Hallucination
A confident, fluent LLM output that is factually wrong or unsupported by any source, produced because the model predicts plausible text rather than checking truth.
Definition
A language model generates the most plausible continuation of the text it has been given. Plausible and true are usually the same thing, but not always. When the model lacks the relevant fact, or the fact is rare, it still produces a fluent answer, and that answer can be invented: a citation that does not exist, a product feature you never shipped, a legal clause that was never written. Some researchers prefer the term "confabulation".
Hallucinations are more likely with vague questions, with topics the model saw little of in training, with long outputs and with high temperature. They are less likely when the model is given the source material (RAG), told it may say "I don't know", and asked to quote rather than recall.
Example: asked for "the cancellation terms in our enterprise plan", an ungrounded assistant drafts a convincing 30-day clause. The real contract says 90 days. A grounded assistant quotes the actual clause or reports that it could not find one.
Why it matters for your business
- Wrong answers delivered confidently damage trust faster than an honest "not found".
- In regulated domains (finance, health, legal) a hallucinated figure or clause is a liability, not an inconvenience.
- The remedy is architectural: grounding, citations, scope limits and evals, not better wording alone.
- Error rates must be measured and monitored in production, because they change as data and models change.
How Teksmobile uses it
Every assistant we build in our AI knowledge base practice is grounded in retrieved sources and returns citations the user can check. Our fine-tuning and evals service sets up the test sets that measure hallucination rates before launch and after every change.
Related terms
Part of the Teksmobile AI Glossary. For longer reads, see our guides.
Frequently asked questions
Will newer models stop hallucinating?
Rates fall with each generation, but the behaviour is inherent to next-token prediction. Design for it: ground answers in sources, constrain scope and measure error rates rather than hoping a model upgrade removes them.
How do we measure hallucination?
With evals: a set of questions with known answers and sources, scored for correctness and for whether each claim is supported by the retrieved text. See evaluating LLM outputs.
Is a hallucination the same as a bug?
No. A bug is deterministic and fixable in code. A hallucination is a statistical failure that you reduce and contain through architecture, data and process.