When it comes to AI, source is only part of the system. The trained model would be a black box. We don't fully understand why some NNs work - so something of this complexity can't simply be cracked open and vetted accurately.
Barring military secrets, perhaps, there's no reason the AI program's current state would necessarily have to be a black box. Nor would it need to be a neural network. I also disagree with the assertion that complex things cannot be vetted.
