Intrinsic and Extrinsic Hallucination Against a Source Document
Understanding the Two Forms of Hallucination
Intrinsic and extrinsic hallucinations represent two distinct categories of incorrect information that can appear in AI-generated responses. Intrinsic hallucinations occur when the AI creates information that contradicts or goes beyond what is stated in the source document. These errors typically involve fabricating facts, dates, names, or claims that simply do not exist in the provided materials.
Extrinsic hallucinations happen when the AI incorporates information from sources outside the provided document, creating responses that seem plausible but are actually based on external knowledge or assumptions rather than the specific content being referenced.

Consider a scenario where a legal assistant receives an AI-generated summary of contract terms. The AI might claim that a contract was signed on a specific date that does not appear in the document, or it might reference clauses that are not present. This represents an intrinsic hallucination.
Alternatively, if the AI mentions a regulatory framework that exists but was not included in the contract materials provided, this constitutes an extrinsic hallucination. Both forms can lead to serious consequences in professional settings where accuracy is essential.
Practical Examples in Workplace Contexts
In healthcare documentation, an AI system might generate patient notes containing fabricated medical conditions or dates that do not appear in clinical records. A clinical assistant reviewing such output must identify whether the AI has created information that contradicts existing patient data. Similarly, in financial reporting, an AI might produce figures or references to transactions that exist nowhere in the provided financial documents.
These errors can have significant implications for compliance and decision-making processes.
- Legal professionals must verify that AI-generated case summaries do not contain fabricated precedents or dates
- Medical coders need to ensure AI responses don’t invent diagnosis codes or patient history elements
- Financial analysts must check that AI-generated reports don’t reference external market data not provided in source documents
- Human resources specialists should validate that AI responses don’t create fictional employee details or policy references
Accounting firms often encounter these issues when AI systems process client data. An AI might claim that a particular tax regulation was applied in a specific year, when the actual documents show no such application. Marketing teams face similar challenges when AI generates campaign data or customer demographics that don’t align with provided market research.
The key is to develop systematic approaches for identifying these discrepancies through careful comparison against source materials.
Detection Strategies and Verification Techniques
Effective detection requires establishing clear verification protocols that focus on source document alignment. Practitioners should develop checklists that examine whether every claim, date, name, or reference exists within the provided materials. When AI responses contain specific details, these must be traceable to the source documents through direct quotation or clear reference points.
The verification process involves cross-referencing AI output against the original content to identify any elements that cannot be substantiated.
Training staff to recognise these patterns involves creating specific examples of both hallucination types. For instance, showing how an AI might claim a company was founded in 1995 when the document states 2001, or how it might reference a policy that exists but was not included in the provided materials.
Regular practice with these examples helps develop the critical thinking skills needed for accurate verification. The goal is to create a systematic approach that makes hallucination detection a routine part of quality assurance processes.
Organisations should implement structured review processes that include multiple verification steps. These processes must account for both the content of the AI output and its relationship to the source documents. Regular training sessions help maintain awareness of these issues and reinforce proper verification techniques.
The focus should always remain on ensuring that AI-generated responses align precisely with provided materials rather than introducing new information or assumptions.
The Decision Maker's Handbook to Data Science
The definitive guide for non-technical executives and entrepreneurs on navigating AI, machine learning, and big data.
