Unfortunately, the nature of the models is that it’s very difficult to get an understanding of the innards. That’s part of the point, you don’t need too. The best we can do is monitor how it’s built and what connects in and out of it.

The open source bits let you see that it’s not passing data on without your permission. If the training data is also open source, you can get for biases e.g. 90% of faces being white males.