On this page
Definition
A hallucination is an LLM output that is fluent and plausible but factually incorrect or unsupported by any reliable source.
Simple explanation
LLMs generate the most likely-sounding next word, not the most true one. When they lack the right information, they still generate — and the result can be wrong while sounding perfectly authoritative.
Why it matters
In consumer chat, a hallucination is annoying. In legal, medical, financial or enterprise contexts, it can be dangerous. Managing hallucinations is central to shipping trustworthy AI products.
How it works
- 1Missing infoThe model has no reliable data on the query.
- 2Predict anywayIt generates the most probable-looking answer.
- 3Sound confidentThe output looks polished, hiding the uncertainty.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- Made-up citationsLLMs sometimes invent plausible-looking academic references.
- Wrong dates and namesEspecially when asked about recent or niche events.
Advantages
- Understanding hallucinations helps you design better systems around them.
Limitations
- Cannot be fully eliminated; only reduced.
- Users often can't tell a hallucination from a fact.
Common misunderstandings
- ClaimHallucinations are bugs.RealityThey are a natural consequence of how LLMs generate text.
- ClaimBigger models don't hallucinate.RealityThey hallucinate less on average, but still frequently on niche or recent facts.
Frequently asked questions
How do I reduce hallucinations?
Ground answers with retrieval, instruct the model to say 'I don't know', and cite sources.
Are hallucinations lies?
No — the model has no intent. It is generating what looks correct.
The Tool Money Lab perspective
Hallucinations are the single biggest reason we recommend keeping humans in the loop for any AI output that affects money, health, safety or legal outcomes. Systems can reduce them; systems cannot yet be trusted to eliminate them.
Conclusion
Hallucination is the defining reliability problem of LLMs. Ground your systems, cite your sources, and never present raw model output as verified fact.