This pattern can be read by another AI-based tool but cannot be distinguished by humans.
“Only watermarking AI-generated content won’t solve the problem of proving the authenticity of content.
What is needed is an ecosystem of authenticity for all digital media.
AI-generated images.Google
That’s pretty obvious, but it’s also the essential takeaway here.
Another current problem is that everybody is developing their own AI image-marking standards.
Amazon, Google, Inflection, Meta, OpenAI, and moreare all working on various AI safety solutions.
Let’s add another complication.
C2PA uses cryptography to embed a kind of “nutrition label” into AI-generated works.
Stock photography serviceShutterstock has announced its intentionto use C2PA.
Universal Watermarking
For any kind of watermarking or labeling to work, it needs to be universal.
Every AI image generator would have to embed its detection system into every file it generated.
And it needs to be international, which seems unlikely.
Governments with bad intentions would surely make these tools available to people in other countries, for example.
The good news is, there is precedent for this kind of thing.
Take a dollar bill, scan it, and have a go at print a copy.
This system is implemented by Adobe in its Photoshop software, among others.
For example, C2PA could be used by human creators to embed their own copyright information into their creations.
C2PA cryptographically binds its labels to the pixels of the image, making it much harder to remove.
