I think it's much more likely whatever scraping they used to get the training data snatched a screenshot of the movie some random internet user posted somewhere. (To confirm, I typed "joaquin phoenix joker" into Google and this very image was very high up in the image results) And of course not only this one but many many more too.
Now I'm not saying scraping copyrighted material is morally right either, but I'd doubt they'd just feed an entire movie frame by frame (or randomly spaced screenshots from throughout a movie), especially because it would make generating good labels for each frame very difficult.
Understanding the math behind it doesn't immediately mean understanding the decision progress during forward propagation. Of course you can mathematically follow it, but you're quickly gonna lose the overview with that many weights. There's a reason XAI is an entire subfield in Machine Learning.