What is AI hallucination, and why do LLMs make things up?

A hallucination is not a bug in one answer; it is the model doing exactly what it was trained to do in a situation where that goes wrong. The model writes the most plausible continuation. When it lacks the facts, plausible and true come apart.

Why do LLMs hallucinate?

LLMs hallucinate because they generate likely text rather than retrieve verified facts, and nothing in the basic mechanism checks the output against reality. A large language model stores knowledge as statistical patterns. For facts it saw many times, the pattern is strong and the answer is usually right. For rare facts, recent events, or specifics like page numbers and exact figures, the pattern is weak, but the model still produces fluent text in the right shape.

Training incentives make this worse. A 2025 paper from OpenAI researchers, "Why Language Models Hallucinate", argues that most benchmarks grade answers as simply right or wrong, so a model that guesses scores better than one that says "I don't know". Like a student who guesses on every multiple-choice question, models learn that confident guessing pays. The authors propose scoring that penalises confident errors more than abstentions.

What are the common types of hallucination?

It helps to separate two kinds. Factuality errors contradict the world. Faithfulness errors contradict the documents you gave the model. Retrieval helps with the first; careful prompting and checking help with the second.

How can you reduce hallucinations?

You cannot eliminate hallucinations, but you can make them rare and catchable. The most effective techniques supply the facts, give the model permission to decline, and verify the output.

1. 1
   Ground the answer in sources

   Use RAG or tool calling so the model reads the facts instead of recalling them.
2. 2
   Allow "I don't know"

   Say explicitly that it is fine to answer "not in the provided documents". Models follow this much more often when told.
3. 3
   Ask for quotes first

   For long documents, have the model extract the relevant quotes, then answer only from those quotes.
4. 4
   Require citations

   Ask the model to cite a passage for each claim, then check automatically that each cited passage exists and supports it.
5. 5
   Lower the temperature

   For factual tasks a low temperature reduces odd, low-probability choices.
6. 6
   Verify with code

   Validate IDs, dates, URLs, prices and function names against real systems before showing them.
7. 7
   Measure it

   Build an eval set of questions with known answers, including some that should be declined, and track the error rate.

For high-stakes answers, add a second check: a separate model call, or a guardrail, that compares each claim against the source documents and flags anything unsupported.

Do newer models hallucinate less?

Generally yes, on the tasks they are tested on, but not to zero, and it varies by task. Larger and better-trained models know more facts and are more often trained to express uncertainty. Reasoning models can catch some of their own mistakes. They can also produce longer chains of confident reasoning, so a wrong premise can travel further. Measure on your own use case rather than relying on a general claim.

• ✓Do: Treat model output about facts, numbers and citations as unverified until checked
• ✓Do: Give the model the documents it needs, and tell it to use only those
• ✓Do: Keep a set of questions it should refuse, and test that it does
• ✕Avoid: Ask a model for sources without checking that the sources exist
• ✕Avoid: Assume a confident tone means a correct answer
• ✕Avoid: Rely on the model to know events after its training cutoff