That's not the source of the problem. In layman terms, most fingers are flanked by two fingers and the ai does what is "statistically more accurate". It doesn't have a solid big picture context and gets lost in local detail, similar to kids' drawings. The role of training data is that hands are extremely expressive and commonly shown in very complex positions in the training images. AI is equally inept with complex body postures and contortions. But as it has been said, there are already models and workflows that improve the results significantly.