Entity-Level Hallucination

The model generates a named entity that does not exist in the source document or training data. This is the most severe form of entity-level hallucination.

• Example: Inventing a researcher named "Dr. Helena Vasquez" who never published the cited study
• Mechanism: The model samples from its parametric knowledge distribution rather than grounding in the provided context
• Detection: Cross-referencing generated entities against the source text reveals zero matches
• Risk: Creates entirely fictional authorities, products, or events that can propagate through downstream systems

The model replaces a correct entity from the source with a different but plausible entity, often from the same semantic category.

• Example: Attributing a quote to "Microsoft CEO Satya Nadella" when the source actually quoted "Google CEO Sundar Pichai"
• Common patterns: Swapping locations (Paris → London), dates (2023 → 2022), or person names within the same domain
• Root cause: The model's internal knowledge graph overrides the provided context during generation
• Detection difficulty: Substitutions are harder to catch than fabrications because the entity is real but contextually wrong

The model incorrectly segments or merges named entities, creating malformed references that distort meaning.

• Example: Splitting "New York University" into two separate entities: "New York" (location) and "University" (organization)
• Example: Merging "John Smith, CEO of Acme Corp" into a single entity "John Smith CEO of Acme Corp"
• Impact: Breaks entity linking pipelines and knowledge graph integration
• Measurement: Evaluated using Named Entity Recognition (NER) F1 scores against gold-standard annotations

The model loses track of which entity a pronoun or reference points to, causing identity confusion across a passage.

• Example: "Alice told Beth she would present" — the model later attributes the presentation to Beth instead of Alice
• Cascade effect: A single co-reference error can corrupt all subsequent entity mentions in long-form generation
• Evaluation: Measured by co-reference resolution accuracy on benchmarks like OntoNotes
• Mitigation: Explicit entity tracking mechanisms in the generation pipeline

The model generates chronologically impossible combinations of entities and temporal expressions.

• Example: Stating "President Obama signed the Civil Rights Act in 1964" — a real person linked to an event that occurred before his term
• Example: "The 2025 iPhone 15 release" — mismatching a future date with a past product
• Detection: Temporal reasoning benchmarks and constraint satisfaction checks
• Significance: Particularly dangerous in legal, financial, and historical applications where chronological accuracy is paramount

The model alters numerical entities—values, percentages, statistics—while preserving the surrounding context structure.

• Example: Changing "revenue grew 12%" to "revenue grew 20%" while keeping the rest of the sentence intact
• Example: "The study included 1,500 participants" becomes "The study included 15,000 participants"
• Risk profile: Numeric distortions can appear superficially plausible while being catastrophically wrong
• Measurement: Factual Precision metrics that specifically evaluate numeric entity accuracy against source documents