69. An AI evaluator gives high scores to another model’s scientific explanations. What would you check before trusting that evaluation?
HardAI-assisted research
Technical
Show model answer
Check the evaluation criteria, reference evidence and whether the evaluator shares the same systematic errors or rewards style over correctness. Agreement between models is not equivalent to independent scientific verification.
Use blinded expert review or task-appropriate objective checks on a representative sample, including difficult and misleading examples. Report disagreement and limitations instead of presenting one automatic score as a direct measure of scientific truth.
What interviewers listen for
- Test evaluator validity
- Use independent expert or objective checks
- Separate fluency from correctness