee to unwatermarked rivals. When Anthropic rolled out Claude watermarking globally two months ago, it triggered immediate user backlash from people insisting the AI was just an assistant.

By restricting the default rollout strictly to the EU and leaving API usage strictly opt-in, OpenAI satisfies Brussels regulators while avoiding a mutiny in the US and beyond. We think it's a calculated, defensive maneuver rather than a confident product leap.

For everyday users and students, the immediate danger of getting flagged is surprisingly low. The watermark doesn't tie back to your account or prompt history, and light paraphrasing easily scrubs the signal. But for the broader ecosystem, this sparks an inevitable arms race between watermarkers and text-spinning tools. As OpenAI itself admitted, absence of a watermark doesn't prove human authorship, and a detected watermark doesn't measure how much human effort went into editing. Until those lines are clearer, text watermarking remains more of a compliance shield than a foolproof truth machine.