AI and automation

What is hallucination?

Also known as: AI hallucination, confabulation

Published September 10, 2026 · Reviewed by the Fidiora team

Definition

A hallucination is an AI-generated statement that is fluent and confident but not supported by any real source. In customer support it typically appears as an invented policy, a feature that does not exist, or a fabricated timeline. The danger is that it reads exactly like a correct answer.

Why it matters

A hallucinated support answer is worse than no answer, because the customer acts on it. An invented refund window becomes a commitment you either honour or refuse, and both outcomes cost more than the ticket you avoided. This risk, not accuracy in general, is what makes grounding non-negotiable.

What to know

  • Hallucination rises sharply when the question is outside the retrieved content, which is why refusal paths matter.
  • Fluency is not correlated with correctness, so reviewers reading for tone will miss it entirely.
  • Numbers, dates, and policy terms are the highest-risk fields and deserve deterministic handling.
  • Source citation converts hallucination from invisible to checkable in one step.
  • A sampled human review of AI answers, weekly, catches drift that aggregate metrics hide.

An example

Example: asked about an enterprise SLA that is not documented publicly, an ungrounded assistant replies with a confident ninety-nine point nine percent uptime commitment. Nobody at the company ever agreed to that number, and a customer now has it in writing.

Common mistakes

  • Reviewing AI answers for tone rather than for factual support
  • Letting the model answer pricing and policy questions from general knowledge
  • Removing the I do not know response because it looks like failure
  • Relying on aggregate accuracy metrics with no sampled human review
Where Fidiora fits

Fidiora will not answer what your content does not support. It says it cannot confirm, routes to a human, and records the gap, which keeps invented policies out of customer conversations and turns each miss into a documentation task.

See Pricing
FAQ

Questions

Can AI hallucination be eliminated?
Not completely, but it can be reduced to a manageable level. Grounding in verified sources, an explicit refusal path, deterministic handling of prices and policies, and sampled human review together cut the practical risk to something a support team can live with.
What is the most dangerous kind of support hallucination?
Anything a customer will act on or hold you to: refund windows, pricing, security commitments, and delivery dates. These should be answered from a single authoritative source or not at all.
How do I detect hallucinations in production?
Sample AI answers weekly and check each claim against your documentation, watch for spikes in reopen rate on specific topics, and read every escalation where the customer disputes something the AI said.

Related terms

Grounding Grounding is the practice of restricting an AI system to answer only from a defined set of verified sources, such as you... Retrieval augmented generation Retrieval augmented generation is a technique where a system searches a document collection for passages relevant to a q... Support quality assurance Support quality assurance is the structured review of conversations against a defined standard, scoring accuracy, tone, ... Large language model A large language model is a machine learning system trained on very large text collections to predict likely continuatio...
Get Started

See Fidiora resolve a ticket in 60 seconds.

No credit card, no sales call required. Connect your docs and watch it work.

Start Free Book a 15-min Demo

No credit card · Live in under an hour