Hallucination: Why AI Sounds Sure and Is Wrong
A chat AI can give you a smooth, confident, well-organised answer in which one name, one date or one source is simply made up. It does not hesitate, it does not hedge, and nothing in the tone tells you which part is wrong. This is called hallucination, and it is the single most important limit to understand if you use AI for anything that matters. This tutorial explains in plain words why it happens, which kinds of error to expect and where they hide, why asking "are you sure?" does not fix it, and what actually reduces it. You will also learn a short checking routine that takes minutes, and a way to match the effort to the stakes so you do not check everything or nothing.
Difficulty: ★☆☆☆☆ (no technical background needed)
Tools needed: Any chat AI you already use, and a web browser for checking
Last updated: November 2026
Overview
Picture a very well-read friend who hates saying "I do not know". Ask them anything and you get a fluent, friendly, detailed answer. Most of the time it is right. Sometimes a detail is wrong, and they say it in exactly the same warm, certain voice. If you could not tell the difference by tone, you would have to check the things that matter, and you would check the ones you could least afford to get wrong.
That friend is a decent picture of a chat AI. The reasons are built into how it works, and they are not a temporary bug that the next version will remove. Tools have become more accurate over the years, and many now search the web and show sources, which helps a great deal. But the underlying habit of producing fluent text remains, so the right response is not fear or avoidance. It is a small set of habits.
The word "hallucination" is a little dramatic, and people use it for several different problems. In this tutorial it means: the AI states something as fact that is not supported by its source material, by the real world or by what you asked, in a fluent and confident way.
The tutorial has seven steps. First, what hallucination really is, with the main types. Second, why it happens, in terms you have already met if you read the tokens tutorial. Third, where it hides, so you know which parts of an answer to distrust first. Fourth, why confidence is not evidence, and why "are you sure?" does not help. Fifth, what really reduces it. Sixth, a short routine for catching it. And seventh, how to match your checking to the stakes.
You will finish with a mental checklist you can run in thirty seconds, a prompt that makes answers easier to verify, and a clear idea of which uses of AI are safe to trust lightly and which are not.
Who This Is Useful For
What You Will Learn
What You Need
1. Any chat AI you can use.
2. A topic you know well, so you can spot mistakes. Your own job, a hobby or your home town all work.
3. A browser, for checking.
4. About thirty minutes, including the small experiment.
A free way to check whether a paper or study an AI mentions really exists. Search the exact title and open the result yourself.
Information as of 14 November 2026. How often tools make things up varies by tool, topic and version, and it changes quickly. This tutorial gives no rates or rankings. The principles and habits here apply across tools, and your own tests on topics you know are the best guide to how a given tool behaves for you.
Step 1 — What hallucination really is
A hallucination is a confident statement that is not backed up. In practice it comes in a handful of recognisable types.
| Type | What it looks like |
|---|---|
| Invented fact | A date, number, name or event that never existed |
| Invented source | A paper, book, article, court case or link that is not real, often with a convincing title and author |
| Invented quote | Words put in a real person's mouth that they never said |
| Distorted fact | A real thing, slightly wrong: the right person with the wrong year, the right law with the wrong detail |
| Misread document | A summary that attributes to a document something it does not say, or leaves out a key qualification |
| Invented step or feature | Instructions for a menu option or a setting that does not exist in the tool you use |
| Bad reasoning | A calculation or logic chain that looks tidy and is wrong |
Two points are easy to miss.
It is not only "made up from nothing". The distorted fact and the misread document are often the hardest to catch, because most of the answer is right. A paragraph that is ninety per cent correct and ten per cent wrong is more dangerous than one that is obviously nonsense.
It happens even when the AI has your document. Giving an AI a text to work from greatly reduces invention, but it does not remove it. A summary can still overstate, merge two points or drop a "not" that mattered.
A well-known real example, widely reported in 2023: lawyers in the United States filed a court document that cited several cases supplied by a chat AI. The cases did not exist. The court noticed, and the lawyers were sanctioned. The cases had realistic names, courts and citations, which is exactly why they were not questioned. The lesson is not that AI is useless. It is that fluent and realistic is not the same as real.
Step 2 — Why it happens
You do not need the technical details, only the idea that makes everything else make sense.
A chat model writes by predicting, over and over, the most likely next piece of text given what came before. It learned those likelihoods from an enormous amount of writing. It is very good at producing text that looks like a correct answer. It is not, by itself, checking each statement against the world.
Four ingredients make a hallucination likely.
1. Fluency is the skill, truth is not built in. The model is trained to produce plausible, helpful-sounding text. Plausible and true overlap a lot, because most real writing is mostly true, but they are not the same thing.
2. Gaps get filled. When the question needs a fact the model does not have, for example an obscure person, a very recent event or a precise figure, the most likely-looking continuation is still a confident sentence. Saying "I do not know" is a different, rarer pattern that tools are trained to produce, but not always reliably.
