Understanding
hallucinations
AI makes mistakes. Why the mistakes happen, when they happen most, and what reduces the rate of mistakes.
Scroll, click, or use the arrow keys to move through.Scroll, or use the arrows and dots to move through.
A confident answer assembled the same way as a correct one.
An LLM generates the most probable next token, over and over, until the answer is done. A hallucination is not a malfunction: it is the same process that most often works correctly, producing content that is not true.
There is no flag inside the machine that says “this one is made up.”
Accuracy and hallucination are the same process.
Where accuracy comes from
The model predicts the next token from everything it has seen. Where the training ground is deep, the most probable continuation is usually the correct one.
Where hallucination comes from
The same prediction on thinner ground. Where the facts are missing, the most probable continuation is what the truth usually sounds like: a plausible date, a real-sounding citation, a function that should exist.
The size
effect
The longer a single step runs, the more likely it drifts.
The probability of hallucination correlates most strongly with the token count of a step. Every token is conditioned on the tokens before it, so one early drift compounds through everything generated after it.
Break the work into steps for the highest accuracy.
The error rate correlates most strongly with the number of tokens, so one long question or request is more likely to have an error.
The same work in small steps means each individual step is more likely to be accurate. Small steps are also easier to check, resulting in greater accuracy.
Reasoning reduces errors. It does not eliminate them.
A reasoning model drafts and checks internally before it answers, reducing the rate of errors delivered. The reasoning that checks for errors uses the same probabilistic methods as the answer, so the floor never reaches zero. And a reasoning model that is wrong hands you a plausible chain of justification along with the wrong answer.
The wrong answers get more persuasive.
Right and wrong arrive in the same tone of voice.
No hesitation in the wrong ones, no extra confidence in the right ones. Nothing inside the answer tells you which you are holding. The check has to come from outside the answer.
Design around it.
Keep steps short
The size effect is the reason. Small steps drift less, and the seams between them are where errors get caught.
Pull data close
An answer grounded in data at hand has less gap to fill. The thinner the ground, the more the model invents.
Check with a fresh call
Make verification its own step with a fresh call. Models are very unlikely to error the same way twice in a row, so loops that re-run or panels that judge the output independently are strong corrective measures.
Hallucination is not a bug to wait out. It is a property you design around.
The rate never reaches zero, but it becomes something you can engineer against instead of something you hope away. Planning for errors is the best way to make it go away.