In fairness, there is no contradiction between "most trustworthy" and it being naïve to then actually trust them.
Our ability to debug or decompile artificial neural networks is only better than doing so for living neural networks because connectome scans are expensive, not because we designed the high-level architecture for the artificial ones and have all the weights in a convenient easy-to-read format: even our best experts in this are still presently akin to a drunk looking under a lamppost for their keys because that's where the light is not where they dropped them.
(They are also well aware of this and will tell you much the same, this is why so many are concerned about AI alignment etc.)