Hallucination: definition, causes and control
Definition
A hallucination is an answer from a language model that sounds fluent and convincing but is factually wrong or entirely made up. The model fills gaps with plausible wording instead of admitting a missing fact. For an AI assistant this is the single most important quality factor, because a confident wrong answer does more damage than an honest I do not know.
Why the Term Matters
An assistant that queries your knowledge brain is only used if people trust its answers. Hallucinations destroy that trust faster than any slow response. Who understands the term builds the right safeguards: answers from approved sources, visible origin and a clear handover to a human as soon as the basis gets thin. That keeps the personal assistant dependable rather than merely eloquent.
When It Is Relevant
- The assistant gives information that someone bases a decision on.
- Answers concern numbers, deadlines, contracts or policies.
- The company brain is also used by people who cannot check the topic themselves.
- An agent derives an action from an answer.
When Not
For purely creative tasks such as a first draft or an idea list, a freely worded answer is wanted, and the term hallucination applies less there. For questions the user immediately verifies, the risk is small too. It only becomes critical when nobody checks any more.
Industries That Use It
- Tax and legal advice: every number and clause counts, an assistant must cite rather than guess.
- Medical technology and pharma: information on approval and use must not be invented.
- Banks and insurers: product and rule information must rest on checked sources.
- Education and training: a learning assistant that invents facts undermines its whole purpose.
Typical Use Cases
- A company brain shows the source for every answer so users can verify the claim.
- The personal assistant states openly when a fact is not in the approved knowledge.
- A check step compares the answer with the source text before it is shown.
- An agent stops and asks back instead of acting on an uncertain assumption.
- Critical information carries a confidence note.
Decision Guide
| Signal | Hallucination is a risk | Alternative |
|---|---|---|
| Answer supports a decision | source binding and check step needed | avoid a free model without grounding |
| User does not verify themselves | force visible origin | a plain chat window is not enough |
| Creative first draft | low risk | free wording is wanted |
| Agent triggers actions | set hard stop rules | do not let it act without approval |
Example
A team binds its assistant strictly to the approved company brain and has every answer carry the source passage. If the assistant finds no supporting passage, it does not answer with a guess but points to the responsible person. This clearly lowers the risk of invented information. The example describes the method, not an asserted error rate.
Service Connection
Citable answers come from a cleanly connected company brain and a personal AI assistant with clear limits, and for actions from controlled AI agents.
Sources
This framing describes common practice for handling hallucinations. How strong the risk is in a given case depends on the model, the connection to sources and the limits set, and must be checked in context.
