SynthID's truce

At Google I/O in May, OpenAI agreed to embed its chief rival's watermark. Every image from ChatGPT, Codex and the OpenAI API now carries SynthID, the invisible signal Google DeepMind has been pushing since 2023, joining NVIDIA's video tool, ElevenLabs' audio and Korea's Kakao. Five companies that compete on everything else now agree on one thing: how to secretly mark what a machine made.

"The model may produce false positives... and false negatives... a probabilistic estimate [that] doesn't guarantee definitive identification." — Google Cloud's AI Content Detection API docs

Google has already tagged more than 100 billion images and videos this way, plus tens of thousands of hours of synthetic speech, and is wiring a checker into Chrome and Search so anyone can right-click an image to ask where it came from.

The catch is buried in Google's own engineering docs. The new business detector that promises to flag content from any model doesn't actually read a watermark when the image came from a rival — there is nothing to read. It falls back to an older trick: an algorithm sniffing for the statistical fingerprints of generation, the same brittle approach people have been fooling for years. A cottage industry of removal tools already strips these marks in a single pass, and one June study found Google's own watermark roughly three times harder to scrub than OpenAI's.

So the convergence is real and the reach is genuinely wide — but it arrives labelled, by its maker, as a probability, not a proof. The useful question is no longer whether the big labs will agree on a watermark. It's whether a mark anyone can remove, and a detector that admits it guesses, can carry the weight the internet is about to put on it.