Why does AI make things up and how do I spot hallucinations?

It makes things up because it generates plausible text rather than retrieving facts — and a fake citation is extremely plausible text. It has the right shape: author names, a journal, a year, a title in the right style. The model is optimising for 'what would a citation look like here', and it has no separate mechanism that checks whether the thing exists. Why the confidence is misleading: fluency and certainty are properties of writing style, which the model reproduces perfectly. There's no internal signal saying 'I'm guessing now' that surfaces as hedging. So confidence carries none of the information it carries when a knowledgeable human says it. Where hallucinations concentrate — treat these as high-risk zones: 1. **Specific citations, references, quotes and page numbers.** The highest-risk category by far. Always verify; a startling proportion are fabricated or mis-attributed. 2. **Precise numbers, dates and statistics.** Especially anything recent. 3. **Niche or obscure topics** where training data is thin. It rarely says 'I don't know'; it interpolates. 4. **Recent events** past the training cutoff, unless the tool is actually searching the web. 5. **Legal, medical and regulatory specifics** — real case names, dosages, statute numbers. Lawyers have been sanctioned for filing AI-invented case law; this is a documented, repeated failure. 6. **APIs, library functions and package names.** It will confidently invent a method that ought to exist. Easy to catch — the code errors — but wastes time, and invented package names have been targeted by attackers who register them. How to spot them in practice: - **Verify anything you'd be embarrassed to be wrong about.** Click through to the source. If the source can't be found in thirty seconds, treat it as fabricated. - **Ask the same question in a fresh conversation.** Genuine knowledge is usually stable; hallucinations often shift between runs. - **Ask 'how confident are you and what would change your answer?'** Imperfect, but it does surface some uncertainty. - **Watch for suspicious specificity.** A precise figure with no source, on an obscure topic, is a warning sign rather than a mark of authority. What reduces the risk structurally: use tools that cite real retrieved sources (search-grounded or document-grounded), give it the source material yourself rather than relying on recall, and use it for tasks where you can check the output — code that either runs or doesn't, drafts you'll edit, explanations of things you can verify. The reliable framing: it's a fast, well-read, occasionally-confabulating assistant. Excellent for drafts, structure, explanation and brainstorming. Not a citation source, and not a final authority on anything that matters.