I've been waiting for something like this to appear. Not just a "plagiarism detector" but something that actually identifies the data in the training pool that most closely represent a particular AI model output. You could do the same for text and images too, and I'm surprised this is the first one that I've heard of.
I'm not a fan of the MAFIAA but if this type of reverse-search tech can hold AI companies to account then it's a step towards reining them in.