Failure Mode

A failure mode is a specific, identifiable way in which a system can fail to meet its intended behavior. The term originated in reliability engineering and failure-mode-and-effects analysis (FMEA), where it denotes each distinct way a component, subsystem, or process can break down, and is used to prioritize mitigation, design reviews, and redundancy. In software engineering and systems design, a failure-mode analysis enumerates the conditions under which a system produces incorrect, unsafe, or unexpected output, often prior to deployment.

In AI and machine learning, the term has been adopted to describe recurring categories of model and system breakdowns — for example, hallucination, reward hacking, prompt injection, distribution shift, mode collapse, and adversarial vulnerability. These are not bugs in the conventional sense but predictable, structural failure patterns that emerge from the gap between training objectives, deployment distributions, and the limits of generalization. AI failure modes are a central concern of AI safety, alignment research, and robust-ML engineering.

This entry is recorded as adjacent to the existing AI/ML entries in this vocabulary — most directly ai-failure-modes (which captures the umbrella category) and alignment-failure (a specific instance). The generic term "failure mode" itself is included for cross-reference completeness: its primary home is reliability engineering, not AI/ML, and it sits on the boundary of this vocabulary's scope rather than in its core.