Why It Makes Things Up
Hallucination Is the System Working as Designed
When a model invents a quotation, a statistic, a court case or a book that does not exist, that is called a hallucination. It is tempting to read it as the model lying, but nothing has gone wrong internally. The model was asked to produce the most plausible continuation and it did. A fake citation looks exactly like a real one — author, year, plausible title — because it was generated from the pattern of real citations. There is no separate fact-checking stage inside the model that could have caught it. Truth is not a variable the system tracks; plausibility is.
- Made-up references, dates and quotes are the classic failure, and they look completely normal
- The model cannot flag them because it has no independent record of what is true
- Risk is highest for specific, verifiable, rarely-written-about details
- Risk is lowest for restructuring or explaining material you supplied yourself
Confidence Is a Writing Style, Not a Signal
Human writers usually hedge when they are unsure. That correlation makes us read confident prose as informed prose, and models write in confident prose almost always, because most of the text they learned from was written by people who were confident. So the usual social cue you rely on — does this person sound sure? — has been decoupled from whether the content is right. Some systems now express uncertainty more, but you should never treat tone as evidence. A fabricated answer and a correct answer arrive with identical body language.
- Fluency and confidence tell you nothing about accuracy — they are style, not evidence
- Asking "are you sure?" often produces a confident correction rather than a genuine confidence estimate
- It may also cave and agree with you when you were wrong, because agreement is a common pattern too
- Trust verification, not tone
Where the Danger Actually Sits for You
The realistic risk is not a wild, obvious lie. It is a small, confident, wrong detail sitting inside four correct paragraphs, in a subject you do not know well enough to notice. That is exactly the situation you are in when you use AI for homework in a topic you are still learning. The defence is not paranoia, it is proportion: verify anything that carries a name, a number, a date, a source or a formula, and relax about phrasing, structure and explanation. Ask yourself what happens if this specific claim is wrong. If the answer is "I lose marks" or "I mislead someone", check it.
- Highest risk: named sources, statistics, dates, legal or medical specifics, anything niche
- Lower risk: rewording your own work, generating practice questions, explaining a concept several ways
- Ask the model to show its reasoning — errors are easier to spot in steps than in conclusions
- If two independent sources cannot confirm it, do not put it in your work
Try It Yourself
Reading about hallucination is not the same as catching one. The quickest way to see it is to ask about something you happen to know better than the model does.
Pick a subject you know inside out — your team's last season, a game you play properly, a band, the streets near you. Ask any AI chat tool a detailed question about it, one whose answer has to contain names, dates or numbers. Then read the reply as the expert you are and mark it.
Tell me about [your specific topic]. Be detailed and specific — include names, dates and numbers.
- At least one detail came back wrong, and you knew it was wrong without looking anything up
- The wrong part was written in exactly the same confident tone as the correct parts
- The errors clustered in the most specific claims — the names, numbers and dates
Turn sampling up and down and watch the output change.
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.