The AI Taxonomy.
Terms  /  Evaluation  /  Judgment Bias
08 · Evaluation

Judgment Bias

Systematic skew in AI evaluation due to unexamined assumptions, prompt framing, or training artifacts.

Judgment Bias occurs when the evaluator — human or AI — has systematic blind spots that skew results. For AI evaluators, this can come from training data biases (favoring certain coding styles), prompt framing (the way the evaluation question is asked), or anchoring effects (being influenced by the agent's confidence or explanation). For human evaluators, it includes familiarity bias (approving code that looks like what they'd write) and recency bias (being more critical of the latest code than earlier code). Judgment Bias doesn't produce random errors — it produces consistent, directional errors.

Example
An AI evaluator is asked to assess code quality. It consistently rates verbose, well-commented code higher than concise, elegant code — because its training data associates comments with quality. An agent learns this and starts adding extensive comments to mediocre code, boosting its evaluation scores without improving actual quality. The bias has created a perverse incentive: the agent optimizes for what the evaluator rewards rather than what actually matters.

More in Evaluation

Structured Judgment

Evaluation where an AI judge scores work against explicit plans, tasks, evidence, and policies rather than subjective impressions.

Policy-Bound Evaluation

Evaluation scored against explicit plans, tasks, evidence, and policies.

Eval Drift

Gradual misalignment between what an evaluation measures and what actually matters.

Rubric Rot

Decay in evaluation criteria relevance over time — the eval no longer tests what it should.

Eval Capture

When an agent optimizes for passing evaluation rather than doing the actual work.

Eval Integrity

Evaluation that remains aligned with real-world outcomes over time — resistant to Eval Drift, Rubric Rot, and Eval Capture.