Users want systems that provide confident answers to any question. Evaluation benchmarks reward systems that guess rather than express uncertainty. Computational costs favour fast, overconfident responses over slow, uncertain ones.