Hallucination
Also: confabulation, fabrication
What is Hallucination?
A hallucination is output from a language model that is fluent and confident but factually wrong or unsupported by any source, such as an invented policy, figure, citation or product detail.
What Hallucination means
A hallucination happens when a model generates text that is plausible in form but false in content. Because the model is predicting likely tokens rather than looking facts up, it will complete a sentence about your refund policy or a drug interaction whether or not it has the right information. The result reads exactly like a correct answer, which is what makes it dangerous.
The main causes are missing context (the model was never shown the fact), ambiguous questions, pressure to answer rather than decline, and long generations that drift from the source. Mitigations are architectural: supply the facts through grounding and retrieval, constrain answers with structured output, require citations, let the model say "not found", and measure groundedness in evals.
Hallucination is not a bug you patch once, and it is not the same as a model being out of date. It is a property of how generation works, so the goal in production is to detect and contain it: keep the model inside your data, verify claims that matter, and route uncertain cases to a person.
Who it really matters to
- Founder / CEO: one invented commitment to a customer can cost more than the feature saves, so containment is a launch condition.
- Support manager: an assistant that fabricates policy creates tickets and refunds it was meant to prevent.
- Compliance officer: fabricated regulatory or medical statements are a liability, and "the AI said it" is not a defence.
- CTO / Head of Engineering: the fix is grounding, evals and refusal paths, which shape the architecture from day one.
Why it exists
The term exists because early deployments discovered that fluency and accuracy are independent: a model can be persuasive and wrong at the same time. Naming the failure made it something teams test for rather than something they discover from customers. The trade-off in containing it is helpfulness versus caution. Tight grounding and refusal rules reduce fabrication but also make the system decline more often, and users notice both. The right balance depends on the cost of a wrong answer, which is why a lending assistant and a marketing copy tool are tuned differently.
Where it is applied
- A customer-service agent that answers only from the help centre and shows the article it used, or escalates.
- A clinical assistant that summarises discharge notes with every statement traceable to the source document.
- A banking chatbot that quotes fees and terms exclusively from a versioned product table, never from memory.
- A student-support agent that declines fee and admission questions it cannot ground in the institution's rules.
- A logistics customer-update system that fills a fixed template from tracking data rather than composing free text.
Is Hallucination a skill?
ConceptA concept every stakeholder should understand, and a failure mode Eazyware designs against in every LLM application: grounding, citations, refusal paths and groundedness scores are standard, not add-ons.
Eazyware service that covers it: LLM Application Development. Starting prices are on the pricing page.
Frequently asked questions
Can hallucinations be eliminated completely?
No. They can be reduced sharply with grounding, structured outputs and refusal paths, and detected with groundedness checks and human review. The realistic goal is a measured, acceptable rate for your use case with a safe path for the remainder.
Does a bigger or newer model fix hallucination?
Better models hallucinate less on general knowledge, but they still cannot know your policies, prices or patient data unless you supply them. For business use, retrieval and grounding matter more than model size.