Testing AI When Specifications Cannot Capture RealityThe hard testing problem in AI is epistemic: a system can satisfy every contract and still miss the relationship that matters. Reliable practice combines automated checks with evidence, counterexamples and ongoing evaluation.
