An engineer tested whether AI image generators actually block copyrighted characters. Disney's Frozen characters were refused at first, but four tweaked prompts produced images identical to the originals. The vulnerability shows how language manipulation can bypass copyright filters in AI systems.
I ran this test with an engineer that has 2 daughters that love Frozen, and copyright is a very big issue with these generative AIs. So we asked it to create an image of Anna and Elsa from Frozen, and it says it can't do it, it's copyrighted. Okay. In 4 prompts, we had an image of Anna, Elsa, and Olaf. Wow. Looked identical to Disney. So you can trick these systems because they're language-based. And somebody that is rather articulate and knows how to use language can use it to divulge information.
