Can't find the reference atm, but I recall this being a thing they were working on, Google even open-sourced their "SynthID"

The idea is that AI generated images can be debunked in cases of slander or falsified documents. The counter-argument is, if you know how to read the watermark then you can also remove it...

It wouldn't surprise me though if they're using steganography in a way that's not publicly known, for law enforcement etc, as a condition of government contracts.

At the end of the day anyone hoping to pass off AI generated images as not AI generated is a scammer and will get pwned.