On Oct 5 OpenAI described its response to the EU AI Act's machine-readable text requirement: API customers worldwide can now opt in to text watermarking for select models (off by default), and in the coming weeks eligible ChatGPT and Codex text output in the EU will carry an invisible watermark across all plans. The technique, textGrain, embeds a statistical signal in word choices; OpenAI says it matched or beat SynthID for text and plans to open-source it. The detector is limited to approved researchers. At 1% false-positive rate, detection is ~80% on 200-token and ~95% on 400-token passages, falling to 17% after 25% synonym replacement; Astra benchmarks show no meaningful quality change.
Key Takeaways
- ✓Scope: invisible watermarks on eligible ChatGPT and Codex text in the EU in coming weeks (all plans); API opt-in worldwide, off by default
- ✓Detection at 1% FPR: ~80% on 200-token passages, ~95% on 400-token; much lower for math-style text
- ✓Robustness: 10% synonym replacement drops detection ~92%→66%; 25% drops it to 17%
- ✓Quality: Astra (max) DeepSWE v1.1 72.80% vs 71.68%, Terminal-Bench 4.0 53.90% vs 56.06%, GPQA Diamond 94.44% vs 93.94% (unwatermarked vs watermarked)
- ✓Openness: textGrain technical report published, open-source release planned; detector limited to approved researchers
Developer 3s Key Decision Metrics
Turn your technical choice into a development budget
Compare 29+ dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
OpenAI's EU text provenance post answers the EU AI Act requirement that generated text be machine-identifiable. Its textGrain technique adds an invisible statistical signal to word choices; a detector checks for it. OpenAI says textGrain matched or beat other approaches it tested, including SynthID for text, published a technical report and plans to open-source it.
Rollout: API customers worldwide can opt in to watermarked text for select models starting now (off by default), with cloud partners to follow; over the coming weeks eligible ChatGPT and Codex text output in the EU will be watermarked on all plans; detector access is limited to approved researchers and expert organizations. Reported numbers: at a 1% false-positive target, ~80% detection on 200-token passages and ~95% on 400-token ones, much lower for math-like low-entropy text; 10% synonym replacement drops detection from ~92% to 66%, 25% to 17%. On Astra (max) benchmarks there is no meaningful quality difference (e.g. DeepSWE v1.1 72.80% vs 71.68%, Terminal-Bench 4.0 53.90% vs 56.06%). For developers: EU-facing products can use the API opt-in for transparency obligations, Codex output in the EU will carry the signal, and a non-detection never proves human authorship.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.