Hallucination

AI concepts
About 1 min read

A phenomenon in which an AI model, in the process of predicting the next word based on statistical probability, logically and confidently generates incorrect or groundless information as if it were the truth.

Also known as
HallucinationAI HallucinationHallucination

Detailed explanation

Hallucination occurs when an LLM relies solely on patterns in its training data, and is characterized by presenting fabricated content in highly plausible sentence structures rather than just simple incorrect answers. It is broadly categorized into 'Intrinsic Hallucination', which ignores or contradicts the provided context, and 'Extrinsic Hallucination', which fabricates information that is not in the context and is factually incorrect. According to recent research, more advanced reasoning models can exhibit more complex forms of hallucination, making it essential in business environments to manage this through technical safety measures such as RAG (Retrieval-Augmented Generation) and Guardrails.

Why it matters in tool selection

When selecting enterprise AI tools, the hallucination rate is directly linked to the 'cost of trust' rather than being a simple performance metric. Hallucinations can lead to legal risks, decline in brand trust, and decision-making errors. Especially with surveys indicating that, as of 2025, knowledge workers spend an average of over 4 hours weekly verifying AI responses, selecting tools that can mitigate hallucinations and clearly provide citations for their answers is key to improving productivity.

What to check when adopting tools

  • Does it provide specific source citations and links for each sentence of the response?
  • Does it natively support RAG (Retrieval-Augmented Generation) architectures based on real-time data?
  • Are there settings to make it answer 'I don't know' or refuse to answer when confidence is low?
  • Is a dedicated Guardrails system built-in to detect and block hallucinations?

Examples of major hallucination types

- Intrinsic Hallucination: When the input meeting minutes state 'revenue of 1 billion', but the summary incorrectly writes 'revenue of 10 billion'. - Extrinsic Hallucination: When it fabricates and cites non-existent research paper titles or legal case numbers.

Confusing terms

Confabulation

A psychological term referring to the phenomenon of unconsciously fabricating stories to fill gaps in memory; its mechanism is highly similar to AI hallucination, and it is sometimes proposed as an alternative term.

Bias

The phenomenon of providing information biased toward a specific group or perspective regardless of factual accuracy, falling under the issue of information 'fairness'.

Related terms

RAGGrounding