Why it happens: the model produces likely-sounding text, and a plausible fake can be ‘likely’.
Warning signs: very specific numbers without sources, references you cannot find, and answers that change when you ask again.
Check routine: (1) Is it important? (2) Find the claim in a reliable source such as an official site, reputable news or a textbook. (3) Check any references actually exist. (4) Ask the AI for its uncertainty. (5) If it matters — health, money, law — ask a qualified person.
Some tools can search the web and show links; still open the links and read them.
Worked example 1
A chatbot says ‘According to a 2019 Oxford study, 83% of cats prefer jazz.’ What do you do?
Very specific statistic plus a vague source — warning signs.
Search for the study on Oxford's site or a research database.
If you cannot find it, treat it as invented.
Answer: Do not use it unless you can find the actual study.
Try it online · print it · check it
Practice questions
Intermediate · ages 11+
AI errors and hallucinations
NameDate
Answer each scenario: trust, check or ask an expert.
1.An AI suggests a synonym for ‘happy’ in a birthday card.
2.An AI gives a quote from a famous person for your essay.
3.An AI tells you how much of a medicine to give a child.
4.An AI lists three books as sources for your report.
5.An AI says what your tenancy law rights are for a dispute.
Intermediate · answersAnswer key
AI errors and hallucinations — answer key
Answer each scenario: trust, check or ask an expert.
1.An AI suggests a synonym for ‘happy’ in a birthday card.trust
2.An AI gives a quote from a famous person for your essay.check
3.An AI tells you how much of a medicine to give a child.ask an expert
4.An AI lists three books as sources for your report.check
5.An AI says what your tenancy law rights are for a dispute.ask an expert
Tips
Common mistakes
Checking only by asking the same AI ‘are you sure?’.
Trusting a confident tone. AI tools sound equally sure when they are wrong.
AI hallucination explained
Why AI tools make things up, and what to do about it
What an AI hallucination is
An AI hallucination is an answer from an AI system that sounds confident and fluent but is wrong or made up. The term is used most for chatbots and writing assistants built on large language models, but image and speech tools can produce similar errors, such as a picture with the wrong number of fingers or a transcript that adds words nobody said.
The word is borrowed loosely from psychology. The AI is not seeing things or lying in a human sense; it has no intention at all. It is producing text that fits the pattern of a good answer without any built-in check that the content is true.
Why language models produce them
A language model is trained on a very large amount of text to predict which words are likely to come next. When you ask a question, it builds an answer one small piece at a time, each time choosing a likely continuation. Most of the time the most likely continuation is also correct, because correct facts appear often in its training text.
Problems start when the model has seen little or conflicting information, when the question is very specific, or when it is asked for something that does not exist. A plausible-sounding author, page number or percentage is still ‘likely text’, so the model can produce it as smoothly as a real one. It does not have a separate store of checked facts to consult unless the tool has been connected to a search engine or document library, and even then it can misread or misquote what it retrieves.
Other causes include training text that was itself wrong, information that has changed since the model was trained, and prompts that assume something false. Ask ‘Why did the council ban bicycles in 2021?’ and a model may invent a reason rather than tell you the ban never happened.
Common types of hallucination
Made-up facts: a wrong date, a person in the wrong job, an event that did not happen, or a law that does not exist.
Fake citations: references to books, articles, court cases or web pages with realistic titles and authors that cannot be found anywhere.
Wrong numbers: arithmetic slips, totals that do not add up, unit mix-ups, or statistics that look precise but have no source.
Invented quotes: words attributed to a real person who never said them, often in their recognisable style.
Confident summaries of documents that include points the document never made, or miss the main point.
Code or instructions that look right but call functions, settings or menu options that do not exist.
What can go wrong in real life
The consequences depend on where the answer ends up. A student who copies an invented quotation into an essay can lose marks or face questions about academic honesty. A professional who submits a report with fake references can damage trust with a client or an employer, and there have been publicly reported cases of documents prepared for courts containing non-existent case law.
Wrong numbers can lead to poor money decisions, such as a budget built on a miscalculated total. Health or safety information that is slightly wrong can be dangerous. A made-up claim about a real person can be unfair or defamatory if it is shared. None of these need bad intentions; they happen when a fluent answer is trusted without checking.
How to check an AI answer, step by step
Decide how much it matters. A suggested synonym needs no checking; a fact for homework, work, money, health or law does.
Pick out the specific claims: names, dates, numbers, quotes and references. These are where errors hide.
Find each important claim in an independent, reliable source, such as an official website, a reputable news organisation, a textbook or the original document. Do not ask the same AI to confirm itself.
Open every reference. Check that the book, article or page exists and actually says what the AI claims.
Redo any calculation yourself or with a calculator, and check the units.
Ask the tool to show its sources or say how confident it is, then treat the reply as a lead to check, not as proof.
If the answer could affect someone's health, money, safety or legal position, take it to a qualified person.
Warning signs to look out for
Very precise figures (such as ‘73.4%’) with no source.
References you cannot find with a simple search, or links that go nowhere.
An answer that changes when you ask the same question again.
Quotes that sound exactly like what you hoped someone would say.
Agreement with a false assumption built into your question.
When not to rely on AI
Do not rely on an AI answer on its own for medical doses, symptoms or treatment; legal rights or deadlines; tax, benefits or financial decisions; safety instructions; or anything you will submit as your own work with references. Use AI for brainstorming, explaining an idea in simpler words, drafting something you will check, or practising questions, and keep the final judgement with a person who can verify it.
Reducing hallucinations is an active area of work for AI developers, and tools that search the web and show links can help. Checking still matters, because even a linked source can be misread.
Questions
Frequently asked questions
It is a loose borrowing of the everyday word: the AI produces something that is not really there. The AI has no intention or awareness, so some people prefer terms such as ‘fabrication’ or simply ‘error’.
Not completely. Connecting a tool to search or trusted documents and asking it to cite sources can reduce them, but any language model can still produce a confident wrong answer.
Search for the exact title and author in a library catalogue, a search engine or the publisher's site. If you cannot find it, or it exists but says something different, do not use it.
Follow your school's rules first. Used for explanations or practice and checked against reliable sources, it can help. Copying unchecked answers, quotes or references is risky and may break academic rules.