3. The question can pull it. A leading question ("Why did the 2019 law ban X?") assumes something. A model may go along with a false premise rather than push back.
4. It was not looking anything up. Unless the tool is connected to search or to your documents, it answers from what it learned in training, which has a cut-off date and gaps. And as the tokens tutorial explained, even a pasted document can be used imperfectly.
Tools improve in this area, and features such as web search with citations, document grounding and "I am not sure" behaviour help. But all four ingredients remain, which is why habits matter more than hoping.
Step 3 — Where hallucination hides
Not every part of an answer is equally risky. Learn the hiding places, so you check there first.
Names and titles. People, companies, products, papers, books and cases. Especially full names with initials, and exact titles.
Numbers. Statistics, prices, percentages, dates, counts and measurements. A number looks precise, which makes it feel trustworthy.
Quotes. Anything in quotation marks attributed to a real person or document.
Sources and links. Citations, references and URLs. A link can look right and lead nowhere, or lead to something unrelated.
Recent events. Anything after the model's training cut-off, unless it searched.
Niche or local detail. Small businesses, local rules, regional facts and less common languages.
Technical specifics. A command, a setting, a menu path, a function name or a legal clause. They look exact and are often slightly wrong.
Long documents. A summary of a long text can drop a qualification, merge two sections or attribute a view to the wrong person.
Calculations. Multi-step arithmetic and logic. A model can write a clean explanation and still get the sum wrong.
Things it cannot know. What happened in a private meeting, what you meant, what is on a page it did not open. It may write as if it knows.
Step 4 — Confidence is not evidence
A chat AI sounds the same when it is right and when it is wrong. That is the heart of the problem, and it explains several habits that do not work.
Tone proves nothing. A fabricated date and a correct date arrive in the same sentence structure and the same calm voice.
"Are you sure?" is unreliable. Many people ask this as a safety check. It does not reliably work, for two reasons. First, the model has no separate fact-checking part to consult, so "checking" tends to be more of the same text generation. Second, models can be eager to please. Ask "are you sure?" about a correct answer and some will back down and change it to a wrong one. Ask about a wrong answer and some will insist on it. Neither tells you what is true.
Asking it to rate its own confidence helps a little, not enough. A request such as "Say how confident you are" can bring useful hedging, but the stated confidence is itself generated text and is not a measurement. Treat it as a hint about where to look, not as proof.
Repetition is not confirmation. Asking the same question three times and getting the same answer does not make it true. It may mean the model reliably produces that error.
A source name is not a source. "According to a 2021 study in a leading journal" sounds authoritative. Without a real, checkable reference it is just words.
What does work is checking against something outside the model: a source, a document, a second tool, your own knowledge or a calculator.
Step 5 — What actually reduces hallucination
You cannot switch it off, but you can lower the odds and make mistakes easier to spot. These work in any tool.
1. Give it the source and limit it to that source. If you have the document, paste or attach the relevant part and say so.
Answer using only the text below. If the answer is not in the text, say "not in the text" and do not guess. After each claim, quote the sentence it comes from.2. Ask for quotes you can check. A quoted sentence from a document you hold can be matched in seconds. A claim with no quote is a claim to verify.
3. Use a tool that searches and cites, and then open the sources. A search-connected tool that shows its sources is a good start, since you can click through. But the summary can still misstate what the page says, so open two sources yourself.
4. Allow "I do not know". Many models answer better when told it is acceptable to admit gaps.
If you are not sure, or the information may be out of date, say so plainly. Do not fill gaps with plausible details. Separate what you know from what you are guessing.5. Ask narrower questions. "Summarise everything about this company" invites invention. "List the three products named on this page" is much safer.
6. Avoid leading questions. "Why is X true?" pushes toward agreeing. Ask neutrally: "Is X true? What evidence is there for and against?"
7. Ask for the assumptions. "List the facts you relied on, and mark which you are least sure of" shows you where to look.
8. Split the job. One pass to gather facts with sources, another to write. Keep the unsourced material out of the final text.
9. Use the right kind of tool for the job. For facts that must be current and sourced, a search tool is better than a pure chat model. For your own files, a document-grounded tool is better. For maths, use a calculator or a tool that runs code. Choosing the kind of tool is covered in the next module.
Step 6 — A checking routine that takes minutes
Checking should be quick enough that you actually do it. This routine takes two to ten minutes for a normal answer.
1. Underline the checkable details. Skim the answer and mark every name, number, date, quote, source and specific instruction. Ignore the general explanation for now.
2. Pick three, and do not pick the first three. Choose one from the middle, one that looks surprising and one that matters most to your decision.
3. Check each against something outside the AI. Search the exact title of a paper in a scholar search. Open the link. Look up the figure on an official page. Match the quote against the document. Compare the statistic with a second source.
4. Look at what the source really says. Not just whether it exists, but whether it says what the AI claims. Read the relevant lines.
