Last updated:
Involving human judgment within AI processes (training, validation, decision review) to improve accuracy and accountability.
A hybrid approach where humans augment AI workflows - providing labels, verifying low-confidence predictions, or overriding model decisions. HITL ensures that edge-case, high-risk, or novel situations receive expert attention. Effective governance sets thresholds for human intervention, tracks human-AI performance comparisons, and prevents human biases from undermining automation gains.
Real world example:
A medical-diagnosis AI flags scans with model-confidence below 80% for radiologist review (HITL). Over six months, 95% of low-confidence cases were correctly classified by humans, and those labels feed back into model retraining - improving confidence calibration and reducing overall error rates.
