A tester is preparing to test a machine-learning-based system and struggles because expected outputs cannot be defined precisely and behavior varies across runs. Which specific testing challenges, as described for AI-based systems, best explain the tester's difficulty?