You would do statistical analysis of the output. E.g. whether it is preferring a certain gender if gender is variable in the prompt, but everything else is the same etc.
Assuming reproducible settings are used, that will give you perfect insight into the behavior in the tested circumstances, but unless it lets you reconstruct the entire network, it won't tell you the impact that untested changes even if they are recombinations of tested elements behave. It is not perfect insight into the biases.