Why AI makes things up (hallucinations)
- verification
- foundations
A hallucination is a fluent, confident piece of AI output that simply is not true. Models generate what plausibly comes next — not what is accurate — so the most convincing sentences can be the least reliable. Knowing where fabrications cluster is most of the defence.
It is not lying — it is completing
LLMs have no internal fact-checker. When the pattern calls for a citation, a number or a quote, the model produces the shape of one: authors, years, statistics, page numbers. The output is optimised to look like a correct answer, because correct-looking text is what its training rewarded.
The rule of thumb follows from the mechanism: plausibility is generated; truth is not.
Where fabrications cluster
- Numbers and statistics — precise figures feel like evidence; an invented one is indistinguishable by style.
- Dates and “the latest” anything.
- Names, titles and institutions — Dr. So-and-so at the Institute of Something.
- Quotes — the more colourful, the more suspect.
- Citations, links and DOIs — reference-shaped text is easy to generate and impossible to click.
- Niche or very recent facts — thin training signal, strong pressure to answer anyway.
A bad example
Ask: “Give me a peer-reviewed study showing that open-plan offices reduce creativity, with a DOI.”
You will get a full citation — authors, journal, year, DOI. It may not exist. The model was completing a citation pattern; nobody checked the library.
A better example
Ask: “Quote only from the text I paste below. If the answer is not in the text, say ‘not in the text’.”
And separately: when you need an outside fact, look it up in a source you can open — a search, a database, the original document — instead of asking for recall.
Why it works
Two habits cover most of the risk:
- Treat specifics as claims, not facts. Numbers, names, dates and quotes get checked — that is exactly where failures live.
- Set the verification budget by the consequence. A draft email: a glance. A number in a report, or anything about money, health or law: primary source or a human.
The goal is not blanket suspicion — it is calibration: fast trust for low stakes, real verification where it counts.
One honest note on scale: this is not a bug you can wait out. Catching fabrications is a skill — the checklist in the research track is the same one every careful reader uses, and it works on any model, current or future. The tool changes; the verification habit travels with you.
Practice
Spot the fabrication
An AI assistant answered a question about workplace research. You cannot open any sources right now — you have only the text below.
Select every part of the answer that should not be trusted as-is.
The AI answer
Hybrid work has become common in many companies since 2020. A 2023 study by the Institute for Workplace Research found that hybrid teams were 41% more productive than fully in-office teams, and the study's lead author, Dr. Helena Marsh, called it “the largest productivity gain since the typewriter.” Hybrid schedules usually combine office days with remote days.
Hint
Specifics are the suspects: names, numbers, dates, quotes, citations.
Generic statements with no measurable detail are rarely the problem.
Why this is the answer
Hallucinations cluster around specifics: names, dates, numbers and quotes, plus any citation you cannot open. The broad statements around them are usually fine — which is exactly what makes the fabricated details easy to miss. Treat specificity as a reason to verify, not a reason to believe.
Transfer
- A deck with five statistics: every number gets a source.
- “A quick summary of the regulations”: find the regulation.
- An email draft: read it and send — low stakes.
- Anything with a DOI or a link: click it, or delete it.
Next
Next: Models vs products — why the same engine behaves differently in different apps. The hallucination glossary entry is the one-line version.