OpenAI Adds AI Watermarks for EU Users, Optional Elsewhere

Why this matters

The default-off setting outside the EU tells you where the real pressure is coming from: regulation, not conscience. OpenAI built textGrain because Brussels made it, not because the company suddenly decided invisible provenance tracking was a feature users wanted. That's worth remembering every time a vendor frames compliance work as innovation.

For developers building on the API, the calculus is simple: watermarking is off by default, which means most generated text in the wild will carry no statistical fingerprint unless you deliberately turn it on. Anthropic's SynthID-style approach for Claude works similarly but OpenAI hasn't published detection rates the way Anthropic has made some claims about edit-resistance. That gap matters if you're building anything where provenance verification is the point, not an afterthought.

The Astra performance claims deserve the same skepticism we'd apply to any vendor-run benchmark on its own model. No independent testing, no third-party detection numbers. Until outside researchers can actually probe textGrain's robustness against paraphrasing or translation, treat "no quality loss" as a marketing line, not a settled fact.