Instruction-gated vulnerability to external reasoning traces in large language models

Decades of research on automation bias have shown that humans defer to external advice even when incorrect—yet whether this vulnerability increases or decreases with capability remains unresolved. Here we reveal a capability-compliance paradox: high capability can paradoxically increase vulnerability to external influence. Using language models as a controlled testbed enabling counterfactual comparisons impossible with human participants, we investigate across four studies how agents respond to reasoning traces containing deliberate errors. We reveal a phenomenon we term compliance-induced failure (CIF): the abandonment of correct independent reasoning in favor of erroneous external input. Study 1 (N = 99, 3 models) demonstrates that high-capability agents are most vulnerable: Claude 4 Sonnet showed 40.0% CIF with zero recovery, while low-capability Claude 3.5 Haiku showed the reverse pattern (11.4% CIF, 90.9% recovery). Study 2 (N = 99) shows that explicit verification cues—a metacognitive intervention—eliminate CIF entirely, restoring performance from 57.6% to 96.0%. Study 3 (N = 198, 3 models) reveals that contamination type matters critically: coherent-but-wrong reasoning causes catastrophic failures (GPT-4o: 7.1%), while noise-like contamination shows milder effects. Study 4 (N = 199, 6 models) confirms the capability-compliance paradox across a broader range of agents. These findings establish a general principle: capability itself can become a liability when agents defer to external guidance without verification. This has immediate implications for human-AI interaction, expert decision support systems, and any context where capable agents receive external advice.