Watermarks Won't Settle It. Provenance Might Help.

Every few months, someone proposes the same fix for AI-generated fakes: make every AI image, video and audio clip carry an invisible watermark, and the problem goes away.

It's an appealing idea. It's also not enough on its own, and understanding why helps make sense of a quieter approach that may matter more.

Why watermarks struggle

A watermark is a hidden signal embedded in content when it's generated. The trouble is that content travels. It gets cropped, compressed, screenshotted, re-recorded and remixed. Each step can weaken or erase the signal. And watermarks only exist if the generating tool adds them; open-source tools and bad actors can simply skip that step.

There's also a deeper problem. A watermark can, at best, tell you something was generated. It says nothing about the far larger universe of content that wasn't — and it's the authentic photos and recordings we most need to trust.

Provenance flips the question

Instead of trying to label every fake, provenance focuses on documenting the real. The idea: cameras, editing apps and publishers attach signed records to content describing where it came from and what was done to it. Taken on this device. Cropped and color-corrected in this app. Published by this outlet.

An open technical standard from the Coalition for Content Provenance and Authenticity (C2PA) is one widely discussed approach; its records are often presented to people as "Content Credentials." When provenance data is present and intact, you can check a piece of content's history instead of guessing.

The honest limits

Provenance isn't magic either. Metadata can be stripped, and most content online today carries none. It only helps if cameras, platforms and newsrooms adopt it widely, and if people know to look.

The likely future isn't one technique winning. It's layers: watermarks where they work, provenance for authentic content, platform labels, and the oldest tool of all — asking where something came from before you share it.