OpenAI rolls out invisible textGrain watermarking in ChatGPT and Codex, EU first
The Verge AI · rss · 2026-10-06
OpenAI is rolling out textGrain — an invisible, machine-readable watermark in text output — to ChatGPT and Codex users, starting with the European Union.
- Why: to meet EU AI Act content-labeling requirements. Anthropic's August-announced text watermarking was similarly compliance-driven, and builds on the same approach as Google DeepMind's SynthID.
- Performance: OpenAI says textGrain "matched or exceeded" alternatives like SynthID, and published benchmark scores showing watermarked and unwatermarked text perform similarly.
- Caveat: OpenAI itself notes textGrain "does not guarantee" reliable detection, and not all users welcomed the addition.
More from Safety
- OpenAI Tells NYC Council Staff Can Now Flag Misalignment for Public Review — ryanmerket · 2026-10-06
- Bipartisan Senate bill would hold AI agent operators and developers criminally liable for hacking — StephenLCasper · 2026-10-06
- Fearing Data Misuse, Companies Turn to Open Models and 'Sovereign AI' Over OpenAI and Anthropic — fortune · 2026-10-06
- Researchers puzzled by absence of an org doing open alignment science — sebkrier · 2026-10-06
- ChatGPT user dropped into malicious Custom GPT mid-task in ClickFix malware attack — BarracudaLittle4376 · 2026-10-06
- Malaysia has no AI law yet but is building governance on existing rules — Comfortable_Gene5180 · 2026-10-06