Why do LLMs hallucinate, and how can you reduce it?

Question

Accepted Answer

An LLM **hallucinates** when it produces text that sounds confident and plausible but is factually wrong or invented. To understand why, you first have to understand what the model actually does.

## How an LLM works (briefly)

At a high level, an LLM is a **next-token predictor**. Given the text so far, it predicts the most likely next token (word/piece) based on **statistical patterns learned during training** — not by looking facts up in a database.

```text
input: "The capital of Australia is"
model: P(next token) → "Canberra" 0.71, "Sydney" 0.18, ...
→ samples a token, appends it, repeats
```

## Why hallucinations happen

- **It generates, it doesn't retrieve.** The model produces *plausible* text, not *verified* facts. Fluency and truth are independent.
- **No built-in truth check.** Nothing inside the model compares output against reality.
- **Gaps get filled confidently.** When training data is thin, ambiguous, or missing (rare APIs, recent events, obscure people), the model still emits the statistically likely continuation — which can be wrong.
- **Confidence is not correctness.** A fabricated citation looks exactly as fluent as a real one.

## How to reduce it

- **Ground with RAG / sources.** Retrieve relevant documents and put them in the prompt, so the model answers *from* real text instead of memory.
- **Ask for citations** and check them — fabricated references reveal hallucination.
- **Lower the temperature** for factual tasks, so the model picks high-probability (safer) tokens.
- **Permit "I don't know."** Explicitly instruct the model to say so when unsure, removing pressure to invent.
- **Verify with tools.** Run the code, query a calculator/search/database, or call an API rather than trusting the model's recall.
- **Narrow the scope.** Specific, well-bounded prompts hallucinate less than open-ended ones.

## Why it matters

LLMs are increasingly embedded in products that real users trust. Because the model **generates plausible text rather than retrieving facts**, hallucination is an inherent behavior, not an occasional bug — so engineers must design around it. Knowing *why* it happens (next-token prediction with no truth check) tells you *how* to mitigate it: grounding, citations, lower temperature, permission to abstain, and external verification. Treating LLM output as a draft to be checked — not an oracle — is the difference between a reliable feature and a confident liability.