Human-in-the-loop simulation

Measure how the confidence threshold, reviewer capacity and accuracy, and automation bias affect final error per group.

AdvancedResponsible AI Lab
العربية

The idea

Cases with stated model confidence are generated; the least confident go to a human reviewer within capacity.

Data description

Synthetic data generated in your browser from stated distributions; no real personal data.

Suggested next experiment

Concepts in this experiment

Human oversight (HITL)

Involving a person able to understand and change the decision. Effectiveness depends on capacity, information, authority, time and avoiding automation bias.

Automation bias

The human tendency to accept automated recommendations without enough scrutiny, making oversight a formality. Mitigated by training, showing uncertainty and review designs that present evidence before the recommendation.

Scientific sources for this experiment

  1. Artificial Intelligence Risk Management Framework (AI RMF 1.0)NIST (2023) — Framework

More in this module