NIST AI Risk Management Framework
Primary guidance on governance, measurement, human oversight, and managing AI risks.
← Back to Compute College
Compute College
Human review and model-based graders
Combine human judgment and model-based grading without mistaking either one for ground truth.
Plain-English definition
Human review uses people to judge outputs against a rubric; a model-based grader uses another model to apply or summarize that rubric. Human review is slower and expensive but can handle ambiguity. Model-based grading is faster and more scalable but can share blind spots with the system being evaluated and must be calibrated against people.
Memory trick: Automation scales judgment only after judgment is calibrated.
Why it matters
A single automated score can hide disagreement, bias, or a grader that rewards fluent but unsupported answers. A review program should state which cases require people, how graders are calibrated, and when automated results are sampled for audit. This protects both quality and the operating budget.
- Human reviewers establish whether an automated grader is aligned with the intended rubric.
- Model graders can prioritize cases for review but should not silently define safety policy.
- Agreement, uncertainty, and escalation rates are useful metrics alongside the score.
Simple example
A customer-support assistant uses deterministic checks for format, a model grader for completeness, and human review for policy-sensitive advice. Ten percent of automated passes are sampled weekly; disagreements are labeled and added to the calibration set.
- Review effort is concentrated where the consequence or uncertainty is highest.
- The sample checks whether the grader drifts as models or traffic change.
- Escalation is part of the workflow, not an evaluation failure to hide.
Example figures are illustrative calculations, not current quoted market prices.
Current example
Primary reference
This source explains the method or risk boundary. It does not replace testing the actual model, data, provider, and workload conditions in production.
Source discipline: examples on this page are illustrative; measure the real workload before making a release or capacity decision.
Common mistake
A model grader is not an objective judge. It is another probabilistic system with preferences and failure modes, so validate it against people and keep critical decisions reviewable.
Practical takeaway
What you can do with this
Choose one nuanced criterion, write a rubric, collect a human-reviewed calibration set, and compare the model grader against it. Record disagreements instead of averaging them away.
- Use at least two reviewers for difficult calibration cases.
- Track inter-reviewer agreement before trusting an automated score.
- Route uncertainty and policy-sensitive cases to people.
Decision check: can you show where the automated grader agrees with trained reviewers and where it is not trusted?
Compute College learning path
AI Engineering
Step 23 of 48: Human review and model-based graders