SynthID will cause more trouble than it's solving

2026-05-21

Watching the Google I/O, they announced that more partners than ever are using synthID as way to watermark content generated by an AI.

This sounds like a good idea on the surface as it suddenly makes it trivial to verify if some content has been generated or not, provided that it was generated by one of these partners.

However, this creates an incentive for all these powerful companies to label any open source model as dangerous, since they won't be able to integrate the watermarking system.

So I can see how in a few years we'll get law proposals against open-weight models, while any bad actors will still have access to models capable of generating propaganda material that doesn't sign its content, and I can't think of a way to prevent this.