Types of AI Hallucinations: Intrinsic vs Extrinsic

The Intrinsic vs Extrinsic Distinction

The most important distinction in hallucination taxonomy is whether the model had access to the correct information. Intrinsic hallucinations occur when the correct information was present in the model's context window (system prompt, retrieved documents, conversation history) but the model's output contradicts it. If you provide a document stating a product costs $49 per month and the model responds with "$29 per month," that is an intrinsic hallucination. The model had the right answer and produced the wrong one.

Extrinsic hallucinations occur when the model generates claims that go beyond the provided context. If you provide a product description and the model's summary includes features that are not mentioned anywhere in the source material, those additions are extrinsic hallucinations. The model is not contradicting the source; it is adding information that the source does not support. Extrinsic hallucinations are more dangerous in many applications because they look like additional helpful information rather than obvious errors.

This distinction matters for mitigation because the two types have different causes and different solutions. Intrinsic hallucinations are attention failures: the model did not properly attend to the relevant part of its context. They become more common as context length increases, especially when relevant facts are buried in long documents. Solutions include placing critical facts early in the context, using structured formatting that highlights key information, and adding post-generation checks that verify output against input. Extrinsic hallucinations are generation failures: the model filled knowledge gaps with plausible fabrication. Solutions include constraining the model to answer only from provided context, requiring citations, and implementing fact-checking against external knowledge bases.

Entity Fabrication

Entity fabrication is when the model invents proper nouns: people who do not exist, companies that were never founded, products that were never built, papers that were never published. This is one of the most common and most damaging hallucination types because fabricated entities can be difficult to detect without domain knowledge and can lead users to take action based on nonexistent things.

The most well-documented case is citation fabrication. When asked to provide references, models frequently generate plausible-looking academic citations with real-sounding author names, journal titles that exist, and publication years that fall within expected ranges, but the specific papers do not exist. The citations are convincing because the model understands the format and conventions of academic citation, and it generates titles that sound like real research. A legal assistant that fabricated case citations led to court sanctions when the fabricated cases were included in actual filings.

Entity fabrication extends beyond citations. Models fabricate API endpoints that do not exist in a real library, configuration options that a tool does not support, historical events that never happened, and organizations that were never founded. In each case, the fabricated entity is plausible given the context, which is what makes it dangerous. A developer who receives a fabricated API method name will spend time trying to make it work before realizing it does not exist.

Knowledge graph grounding is the most effective mitigation for entity fabrication. When every entity claim must be verified against a graph of known entities and relationships, fabricated entities are caught because they do not appear in the graph. Persistent memory systems like Adaptive Recall maintain entity graphs that grow over time, providing an expanding verification layer that catches entity fabrication for any topic the system has previously encountered.

Numerical and Statistical Fabrication

Models frequently generate specific numbers, percentages, dates, and measurements that have no factual basis. This happens because the model understands the pattern of making quantitative claims (using numbers lends authority and specificity) but generates the actual values probabilistically rather than retrieving them from memory. Common examples include fabricated statistics ("studies show that 73% of users prefer..."), invented dates ("the feature was released on March 15, 2024"), approximate measurements presented as precise ("the latency is 42 milliseconds"), and financial figures that sound plausible but are entirely made up.

Numerical hallucination is particularly insidious because the numbers are often in the right ballpark. A model asked about an API's rate limit might say "1000 requests per minute" when the actual limit is 500. The hallucinated number is the right order of magnitude and sounds reasonable, which means users are less likely to question it than if the model said "one million requests per minute." This approximate-but-wrong pattern makes numerical hallucinations harder to detect through casual review.

Structured data lookup is the primary defense. When your system stores specific numerical facts in a knowledge graph or database, the model can be constrained to use those exact values rather than generating approximate ones. For questions where the precise number is not available in any knowledge base, instructing the model to avoid specific quantities and use qualitative language instead ("the rate limit is in the hundreds per minute") is more honest than allowing fabricated precision.

Temporal Hallucination

Temporal hallucinations involve incorrect claims about when things happened, what the current state of something is, or how things have changed over time. Models frequently confuse timelines, attribute events to wrong dates, describe current states using outdated information, and assert sequences of events in the wrong order.

The most common temporal hallucination is treating outdated information as current. A model trained on data through early 2025 might describe a software tool using information from a 2023 blog post, unaware that three major version updates have changed the tool's features, API, and pricing. The model's response is not fabricated in the sense that the information was true at some point, but presenting historical information as current is a hallucination that can lead users to make decisions based on outdated context.

Persistent memory directly addresses temporal hallucination by maintaining timestamped records of facts as they change. When a memory system records that a tool's pricing changed in March 2026, the model retrieves the current pricing rather than the historical pricing from its training data. The memory system's temporal awareness, knowing not just what is true but when it became true and what was true before, gives the model a timeline that its parametric knowledge cannot provide.

Relationship Fabrication

Relationship fabrication is when the model asserts connections between real entities that do not actually exist. Both entities are real, but the claimed relationship between them is invented. Examples include claiming that a researcher works at an institution they are not affiliated with, asserting that one software library depends on another when it does not, stating that a company is a client of another company when they have no business relationship, and claiming that two research papers cite each other when they do not.

Relationship fabrication is harder to detect than entity fabrication because all the named entities check out individually. If you verify the entities alone, everything looks correct. The error is in the relationship claim, which requires checking the specific connection between entities, not just the entities themselves. Knowledge graphs that store verified relationships are the primary defense: the graph can confirm that Entity A and Entity B both exist but that the claimed relationship between them is not recorded.

Synthesis Hallucination

Synthesis hallucination occurs when the model combines real information from multiple sources in a way that creates false claims. Each individual fact might be accurate, but the combination produces a statement that is not true. For example, if a model retrieves one document about PostgreSQL 15's features and another about PostgreSQL 16's performance benchmarks, it might synthesize "PostgreSQL 15 achieves the performance improvements shown in the benchmarks," combining real features from one version with real benchmarks from another to create a false claim about version 15.

This type of hallucination is especially common in RAG systems where multiple documents are retrieved and the model must synthesize information across them. The model treats all retrieved content as part of one unified context and freely combines information from different documents without tracking which facts came from which source. Citation requirements help because they force the model to trace each claim to a specific source, making cross-source fabrication more visible.

Choosing Mitigation by Hallucination Type

Different hallucination types respond to different mitigation strategies. Intrinsic hallucinations respond best to better context formatting and attention guidance. Entity fabrication responds to knowledge graph verification. Numerical fabrication responds to structured data lookup. Temporal hallucination responds to persistent memory with timestamps. Relationship fabrication responds to graph-based relationship verification. Synthesis hallucination responds to citation requirements and per-source attribution. A comprehensive anti-hallucination strategy addresses all types, but knowing which types are most common in your application lets you prioritize the most impactful mitigations first.