Objective functions treat all misclassifications as equally costly

Evidence

By the lights of cross-entropy loss, misidentifying a stovetop as a car grille, or a green apple as a pear, or an English bulldog as a French bulldog is just as bad as characterizing a human being as a gorilla. In reality, certain types of errors—if only from the fiscal perspective of Google, to say nothing of the human so miscategorized—are probably thousands if not millions of times worse than others. — Brian Christian, The Alignment Problem, 1383

Categorization

• SEED: Essay angle: the moral of ‘error weighting’ — how technical choices about loss functions encode ethical positions, often invisibly.
• corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 1383