But they might be reliable at other tasks. We should not presume they are or are not, we should figure it out by testing. If they are reliable, then we should use them. If not, we should not.

The fact that they are black boxes is also a red herring, imo. First of all, humans are black boxes and we use them to do science all the time. But so too are many tools: I don't know how my computer works at any serious level of detail. But I'm still allowed to do science on it.