Hallucination Taxonomy: Types, Causes, and Targeted Mitigations

AI hallucinations fall into four distinct categories with different root causes and mitigations. Factual confabulation is the classic type: the model invents specific facts — a citation, a statistic, a person's biography — with no grounding in training data. Root cause: the model has learned to produce fluent, contextually appropriate text and generates plausible-seeming specific claims when it lacks actual knowledge. Best mitigation: RAG with source citation requirements and post-hoc fact verification. Reasoning hallucinations occur when the model's chain-of-thought contains logical errors that cascade into wrong final answers — the steps look plausible but a key inference is incorrect. Root cause: models learn to generate reasoning-sounding text, not to verify logical validity. Best mitigation: multi-step verification prompts, self-consistency sampling, and programmatic verification of checkable steps. Instruction-following hallucinations happen when the model's output violates the explicit constraints given in the prompt — wrong format, wrong length, forbidden content. Root cause: instruction following competes with other training objectives and fails under constraint density or unusual combinations. Best mitigation: structured output schemas with validation, and breaking complex instruction sets into sequential prompts. Temporal hallucinations are a category increasingly relevant with deployed models: confidently asserting facts that were true at training time but are now outdated. Best mitigation: web-grounded retrieval for time-sensitive claims and explicit cutoff disclosure.