SRC01-E01 — Hallucination Taxonomy¶
Extract¶
The survey provides a taxonomy of LLM hallucinations including: intrinsic hallucinations (contradicting input context) vs. extrinsic hallucinations (inconsistent with training data or reality), and factuality (absolute correctness) vs. faithfulness (adherence to input). Specific manifestations include "factual errors, contextual and logical inconsistencies, temporal disorientation, ethical violations, and task-specific hallucinations across domains like code generation and multimodal applications." Hallucinations in LLMs "appear in various forms, which are easier or harder for humans to detect, and can be categorized and organized according to how challenging it is to detect them, representing a spectrum of types with increasing severity." The paper identifies contributors across "data, training, and inference stages."
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Contradicts — training treats hallucinations as monolithic; this shows they are a spectrum | Strong |
| H2 | Contradicts — training does mention hallucinations, though superficially | Moderate |
| H3 | Strongly supports — the rich taxonomy known to research is absent from training | Strong |
Context¶
This is the most comprehensive academic taxonomy of LLM hallucinations. The detection-difficulty spectrum is central to Q003 because corporate training treats all hallucinations as equally detectable random errors.
Notes¶
The detection-difficulty spectrum is the critical insight: some hallucinations are obviously wrong, but others match user expectations and are nearly undetectable without independent verification. No corporate training material examined conveys this spectrum.