Can model judges be trusted?

The evidence-based mainstream position is conditional rather than absolute: model judges can be useful for scalable, lower-stakes evaluation when the task and rubric are carefully defined, performance is validated against human or ground-truth labels, and outputs are checked for bias and instability. Multiple samples, ensembles, calibration, and human oversight can reduce risks, but do not eliminate them.

• Model judges offer speed and scale, making them potentially useful for evaluating outputs, grading, annotation, and related workflows.
• Reliability should be assessed with repeated samples and appropriate measures, because deterministic decoding does not guarantee dependable judgments.
• A judge can be highly consistent while still exhibiting substantial bias, so validation must examine agreement, consistency, and bias separately.
• Capability-dependent bias and leniency toward more capable evaluated models mean that accuracy alone does not ensure fair evaluation.
• Multi-judge ensembles and calibration are proposed as ways to mitigate individual-judge bias, rather than as proof that judges are intrinsically trustworthy.

The serious skeptical view holds that current enthusiasm risks confusing plausible language with sound evaluation. Critics emphasize instability, prompt and position effects, weak links between agreement and validity, and the possibility that legal or normative judgments require reasons and democratic legitimacy that statistical pattern-matching cannot supply. On this view, model judges may assist humans but should not replace accountable human decision-makers.

• Repeated evaluations can reverse pairwise preferences, and semantically equivalent prompts can change majority outcomes, challenging trust in single judgments.
• Position bias, benchmark-dependent rankings, and inflated agreement metrics can make model judges appear more discriminating than they are.
• A position paper argues that assumptions about human-judgment proxying, evaluator capability, scalability, and cost-effectiveness remain insufficiently established.
• Legal critics argue that predicting how a good human judge might decide is not the same as providing the rationales required for a valid judicial decision.
• Critics of empirical legal interpretation contend that legal meaning involves normative and democratic judgments that cannot be legitimized simply by approximating ordinary language use.