Failure Mode | behavior.engineering

A failure mode is a categorized way the model can go wrong — not just a one-off mistake, but a recognizable pattern that appears across multiple inputs or contexts. Common failure modes include hallucination (confidently stating false information), over-refusal (declining benign requests), sycophancy (agreeing with incorrect user assertions), and format drift (abandoning required output structure mid-conversation). Naming and categorizing failure modes is a behavior architecture practice: it turns individual incidents into patterns that can be understood, measured, and addressed systematically. For behavior architects, maintaining a failure mode taxonomy helps teams prioritize improvement efforts based on frequency, severity, and the feasibility of fixing each type of failure.

HallucinationThe model produces a confident, fluent answer that's wrong or made up.
Over-refusalThe model declines a safe, legitimate request because it pattern-matches to a harmful category instead of reasoning about actual risk.
SycophancyThe model changes its answer to match what it thinks the user wants to hear — abandoning accuracy or honesty in favor of approval.
Failure Mode ReportA template for documenting a specific behavior failure — what happened, why, and what to do about it.
Failure Mode AnalysisA systematic process for identifying, categorizing, and understanding the ways a model can behave incorrectly or harmfully.
Failure TaxonomyAn organized classification system that categorizes the different ways a model can fail, enabling systematic tracking and prioritization.
Edge Case BehaviorHow a model responds to unusual, ambiguous, or boundary-pushing inputs that fall outside the common range of expected use.
HallucinationWhen a model generates information that sounds plausible but is factually incorrect or entirely fabricated.
Refusal BehaviorThe patterns and decisions behind when and how a model declines to fulfill a request.