New AI model reveals 4-phase bias cycle in subway flood dispatchers

A team of researchers led by Jinfeng Lou and Pingbo Tang (Carnegie Mellon University) and Cleotilde Gonzalez have developed a computational cognitive model that simulates how human operators adapt their decisions during extreme urban infrastructure events – specifically subway flooding. Published on arXiv (2606.06429), the model integrates Instance-Based Learning Theory (IBLT) into civil engineering computing, creating an architecture where operator decision processes are driven by memory retrieval and utility blending. This framework serves as a rational baseline for boundedly rational adaptation, allowing the algorithmic isolation of latent psychological biases like overconfidence and defensive overcorrection.

The team validated the model using a human-in-the-loop microworld experiment mimicking subway flood-induced track suspensions. Dispatchers had to balance passenger safety against service efficiency. Analysis revealed a complex four-phase adaptation cycle: acquisition, overconfidence, overcorrection, and recalibration. Crucially, during the post-accident overcorrection phase, human operators showed immediate, defensive risk overestimation, while the computational model maintained a stable trajectory based on accumulated experience. This divergence proves that operational instability after failure is often driven by acute psychological bias overriding stable memory-based adaptation – a pattern expected to recur across high-stakes environments like air traffic control or power grid management.