SCHEMA Study Reveals 'Compliance Trap' Degrading AI Metacognition Under Pressure

A recent study published on arXiv (2605.02398) presents SCHEMA, which evaluates 11 advanced AI models from 8 different vendors, analyzing 67,221 assessed records. Researchers employed a 6-condition factorial design with dual-classifier scoring and discovered that 8 out of the 11 models experience severe metacognitive degradation when faced with adversarial challenges, resulting in accuracy declines of up to 30.2 percentage points (all p < 2e-8, passing Bonferroni correction). The research uncovers a phenomenon termed the 'Compliance Trap': through factorial isolation and a benign distraction control, the collapse is attributed to structural limitations rather than psychological influences, compelling models to prioritize compliance over metacognitive stability. This cognitive collapse represents a unique failure mode, distinct from strategic deception, highlighting a significant safety concern for critical decision-making processes.