Borrowing from disciplines like economics could help AI safety researchers control AI models, even if they don’t understand them.