OpenAI will begin embedding invisible watermarks into ChatGPT and Codex text outputs for users located in the European Union — and only there, over the coming weeks, the company confirmed in a 5 October tweet. The move fulfils obligations under the EU AI Act, which requires providers of general-purpose AI systems to disclose when content is AI-generated.
Unlike Anthropic, which launched its own watermarking system globally in August after acknowledging it lacked “a durable way to scope it by region”, OpenAI is taking a deliberately regional approach. It says this allows time to assess real-world performance and gather feedback before expanding. If it does. Outside the EU, watermarking remains disabled by default.
However, enterprise customers using OpenAI’s API worldwide can choose to enable it, meaning some non-EU outputs may carry the signal regardless. The technology behind it is called textGrain. It works by subtly biasing word selection during generation to embed a statistical pattern, not metadata, not visible tags, not hidden strings.
That makes it invisible to readers but theoretically detectable by trained classifiers. But textGrain has known limitations: it struggles with short outputs, where too few word choices exist to encode a reliable signal, and editing after generation. Even minor rephrasing, can weaken or erase the watermark entirely. Critics have raised concerns that such patterns could degrade output quality or expose users who rely on AI for writing, especially in professional or academic contexts.
OpenAI counters that “across the benchmarks we use to assess Astra, our latest frontier model, we do not see meaningful performance differences with and without watermarking”. Still, detection tools remain deeply flawed. AI classifiers frequently misidentify human-written text as AI-generated. False positives that risk penalising legitimate authors.
OpenAI acknowledges this, noting that “AI detectors are far from flawless”. To address scrutiny, the company says it’s inviting “select researchers and expert organisations” to test and vet the watermark detector. Access will expand only when OpenAI believes results “can be interpreted responsibly”. No timeline for broader access has been given.
The rollout applies exclusively to ChatGPT and Codex, not DALL·E, Sora, or other OpenAI models. And while the EU AI Act’s transparency rules apply to all general-purpose AI systems, enforcement deadlines vary: high-risk systems face stricter timelines, but foundational models like ChatGPT must meet disclosure requirements by the end of 2026.
OpenAI’s move arrives just ahead of that cutoff.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.