The internals of large models are hard to interpret. Yet evaluating every answer may not require seeing every internal mechanism. A recent preprint frames the problem as an exchange between someone making a claim and a person trying to verify it.
In this view, a model should provide evidence or steps that a person can examine. But verifiability is not free: hard questions take time and attention, and people have cognitive limits and get tired. The paper’s theoretical result depends on the assumptions of its framework.
For practical use, the questions are: what can be checked, how long does checking take, and who is responsible? An answer that looks explainable is not, by itself, evidence that it is correct. The review process should be evaluated separately from the answer.