What is an AI hallucination?
When a model states something false with total confidence. Why it happens, and how to protect yourself from it.
The short version
- → A hallucination is a plausible-sounding answer that is simply wrong.
- → It happens because models predict likely text, not verified facts.
- → You reduce it with sources, tools, and by checking anything that matters.
A hallucination is when an AI model produces information that sounds right but is not true: a fake citation, a wrong date, an invented function that does not exist. The unsettling part is the confidence. The model does not hedge, it just states it.
It is the most important failure mode to understand, because it is built into how these models work rather than a bug that will be patched away.
Why it happens
A language model is trained to produce the most likely next piece of text, not the most truthful. When it does not have a real answer, it does not stop. It generates the most plausible-looking one, which can be completely fabricated.
It has no sense of not knowing
The model does not track a boundary between what it knows and what it is guessing. Everything comes from the same next-token machinery, so a wild guess and a solid fact are generated the same way and sound equally sure.
Where it shows up most
Specific facts the model was not reliably trained on: exact quotes, citations, statistics, dates, obscure names, and anything after its knowledge cutoff. Open-ended reasoning is safer than precise recall.
How to reduce it
Give the model the source material to work from, use tools like web search for current facts, ask it to show its reasoning, and verify anything factual or high-stakes yourself. Retrieval-augmented generation exists largely to attack this problem.
An analogy
It is like a confident student who never learned to say "I do not know." Asked something they missed, they give a fluent, sensible-sounding answer rather than admit the gap. Often close, sometimes completely made up, always delivered with a straight face.
Go deeper
Are LLMs stochastic parrots?
Where the phrase came from, what it gets right about next-token prediction, what researchers have since found inside the models, and what to do with the answer. 15 min read →
Where Berges AI fits
Berges AI cannot make hallucination disappear, because no product can. What it does is tell the model, on every request, to say so when it is unsure, and the WebSearch pipeline can pull in real sources when a question needs current facts rather than the model's memory.
Related concepts
What is an LLM?
The kind of AI behind chat assistants, explained without the jargon: what it is, how it produces text, and where it falls short. Read →
What is RAG?
Letting a model look things up. Instead of answering from frozen memory, it fetches the relevant text first, then answers from that. Read →
What is a context window?
The model's working memory: how much text it can hold in mind at once, and why it seems to forget once you go past it. Read →
Things people ask
Can hallucinations be fixed completely?
No. They can be reduced with sources, tools, and better training, but they are a consequence of how the technology works. Treat any factual claim as something to verify.
Why does the model sound so sure when it is wrong?
Because it optimizes for plausible text, not truth. Confidence is a property of the writing style, not a signal that the answer is correct.
Which tasks are safest from hallucination?
Rewriting, summarizing text you provide, brainstorming, and reasoning over material in front of the model. Precise recall of facts, citations, and numbers is where to be most careful.