Hallucination & Mitigations
Mitigation 4: Evaluation + Human Review
PM perspective: quantifying hallucination risk, designing tiered review workflows, and making Human-in-the-Loop your product's safety net. Strategy 04 / 04 · Detect + Correct.
Three Stages Every PM Must Design
Stage 1 · Pre-launch
Evaluation Baseline: Build a Hallucination Test Set
Test the model with questions that have known correct answers, quantify hallucination rate, and set acceptance thresholds.
Stage 2 · Post-launch
Tiered Review: Human Oversight for High-Risk Content
Automatically route by risk level: low-risk responses go out directly; high-risk ones are reviewed by humans first.
Stage 3 · Continuous Improvement
Error Feedback Loop: Iterate with Live Data
Collect hallucination cases found during review as Bad Cases, feeding them back into model optimization and Prompt refinement.
Key PM Insights
Key PM Insights
Hallucination rate ≠ 0: Every LLM hallucinates. The PM's goal is to keep it within business-acceptable thresholds—aiming for zero is unrealistic.
High-risk = human safety net: In healthcare, legal, and financial contexts, AI only produces a draft; a human must sign off on the final output.
Metrics belong in the PRD: "Hallucination rate < 3%" should be a measurable acceptance criterion—just like "load time < 2s".