5. Scan for what is missing. Ask yourself whether an important caveat, an opposite view or a recent change is absent. Omissions do not trigger any warning.
6. Decide what to do with each claim. Keep it, correct it with what you found, or drop it. Never keep a claim you could not verify in something that matters.
7. Tell the AI what you found. "Source 2 does not exist. Please redo the answer without it, and mark anything you cannot support." Then check the new answer the same way.
A single failed check is information about the whole answer. If one of your three is wrong, check three more.
Step 7 — Match the checking to the stakes
Checking everything is slow, and checking nothing is risky. Decide by what happens if it is wrong.
| Stakes | Examples | How much to check |
|---|---|---|
| Low | Brainstorming names, a rough first draft, explaining a concept you will verify later | A glance; stay alert for anything odd |
| Medium | A report for colleagues, a study summary, research for a decision | The routine above: three checks, a scan for missing items |
| High | Anything with money, health, legal matters, safety, a public post or a message to a customer | A person reads every line; every number and source is verified, and an expert checks it where needed |
Two things move a job up a level without you noticing.
For medical, legal and financial questions in particular, use AI to understand the terms and prepare questions, and take the decision to a qualified person. This tutorial is not professional advice, and neither is a chat AI's answer.
Experiment to Run
This takes fifteen minutes and shows you the behaviour for yourself.
1. Pick a topic you know very well, such as your own job, a hobby, your hometown or a book you love.
2. Ask five questions, two easy, two detailed and one about something you suspect it does not know, such as a small local business or a minor event.
3. Mark each answer: right, partly wrong or made up. Write a line on how sure it sounded.
4. Try a leading question, such as one that assumes something false about your topic. See whether it corrects you or goes along with it.
5. Ask it to cite sources, and check two of them.
6. Repeat with the safeguards from Step 5: give a source, allow "I do not know" and ask for quotes. Compare.
Notice two things: how little the confident tone told you, and how much the safeguards changed the result. Your own notes on a topic you know are the best guide to how much to trust a tool.
Worked Example: a fake citation in a study summary
Ana is preparing a short brief on remote work for her manager. She asks an AI for the research on productivity and remote work, with sources. It returns a tidy summary and four references, each with authors, a year and a journal.
She uses the checking routine. She underlines the checkable details: a claim of "a 13 per cent productivity increase", a quote from one author, and the four references. She picks three: the percentage, the quote, and the reference that matters most to her argument.
She searches the exact title of the key reference in a scholar search. Nothing comes up with that title and those authors. She tries the journal's website. No such article. She marks that reference "not found".
She checks the percentage and finds a real study with a figure of that size, but for call-centre staff in one company, not for remote work in general. The AI had stretched it. She corrects her claim, and says so.
The quote is not in the document it was attributed to. She drops it.
Then she goes back to the AI: "Reference 2 does not exist and the quote is not in the source. Redo the summary using only the two sources I have opened, which I will paste. Quote each claim from them, and say 'not in the text' if it is not there." The new summary is shorter and plainer, and every claim has a quote she can verify. The check took twelve minutes, and it kept an invented reference out of a brief her manager would have relied on.
Common Mistakes to Avoid
Three habits do most of the damage.
Mistake #1: Trusting a confident tone. A smooth, specific answer feels reliable, and a wrong one sounds identical to a right one. Fix: treat names, numbers, quotes and sources as claims to check, whatever the tone, and be most suspicious of precise details on obscure topics.
Mistake #2: Using "are you sure?" as the check. The model has no separate fact-checking part, and it may back down from a correct answer or insist on a wrong one. Fix: check against something outside the AI, such as a source, a document, a second tool or a calculator, and give the AI the source up front.
Mistake #3: Checking the first three details, or only what is there. The first details are the easiest, and an omission triggers no warning. Fix: choose one from the middle, one surprising and one that matters most, and scan for what is missing. For anything high-stakes, a person reads every line.
Going Further
Build a personal "never trust blindly" list. A short list of the things you will always verify: any number you will repeat, any quote, any source, any instruction you will follow. It takes the decision out of the moment.
Use a second tool as a cross-check, carefully. Ask a different tool the same question, with the source in hand, and compare. Where the two differ, check by hand. Where they agree, remember that two tools can share the same error.
Learn what your tools do. Look for the features that help: search with citations, document grounding, a way to show sources. Test them on topics you know.
What to learn next. The next module looks at the landscape of tools and how to choose between them. Choosing the right kind of tool for a job is one of the best ways to reduce hallucination, since a search tool, a document tool and a chat model make different kinds of mistake.
Key Takeaways
In this tutorial you learned:
Before you move on, a challenge. Run the fifteen-minute experiment on a topic you know well, then run your checking routine on one real answer from your own work this week. Write down one thing the AI got wrong and how you caught it. Doing it once for real makes the habit stick.