There are no benchmarks worth the paper they're written on, so in a real sense nobody knows what they're talking about -- but my intuition is this as well.

If you use normal software, you are dealing with a constrained API. std::vector has .push_back() but has no .launch_icbm(), so it's rather easy to tell what it can or can't do.

These things come with an "ask me anything" interface, but that doesn't mean they can do everything, just that it's harder to see the limits.

Clever Hans effect + black box inside + unconstrained API + nondeterminism.