1.16.1
Types of Hallucination
What is it?
LLM hallucination occurs when a model generates fluent text that is incorrect or does not properly reflect the available information. The source identifies three types: intrinsic hallucination, extrinsic hallucination and faithfulness errors. Understanding these categories helps identify how an output can become unreliable.
Key points
1. Intrinsic hallucination: The output contradicts the provided input or context.
2. Extrinsic hallucination: The model generates claims that cannot be verified from the input and are factually wrong.
3. Faithfulness error: The output diverges from the instruction or specified constraints.
4. Hallucinations can occur even when an answer sounds confident and well-written.
5. Identifying the type helps select an appropriate detection method.
Identifying different hallucination types
Intrinsic: A source document says sales increased, but the model summary says sales decreased.
Extrinsic: The model adds a specific company revenue figure that is not supported by the provided document and is factually incorrect.
Faithfulness: The user requests a three-point summary, but the model returns a long essay that ignores the requested format.
Real-world application
When building AI assistants, classify output errors before deciding how to handle them. Compare summaries against source documents, verify factual claims and check whether the response follows the requested instructions.
Key takeaway
Hallucinations can contradict the source, introduce unsupported false claims or fail to follow instructions.