OpenAI adds opt-in text watermarking; ChatGPT auto-marks in the EU
Developers can now opt in via textGrain, but the detector's expert-gated and paraphrasing defeats it—while Anthropic has marked Claude worldwide since August.
Copy markdown
Opt-in watermarking lands in the OpenAI API
textGrain embeds an invisible statistical signal into a model's word choices. It's off by default and you flip it on per request for select models, worldwide; over the coming weeks ChatGPT and Codex outputs get marked automatically, but inside the EU only. If you ship OpenAI-generated text and face EU AI Act transparency duties, you finally have a switch for it.
The detector is expert-only, and a thesaurus beats it
At a 1% false-positive rate, detection runs ~80% on 200-token passages and ~95% at 400—but replacing 25% of words with synonyms crashes it to 17%, and math-heavy text scores far lower. OpenAI opened applications today and is gating the detector to approved researchers. You can mark your own output, but you can't reliably verify anyone else's yet, and a hit proves neither authorship nor accuracy.
Claude has quietly watermarked worldwide since August
Anthropic switched on SynthID-Text marking for Claude on Aug 14 and applies it globally—it says it can't region-scope, so there's no opt-out. If keeping model text invisible to AI detectors matters to your product, your provider choice has already decided it.
The EU AI Act made provenance table stakes
The Act's transparency rule took effect Aug 2 and ~190 orgs signed the EU code of practice in July; Google, Meta, Microsoft and Mistral are all shipping their own marking. Expect provenance signals baked into more APIs by default—and watch for the quality or latency tradeoffs to surface in your outputs.