Curiosity · Concept

Hallucination

LLM hallucination is the phenomenon of a model producing fluent, confident, but false output — invented citations, wrong dates, nonexistent APIs, fabricated quotes. It's rooted in how models are trained: to predict plausible continuations, not to verify facts. Mitigations include grounding via RAG, tool use, decoding controls, self-consistency, and training techniques that penalize overconfident guesses.

Quick reference

Proficiency
Beginner
Also known as
LLM hallucination, confabulation, fabrication
Prerequisites
LLM basics

Frequently asked questions

What is an LLM hallucination?

It is when an LLM produces content that is wrong, fabricated, or unsupported — stated with the same confidence and fluency as correct content. Examples: inventing a court case, citing a nonexistent API, misquoting a real paper, or getting a simple date wrong.

Why do LLMs hallucinate?

They're trained to predict plausible next tokens, not to verify facts. When the model's parametric memory is thin (long-tail facts, post-cutoff events, ambiguous entities), it pattern-matches something that looks right. Sampling with temperature also introduces randomness that can tip into wrong outputs.

How can I reduce hallucinations in production?

Ground the model: use RAG to supply real sources, use tool calls for facts you can look up, require the model to cite retrieved passages, lower temperature, use structured outputs, and have the model explicitly say 'I don't know' when confidence is low. LLM-as-judge or groundedness checks post-hoc are also common.

Are reasoning models less prone to hallucinations?

Mixed. Reasoning models (o1, o3, R1, Claude thinking) can catch their own factual errors during deliberation, and tend to do better on math and logic. But they still hallucinate facts they don't know, and longer reasoning can sometimes let errors propagate. Grounding and tool use still matter.

Sources

  1. Huang et al. — A Survey on Hallucination in LLMs — accessed 2026-04-20
  2. Lin et al. — TruthfulQA — accessed 2026-04-20
  3. OpenAI — SimpleQA — accessed 2026-04-20
  4. Google — FACTS Grounding benchmark — accessed 2026-04-20

Related

  • LLM-as-Judge — LLM-as-judge is an evaluation pattern where a language model grades or ranks another model's outputs, serving as a scalable — if imperfect — substitute for human evaluation.
  • Prompt Injection — Prompt injection is an attack where adversarial instructions hidden in untrusted input — a document, webpage, email, or tool output — override the developer's intended prompt and cause the LLM to behave maliciously.
  • Retrieval-Augmented Generation (RAG) — Retrieval-Augmented Generation (RAG) is a pattern where an LLM is grounded on retrieved passages at query time — fewer hallucinations, up-to-date answers, no retraining required.
  • All Concepts — Browse the complete concepts library
  • VSET — Programmes offered — B.Tech programs (CSE, AI&ML, AI&DS, Electronics, IIoT)