I love all these colorizations, especially videos (easy to find on YT).

One thing I've been wondering about: Is it possible that the ML algo turns the faces into more "modern" faces because it's presumably trained on later color footage?

It's an inherent problem of all ML systems to extrapolate from training data only.

You can get dramatically different results depending on your training data set.

The only real issue with that is treating any of the outputs as "realistic" representations in terms of "close to the original scene". It's just a modern interpretation and the viewer needs to keep that in mind at all times.