Hallucination

If you subscribe to a service from a link on this page, we may earn a commission.

In artificial intelligence, a hallucination is an output that is fluent and confident but factually wrong, fabricated or unsupported by any source. A hallucinating chatbot may invent statistics, cite academic papers or court cases that do not exist, attribute made-up quotes to real people or describe product features that were never released – all in perfectly plausible language.

Hallucinations happen because of how large language models work. An LLM does not look facts up in a database; it generates the statistically most likely next word based on patterns learned during training. When the model lacks reliable knowledge about a topic, the most probable-sounding continuation can be entirely false. The confident tone is part of the problem: the model presents guesses and facts with the same fluency.

Common causes and contributing factors include:

  • Gaps or errors in training data – The model cannot know events after its training cutoff or facts that were rare, wrong or missing in its data.
  • Ambiguous or leading prompts – Questions that presuppose false information (“describe the 2019 merger between X and Y”) often get a fabricated answer rather than a correction.
  • Optimization for helpfulness – Models trained to be agreeable and complete may prefer producing an answer over admitting uncertainty.

Strategies to reduce hallucinations include:

  • Retrieval-augmented generation (RAG) – Connecting the model to a search engine, knowledge base or document store so answers are grounded in retrieved sources.
  • Grounding and citations – Requiring the model to quote or link the sources behind each claim so users can verify them.
  • Better prompting – Explicitly instructing the model to say “I don’t know” when unsure, and asking for sources.
  • Human review – Keeping a person in the loop for high-stakes uses such as legal, medical and financial content.

Hallucination is the main reason AI-generated content should be fact-checked before publication, and reducing it remains one of the central open research problems in generative AI.

Avatar photo

Fritz

Our team has been at the forefront of Artificial Intelligence and Machine Learning research for more than 15 years and we're using our collective intelligence to help others learn, understand and grow using these new technologies in ethical and sustainable ways.