AI hallucinations: what they are, why they happen, how to spot one.
The word "hallucination" sounds dramatic. The reality is more mundane and more useful to understand.
6 min read
When an AI produces a confident statement that is wrong, that is a hallucination. Most people first hear the word and assume the AI is broken. It is not broken. It is doing exactly what it was built to do, and that fact is the whole story.
The technical reality
A large language model produces a probable next response given the input. It does not look up facts. It does not check sources. When asked something it does not know, it does not say "I do not know" by default; it says the most plausible-sounding thing it can construct. That construction is sometimes right and sometimes wrong, and from the outside both look identical.
That is the hallucination. There is nothing mystical about it.
Where hallucinations happen most
The hot spots are predictable.
- Specific quotes and citations. AI will produce a plausible-looking quote from a real person, or a plausible-looking citation to a real paper, that does not exist. - Specific numbers and dates. The model will give you a confident date or figure that is plausible but wrong. - Legal and medical specifics. Cases that never existed. Statutes that say something else. Dosages that look right. - Software API details. Function signatures that look real but do not match the library. Configuration options that were deprecated. - Anything at the edge of what was in training data. The further from common knowledge, the higher the risk.
How to spot one
The trick is not to read with the assumption that the AI is right. Read with the assumption that any specific factual claim is suspect until you verify it.
1. Citations get clicked. If the AI cites a paper or case, click the link. If there is no link, ask for one. If the link is dead or goes to something different, that is a tell. 2. Numbers get checked. If the AI says "78 percent of agencies report X," ask where that came from. Probably nowhere. 3. API calls get tested. If the AI writes code that calls a function, look at the actual library to confirm that function exists with that signature. 4. Confidence is not correctness. AI is confident by default. That tells you nothing about whether it is right.
What to do about it
Use AI for what it is good at: drafting, summarizing, restructuring, analyzing content you provide. Verify any specific factual claim before you act on it. Build that verification step into every workflow that touches AI output.
The Prompt to Product book devotes an entire chapter to the verification patterns that work in practice.