AI watermarking is fragile and already being bypassed
A thread analyzing Meta's AI image watermarking scheme, called 'Un-Stable Signature', found the technical claims poorly supported. More directly, commenters linked to a tool at twotensors.ai that already bypasses it. The EU AI Act mandates watermarking that is either technically irremovable or legally protected, but the technical side of that requirement appears to be failing before it even becomes law.
The discussion also raises the deeper problem: watermarking approaches that are detectable enough to be useful are also detectable enough to be removed. The ones that are subtle enough to survive manipulation may not be reliable enough to prove provenance in a legal context.
This is not an academic problem. Publishers, news organizations, and platforms are currently making product and policy decisions based on the assumption that AI watermarking will be a reliable signal. If that assumption is wrong, those decisions need revisiting.
So what?
Founders building content authenticity or provenance products should not treat watermarking as a solved problem. The technical reality is that watermarks are more fragile than the policy conversation assumes, and any product promise built on watermark reliability is on shaky ground.