Hallucination Taxonomy

The most atomic form of factual error, where a model invents, substitutes, or conflates a specific named entity. This occurs when the generated text references a person, location, organization, date, or other proper noun that has no basis in the source context.

• Invented Entities: Generating a name like 'Dr. Alistair Finch' that does not exist in reality or the source material.
• Entity Substitution: Replacing 'Paris' with 'Berlin' when summarizing a document about the Eiffel Tower.
• Conflation: Merging two distinct entities, such as attributing a discovery by 'Company A' to 'Company B'.

This category is critical for high-stakes domains like legal document review and medical summarization, where a single incorrect entity can fundamentally alter meaning and create liability.

An error where the entities themselves are real and correctly identified, but the predicate linking them is fabricated. The model invents an interaction, association, or event that never occurred between the valid subjects.

• Fabricated Interactions: Stating 'Tesla acquired SpaceX' when both companies exist but the acquisition never happened.
• False Attribution: Claiming 'Marie Curie discovered the structure of DNA'—both the person and the discovery are real, but the relationship is false.
• Temporal Distortion: Asserting an event happened in the wrong sequence, such as 'The CEO resigned after the product launch' when the resignation preceded it.

Relation-level hallucinations are particularly dangerous in financial analysis and intelligence reporting, as they create plausible but entirely fictional narratives from real building blocks.

A hallucination where a single generated sentence directly contradicts another statement within the same output or contradicts the provided source document. This represents a failure of internal logical coherence.

• Intra-Output Contradiction: Paragraph one states 'Revenue increased by 10%' while paragraph three states 'The company reported a decline in revenue.'
• Source Contradiction: The source document says 'The patient has no known allergies,' but the summary states 'The patient is allergic to penicillin.'
• Negation Failure: The model drops a critical negation word, transforming 'The treatment was not effective' into 'The treatment was effective.'

This category is a primary target for Natural Language Inference (NLI)-based evaluation metrics, which are specifically designed to detect contradiction relationships between text pairs.

The generation of entirely unverifiable or fictional content that cannot be traced to any provided source or established world knowledge. Unlike entity or relation errors that distort existing facts, extrinsic fabrication creates information ex nihilo.

• Fictional Citations: Inventing a plausible-sounding academic paper title, author list, and journal name to support a claim.
• Imaginary Data Points: Generating a specific statistic like '78.3% of users prefer...' with no empirical basis.
• Hallucinated URLs: Creating a functional-looking web address that leads to a non-existent page.

This is the classic 'hallucination' in public discourse. Detection relies on semantic entropy techniques, as fabricated content exhibits high stochastic instability—the model will generate inconsistent details across multiple sampling runs.

A domain-specific category where a summarization model generates text that is grammatically fluent and factually plausible but not supported by the source document. The output diverges from the original text's meaning while maintaining surface-level coherence.

• Extrinsic Addition: Adding background context or commentary not present in the source, such as explaining the history of a company when the source only discusses its quarterly earnings.
• Semantic Distortion: Altering the emphasis or implication, such as summarizing a cautiously optimistic report as a definitive success.
• Omission of Key Constraints: Dropping qualifiers like 'potentially,' 'allegedly,' or 'in some cases' that are critical to the source's factual posture.

This is measured by faithfulness metrics and is the central challenge in abstractive summarization systems, where compression inherently risks information loss.

A failure to correctly anchor generated facts to the appropriate time period, resulting in anachronisms or outdated information presented as current. This is distinct from entity-level errors because the facts may have been true at some point but are no longer accurate.

• Anachronistic Statements: Claiming a living person holds a position they left a decade ago.
• Outdated Knowledge: Stating a current population figure using census data from 2010 without acknowledging the temporal context.
• Future Projection as Fact: Presenting a speculative forecast ('Revenue is projected to reach $5B by 2026') as a current or achieved state.

This category is critical for Retrieval-Augmented Generation (RAG) systems, where the freshness of the retrieved context directly impacts temporal accuracy. It highlights the difference between a model's parametric knowledge cutoff and real-time information needs.

A specific type of factual error where a language model invents or substitutes named entities—such as people, locations, or organizations—that do not exist in the source context. This is often the most immediately detectable form of hallucination because it involves concrete, verifiable nouns.

• Example: Stating 'Dr. Alistair Finch' won the Nobel Prize when no such person exists.
• Detection: Cross-referencing generated entities against a knowledge graph or source document.
• Impact: Severely damages trust in domains like legal document review or medical summarization.

An error where the entities are real, but the predicate or relationship connecting them is fabricated. The model correctly identifies two valid entities but asserts a false interaction, property, or temporal ordering.

• Example: Claiming 'Tesla acquired BMW in 2022'—both entities are real, but the acquisition relationship is false.
• Detection: Requires Natural Language Inference (NLI) to check if the relationship is entailed by the source.
• Nuance: Harder to detect than entity-level errors because the individual components pass basic validation.

A hallucination where a generated sentence directly contradicts another sentence within the same output or the provided source context. This represents a failure of internal logical consistency.

• Example: Paragraph 1: 'The patient has no known allergies.' Paragraph 2: 'Administer penicillin with caution due to the patient's allergy.'
• Detection: SelfCheckGPT or Chain-of-Verification (CoVe) can identify these inconsistencies by sampling multiple reasoning paths.
• Significance: Indicates a breakdown in the model's ability to maintain a coherent world state over long contexts.

A high-level taxonomic split based on the error's origin relative to the source material.

• Intrinsic Hallucination: The output contradicts the provided source. This is a failure of faithfulness. Example: A summary stating a product costs $50 when the source says $30.
• Extrinsic Hallucination: The output cannot be verified from the source. It adds unverifiable claims. Example: A summary adds 'the product is manufactured in Berlin' when the source never mentions location.
• Metric: Faithfulness Metric and Grounding Score are used to quantify intrinsic errors.

A taxonomy based on the operational context of the model.

• Closed-Domain Hallucination: Errors made when the model is restricted to a specific, provided corpus (e.g., summarizing a PDF). The ground truth is finite and known. Measured by Factual Consistency.
• Open-Domain Hallucination: Errors made when answering from general world knowledge (e.g., chat). The ground truth is vast. Measured by benchmarks like TruthfulQA or FActScore against a reference like Wikipedia.
• Risk Profile: Closed-domain errors are engineering failures; open-domain errors are often knowledge boundary failures.

A risk-weighted taxonomy metric that isolates hallucinations with high potential for harm. Not all factual errors are equal; this classification filters for errors that fundamentally alter meaning or pose safety risks.

• Examples: Fabricating a drug dosage, inventing a financial liability, or creating a false legal precedent.
• Detection: Combines standard hallucination detection with a domain-specific risk classifier.
• Application: Used in Guardrails frameworks like NeMo Guardrails to block outputs that contain these specific, high-severity error types before they reach the user.