Glossary · Industry

AI Hallucination

An AI hallucination is when an LLM generates plausible but false information. Definition, why it happens, and how to mitigate it in production.

By Kadin Nestler · May 28, 2026 · Updated May 28, 2026
On this page

Why hallucinations happen

LLMs are trained to predict the most likely next token, not to be truthful. When the model has no relevant training data for a query — or has contradictory data — it generates the most plausible-sounding continuation, which may be invented. Hallucinations are most common on niche facts, recent events outside the training cutoff, numerical reasoning, and any domain where the model is asked to be specific about something it half-knows.

Common hallucination patterns

  • Citation fabrication — invented book titles, paper authors, court cases, URLs.
  • Specification drift — model confidently states a product feature, version number, or API endpoint that does not exist.
  • Numerical errors — confident but wrong calculations, statistics, dates.
  • Persona drift — model claims capabilities or facts about itself that are not accurate.
  • Faithful but wrong — output follows the prompt structure but the substance is fabricated.

Mitigation techniques

  • RAG — ground answers in retrieved documents instead of model memory.
  • Citation requirement — make the model cite the source for every claim.
  • Tool use — for facts that require precision (math, dates, queries), call a tool instead of generating.
  • Lower temperature — sampling closer to the most-likely token reduces creative invention.
  • Self-consistency — generate multiple answers and only return claims that appear across runs.
  • Verification layer — a second model or rule-based check validates outputs before they ship.

Why SMBs need to care

A customer-facing chatbot that fabricates a refund policy creates legal exposure. An AI receptionist that hallucinates business hours costs you bookings. A legal AI that invents case citations gets the lawyer sanctioned — this has happened repeatedly in 2023-2025 (Mata v Avianca, Schwartz v Avianca). Any production AI workflow needs an explicit answer to "what happens if the model hallucinates?"

What it means for your business

Hallucinations are not a vendor problem you can solve by picking a smarter model. They are an architecture problem. The right answer is grounding the AI in documents you control and refusing to answer when grounding is thin.

  • AI Grounding — Grounding is the practice of tying AI outputs to verified source material. Definition, techniques, and why it is the primary defense against hallucination.
  • Retrieval-Augmented Generation (RAG) — RAG is the technique of fetching documents from a database and feeding them to an LLM before it answers. Definition, architecture, and SMB use cases.
  • AI Guardrails — AI guardrails are runtime rules and filters that constrain LLM behavior. Definition, types, and how SMBs should use them in production.
  • Large Language Model (LLM) — A Large Language Model is a transformer-based neural network trained on trillions of tokens to predict the next token. Definition, key models, and business use.
  • AI Evaluation — AI evaluation is how you measure whether an AI system actually works. Definition, methods, and why evals are the bottleneck in production AI.
← Back to the full glossary