📰 Key Takeaway

OpenAI announced it will add invisible watermarks to text generated by ChatGPT and Codex users across all plans in the EU within the coming weeks, to comply with the EU AI Act’s transparency requirements that took effect on August 2. Global API developers, meanwhile, can manually turn this on for select models starting today — it’s off by default, not a global mandatory default. The watermark isn’t a visible symbol — it works by fine-tuning the model’s word-selection probabilities to leave a pattern in the text that’s invisible to the human eye but detectable by specialized tools. It survives copy-pasting, doesn’t identify who the user is, and causes no noticeable performance hit when enabled. OpenAI also published a technical report, textGrain, co-authored with researchers from the University of Pennsylvania and Yale, explaining how the method works: ranking next-token predictions with a cryptographic key, then stacking hundreds of these adjustments through fine-tuning to produce a detectable pattern. Testing showed the watermark can be broken by editing — swapping out 10% of the text for synonyms dropped detection rates from about 92% down to 66%, and short text, math answers, and translated text are also harder to detect. Because of this, OpenAI is currently only giving detector access to approved researchers and professional institutions. OpenAI also noted that the absence of a watermark doesn’t mean content was written by a human. This move comes close on the heels of Anthropic’s announcement two months ago to watermark Claude’s text globally — which sparked some pushback from users at the time.

💬 JudyAI Lab’s Take

OpenAI is about to add invisible watermarks to text generated by ChatGPT and Codex users in the EU, to comply with the transparency requirements under the EU AI Act that took effect on August 2. Global API developers can turn this on manually, but it’s off by default.

This kind of watermarking reflects a design philosophy around “traceability” that’s taking shape across the AI industry: instead of a visible marker, it leaves a statistical pattern — detectable only by specialized tools — baked into word-choice probabilities through fine-tuning, one that survives copy-paste. But the original summary also exposes the real-world limits of this approach — swap out just 10% of the text for synonyms and detection drops from 92% to 66%, with short text, math answers, and translated content even harder to catch. That’s a reminder that watermarking tools right now function more as “after-the-fact assistance” than “absolute proof.” For AI builders, designing any kind of content provenance mechanism means being honest about this fragility, not treating it as a silver bullet. Anthropic rolled out global watermarking for Claude two months ago, and now OpenAI is following suit — suggesting this could become standard industry baseline rather than an optional feature.

If your product also handles AI-generated content, now’s the time to think about whether users might accidentally break the watermark just by editing text — and how your system should handle that.

📅 Original Source Info

🔗 Further Reading