“Creating benchmarks to evaluate AI models on more ambiguous, non-multiple-choice tasks is significantly harder and more expensive but represents a valuable area for research.”