A FaceBook on SAFE could actually work in a way where each user owned their own data. For training an AI to do for example facial recognition that would have some disadvantages, though perhaps there might be way around them.
Google Photos can train on all the faces of all photos of millions of Android users. Apple Photos take a different approach where they to do it locally on the app and is thus more privacy preserving, it doesn’t work as well as Google’s, but it does work.
So in practice training an AI on SAFE could be kinda like Apple Photos, it would use the users own computer with the users own data. Perhaps a way could be found to extract features from a trained network and combine it with those of other networks in a privacy preserving way, would be a bit of research to figure out if that is possible or not I guess.
Once SAFE get compute capabilities the network itself could be used for the training, instead of only the users own compter, but then the data would have to be decrypted so the nodes doing the computation could potentially see the data. Fully homomorphic encryption could be a way around this if someone finds a way to do it efficiently. Perhaps another way could be to split the data into small parts and spread it to random nodes to do the computations. You would decrypt the data on a trusted node that you run, then split the data into small parts and do non-sensitive computations that could be combined again on the trusted node.