Threshold Setting
Defining boundaries or cut-off values in AI decision rules (e.g., confidence scores) to balance risks like false positives versus false negatives.
Definition
Involves choosing decision thresholds—confidence, probability, or score cut-offs—that determine when an automated action is taken versus when human review is required. Thresholds are set based on risk tolerance, cost-benefit analyses, and stakeholder preferences. Governance processes require threshold-setting to be documented, tested across subgroups, and periodically recalibrated as conditions change, ensuring optimal trade-offs.
Real-World Example
A medical-diagnosis AI flags scans for biopsy recommendations only if its confidence exceeds 90%, reducing false alarms. Scans with 70–90% confidence are routed to radiologists for review. Quarterly calibration ensures these thresholds maintain a target under-diagnosis rate below 1% and over-diagnosis rate below 5%.
.svg)
