I came across this thing while scrolling the updates so let’s see where is this exactly going.
What’s happening
OpenAI is rolling out an invisible watermark on eligible ChatGPT and Codex text, but only in the EU for now. API users can opt in on select models too, though it stays off by default.
Why the EU specifically
The EU AI Act requires this kind of transparency for AI-generated content, and the deadline for systems already on the market is December 2, 2026. So this isn’t OpenAI being proactive out of goodwill, it’s a regulatory deadline.
The part that caught my attention
OpenAI tested how well the watermark survives editing, and the results are honestly not that reassuring:
- Swap 10% of the words for synonyms, and detection drops from about 92% to 66%
- Swap 25%, and it falls to 17%
So a light edit weakens it, and a real rewrite probably kills it. Also worth knowing: the watermark detector isn’t public. Only approved researchers can check it, not publishers, not teachers, not platforms.
Why this matters beyond the EU
- A missing watermark doesn’t prove a human wrote something
- A detected one doesn’t tell you how much a human contributed
- OpenAI said this themselves, so any policy or contract treating the watermark as proof of authorship is relying on something it was never built to answer
How it compares to Claude
Anthropic already watermarks supported Claude models worldwide, not just one region, using a method based on Google DeepMind’s SynthID-Text. So right now the two companies are taking different paths, one regional and opt-in, one global and on by default.