Human-like performance on narrow benchmarks (word labeling, sentiment) systematically underdetermine...