If I had to guess. It takes source data from images with a hand in a similar position, but the overlap leads it to draw extra. This can usually be solved with negative prompt of "extra fingers" though.

Part of my job is machine learning data generation. Essentially, to train a machine learning algorithm, you give it a bunch of source data. For example, if I wanted to train an Ai to recognize cars from an image I would give it 10s of thousands of images of cars. I would also give it image segmentation data, Essentially like another image where everything is black, and the car is one solid color. And then a json file that says, "the pixels that are red are the car". It's kind of like flashcards. The thing is, you could train two machine learning algorithms on the exact same data, and get to AIs that perform at different proficiencies. In the end, the Ai learns a pattern, it doesn't actually know what a hand or car actually is, it just knows there a a set of patterns that we want. Eventually you could tweak the data and get to a model that always does 5 fingers, but maybe messes up. A face since each thing is a seperate algorithm of patterns. Its hard to make an ai the is highly generalized and highly accurate.

Interesting.

An AI watermark?

Would be pretty fucking cool if all AI visualization tools always added a 6th finger as a method of built-in protection.