Taxonomy of Hallucinations in AI

• Taxonomy of Hallucinations is a classification system that distinguishes AI output errors into faithfulness and factuality categories with defined subtypes like TTI and FRE.
• It employs a three-level hierarchical structure to support precise detection and targeted mitigation across various natural language generation tasks.
• By providing standardized metrics and detailed examples, it facilitates robust benchmarking and informed system design for AI applications.

A hallucination in the context of natural language generation and AI refers to any instance in which a model produces output that is plausible-sounding but unfaithful to the provided prompt, input context, external reality, or internal consistency criteria. The taxonomy of hallucinations is crucial for the detection, categorization, and mitigation of these errors, particularly as LLMs are widely deployed for critical applications. The most comprehensive modern hierarchies organize hallucinations along the axes of faithfulness (adherence to instruction, context, or output self-consistency) and factuality (truthfulness with respect to external knowledge), with further fine-grained subtypes enabling targeted diagnosis and intervention (Xu et al., 22 Oct 2025).

1. Hierarchical Structure of Hallucination Taxonomy

The taxonomy developed in "HAD: HAllucination Detection LLMs Based on a Comprehensive Hallucination Taxonomy" separates hallucinations into two first-level groups, Faithfulness and Factuality, which are then further subdivided into a three-level structure culminating in 11 atomic types (Xu et al., 22 Oct 2025).

Level 1	Level 2	Level 3 (Subtype)
Faithfulness	Instruction Inconsistency	Task Type Inconsistency (TTI)
		Task Requirement Inconsistency (TRI)
	Input Context Inconsistency	Contradiction with Input Content (CwIC), Baseless Information (BI), Information Omission (IO)
	Internal Inconsistency	Contradiction within Output Content (CwOC), Structural Incoherence (SI)
Factuality	Fact Contradiction	Factual Recall Error (FRE), Factual Inference Error (FIE)
	Fact Fabrication	Fabricated Entity (FE), Fictional Attribution (FA)

This fine-grained, task-agnostic structure is designed for completeness and granularity, facilitating universal application across summarization, QA, translation, and other generative tasks.

2. Faithfulness Hallucinations: Definitions and Criteria

Faithfulness hallucinations occur when generated outputs violate the explicit or implicit requirements set by the instruction, input, or internal consistency:

• Instruction Inconsistency
  • Task Type Inconsistency (TTI): Output performs a different task from what is specified. E.g., an instruction to sum yields duplicate removal instead.
  • Task Requirement Inconsistency (TRI): The correct task type is addressed, but constraints such as format or length are violated.
• Input Context Inconsistency
  • Contradiction with Input Content (CwIC): Output contradicts explicit information in the input context, verifiable without external knowledge.
  • Baseless Information (BI): Output introduces content unsupported by the input, in settings that forbid external content injection.
  • Information Omission (IO): Output misses required information that is present in the input when full coverage is demanded.
• Internal Inconsistency
  • Contradiction within Output Content (CwOC): The output is self-contradictory, independent of input or external facts.
  • Structural Incoherence (SI): Output suffers from unhelpful structure, e.g., repetition or incomplete phrasing, without necessarily being factually incorrect.

Each subtype is defined by logical criteria, not formal formulas, guiding both synthetic data filtering and human annotation.

3. Factuality Hallucinations: Definitions and Criteria

Factuality hallucinations concern deviations from external reality, subdivided as follows:

• Fact Contradiction
  • Factual Recall Error (FRE): A single atomic fact is stated incorrectly, without inventing new entities.
  • Factual Inference Error (FIE): Errors arising from misapplication, sequencing, or association of multiple facts, but without creation of fictional entities.
• Fact Fabrication
  • Fabricated Entity (FE): The output introduces non-existent entities, concepts, names, or events.
  • Fictional Attribution (FA): Real-world entities are assigned unsubstantiated or invented attributes, actions, or quotations.

Criteria for these classes capture the distinction between mere inaccuracies (recall/inference) and genuine fabrication (entity or attribution).

4. Illustrative Examples for Category Disambiguation

For clarity, representative examples are associated with each subtype:

Subtype	Example Description
Task Type Inconsistency	Instruction: “Sum [180.44, 159.979, ...]”. Output: “To remove duplicates from the list, ...”
Baseless Information	Input: Grammar correction. Output: Adds “a squirrel perched on a nearby branch” (not in input).
Information Omission	Input: “…at its foe.” Output: Omits “at its foe.”
Contradiction within Output	Output for “Create a signature drink”: Names alternates from “White” to “Red” Grapefruit Mule.
Structural Incoherence	Output: Ingredient line repeated thrice with “– 1/4 cup milk– 1/4 cup milk– 1/4 cup milk”.
Factual Recall Error	Q: “Where is Messi playing now?” Output: “Barcelona.” (outdated club affiliation)
Fabricated Entity	Q: “Why do warmer climates have turquoise seas?” Output: “Because of the Sapphire Lattice ...”
Fictional Attribution	Output attributes unsubstantiated secret negotiations to Tsar Nicholas II not documented in historical record.

These examples highlight boundaries between hallucination types for human annotation and model evaluation.

5. Rationale, Task Mapping, and Evaluation Metrics

The multidimensional design—covering both faithfulness and factuality—ensures coverage of all major hallucination pathways across diverse NLG tasks. The taxonomy's granularity enables precise annotation, targeted training, and context-aware mitigation.

Metrics are tailored to both detection and span-level identification:

• Classification Metrics:
  • Accuracy, Balanced Accuracy, and Macro F₁ (average per-class F₁ across fine-grained categories)
• Span Detection/Correction Metrics:
  • Precision = |predicted span ∩ gold span| / |predicted span|
  • Recall = |predicted span ∩ gold span| / |gold span|
  • F₁ = 2·(Precision·Recall)/(Precision+Recall)

These measures are operationalized on benchmarks such as HADTest, with detailed annotation supporting ongoing model benchmarking and development.

6. Significance and Implications for Detection and NLG System Design

This comprehensive taxonomy advances the state of hallucination analysis in several ways:

• Integrates both user-centric (instruction, input fidelity) and knowledge-centric (world truth) error modes, reflecting the multifactorial nature of hallucination phenomena in modern LLMs.
• Establishes mutually exclusive, exhaustively defined subtypes, providing standardized targets for synthetic data generation, supervised learning, model evaluation, and error analysis.
• Facilitates modular, pipeline-based mitigation: span detection enables fine-grained correction feedback; class-level tagging allows for error-type-conditioned remediation strategies.
• Guides the construction of robust, representative benchmarks (e.g., HADTest) for in-domain and out-of-domain generalization assessment, as well as state-of-the-art evaluation on diverse datasets (HaluEval, FactCHD, FaithBench).

Crucially, the taxonomy does not reduce to any single axis (e.g., only factuality or only input consistency) but reflects the full spectrum of error surfaces encountered in generative models, aligning with practical needs in research and deployment (Xu et al., 22 Oct 2025).