While having racial bias in training data is a massive issue, it is actually a symptom of a much larger issue with neural networks: we have no idea why they do what they do. There is no actual way to be certain why the model is giving what output. Even if you track the data from one side of the model to the other, each parameter does not necessarily imply anything in particular. It is, after all, just a statistical model that is prone to the same issues that simply linear regression is, but it's no longer as simple as saying correlation doesn't imply causation. Often there are as many if not more variables than there is training data, and way more parameters. This means that overfitting is a massive issue and the model cannot be used to make inferences on data that is in any way outside of the domain of the data it was trained on. So the question becomes what are neural networks learning? Are they even learning or can we just give them enough randomly assigned parameters that we end up with a 'golden ticket' path in the network? All we can really say is that this network with these weights hits a (local!) minimum on some loss function.
Deepfakes and stuff seem like this crazy technology, but ultimately it's just a parlor trick. It's literally training the network to do this one thing for one face and one video clip. NNs mean nothing of substance and should NOT be used in any manner where they are depended on for the safety of a human and especially the public as a whole until these issues are solved.