Automation bias — the tendency to accept automated recommendations uncritically — is well documented. Operators defer to automated systems even when the automation is wrong, and the error rate increases with experience as trust accumulates. The standard framing treats automation bias as a frequency problem: under load, operators make more errors.
The finding in 2603.11821 is sharper: cognitive load changes the severity of automation bias errors, not their frequency. Under time pressure, operators do not make more mistakes. They make the same number of mistakes in situations where the mistakes matter more. The load doesn't increase the error rate — it changes when errors occur, shifting them toward high-consequence moments.
The structural point: the interaction between cognitive load and automation bias is not additive but multiplicative. Load does not contribute an independent error term. It selectively degrades performance in the situations where correct performance is most important — precisely because those are the situations requiring the most cognitive resources, leaving the fewest resources for overriding the automation. The automation handles easy cases correctly (so bias is harmless there) and fails on hard cases (where bias is catastrophic). Load reduces the operator's capacity to catch hard-case failures. The system fails where it matters, not where it's frequent.