Skip to content
deen
← All terms
Working with language modelsLast reviewed

Hallucination

Hallucination is when a language model makes up statements that sound convincing but are false or unsupported.

Also known as: Confabulation

How it works

A language model predicts chunks of text, it does not check facts. If it lacks knowledge or the question is unclear, it fills the gap with whatever sounds plausible. This is how invented numbers, false source references or claims that sound like facts come about.

A practical example

You ask for a study on a topic and get authors, title and year — but that study does not exist.

What you should know

  • Hallucinations can be reduced but not ruled out completely.
  • Helpful measures: base answers on your own sources (RAG), have sources shown, write precise prompts and have a human check important content.
  • Especially with numbers, quotes, and legal and medical topics: verify before you pass anything on.

Matching tools

View all termsTo AI integration

From the blog

More on this from practice.

Articles from our blog that explain the term in context.

Artificial Intelligence

ChatGPT in Business: What Works — and What Does Not

Three forms of use, typical use cases, common pitfalls and a five-step approach: how to use ChatGPT and similar models in your company in a sensible way.

Read on →

Related terms

More on the topic.

These terms are closely related to this entry.

Working with language models

Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation (RAG) connects a language model with your own sources: for each question, matching passages of text are looked up and handed to the model as the basis for its answer.

Learn more →

Agents & automation

Human-in-the-Loop

Human-in-the-loop means that a person checks or approves important steps before an AI decision takes effect.

To the definition →

Fundamentals

Large Language Model (LLM)

A large language model (LLM) is a language model trained on enormous amounts of text. It generates text by predicting, piece by piece, the most likely next chunk of text.

Learn more →

Questions about your project?

Get in touch →