Why does AI make things up? Hallucinations explained

Learn in your browser · Chapter 01 of 06

Ask an AI chatbot a question and it will always give you an answer, in the same confident voice, whether that answer is true or invented. This chapter names the problem clearly, with real examples, before the rest of the workbook explains why it happens and what to do about it.

A hallucination is when an AI states something false or made up as though it were a plain fact. It is not the chatbot lying to you on purpose. It is producing text that sounds right, in the same fluent and assured tone it uses for everything else, whether the underlying claim is true or not.

Scroll sideways to see every column.

Learn in your browser · Chapter 02 of 06

The mechanism behind a hallucination is simpler than it sounds. Once you understand it, hallucinations stop feeling mysterious and start feeling predictable.

A large language model is not a database with a search box attached. During training it is shown an enormous amount of text and learns one skill, repeated billions of times: given some text so far, predict a likely next small piece of text, called a token. It does this once, adds the result, then does it again, and again, until it has written a full reply.

That is genuinely all it is doing at the point it writes each word. There is no separate step where it looks a fact up in a trusted reference and reads it back to you. It generates text that fits the patterns it learned, and most of the time those patterns line up with what is actually true, because most of its training text was accurate. When they do not line up, you get a hallucination, produced by exactly the same process as every correct answer it has ever given you.

Two more pieces complete the picture. A model's training data has a cutoff date, so on its own it does not know about anything that happened after that point, and it will sometimes still answer as if it does unless a tool adds live search. And a model has no built-in step that checks its own output against a trusted source after it is written, unless the tool wrapped around it specifically adds one. A later chapter covers the tools that do add that step.

Learn in your browser · Chapter 03 of 06

The single habit that causes the most damage is trusting an AI answer because it sounds sure. This chapter explains why tone tells you nothing about accuracy on its own, and what does.

A language model has no internal dial that measures how certain it actually is, and no built-in habit of hedging only when it should. It writes in whatever tone fits the pattern of a good answer, which in most training text is a confident, complete-sounding one, whether the underlying content in this particular case happens to be right or wrong.

There is also a small amount of randomness built into how a model chooses each next token, which is why asking the exact same question twice can produce two different answers, sometimes one accurate and one not. This is a deliberate design choice that makes writing feel natural and varied, not a fault, but it means consistency across repeated answers is a somewhat better signal than confidence in any single answer.

Scroll sideways to see every column.

Learn in your browser · Chapter 04 of 06

You now know why hallucinations happen. This chapter turns that understanding into a short, repeatable checklist you can run on any AI answer that matters.

Not every AI answer needs the same level of checking. The sensible first step is always to weigh how much it would actually cost you if this particular answer turned out to be wrong.

1. Ask yourself first: if this is wrong, does it actually matter? A low-stakes creative suggestion needs far less checking than a number that will go into a report, an email or a decision.
2. Isolate the specific, checkable claims in the answer: names, dates, numbers, quotations and anything stated as a fact rather than an opinion.
3. Check each one against a source you find and trust yourself, not merely a source the AI names.
4. If the AI gives you a link, a citation or a reference, open it. Confirm it actually exists and says what the AI claims it says.
5. Ask the AI to check its own answer, or to answer the same question again in a different way, and compare. Disagreement is a clear signal to dig further; agreement is a mild reassurance, not proof.
6. For anything involving health, law, money or safety, treat the AI's answer as a starting point for better questions, not as the decision itself, and involve a qualified person.

Learn in your browser · Chapter 05 of 06

None of this means AI is unusable for factual questions. Several genuine, well-established techniques measurably cut how often you hit a hallucination. This chapter covers the three that matter most to you as a user.

Grounding means giving the model real source material, your own document, a specific webpage, a report, and asking it to answer using only that material rather than relying purely on what it remembers from training. A grounded answer is far more likely to be accurate, because the model is now summarising and quoting text in front of it rather than recalling a pattern from memory alone.

Retrieval-augmented generation, usually shortened to RAG, is the technical name for a system that automatically finds relevant real documents and feeds them to the model before it answers, so every reply is grounded in something a human can check, often with the source shown alongside the answer. Many modern chatbot tools use a version of this automatically when they search the web or read a file you upload. A dedicated Trust Agent workbook covers building one of these systems in full technical depth; this chapter only covers what it does for you as a user.

Asking for sources, and actually checking them, does two things: it sometimes nudges a tool toward search-backed answers rather than pure memory, and it gives you something concrete to verify. The checking step is not optional. An invented-sounding citation is itself a hallucination.

Cross-checking across tools, or better, with a trusted human source, is a simple and effective habit. Genuine facts tend to agree across independent sources. Invented ones often do not survive a second, differently-worded question.

Learn in your browser · Chapter 06 of 06

You now understand why AI hallucinates and what genuinely helps. The last step is turning that understanding into a habit you actually keep.

You can now explain, in plain terms, why a language model can state something false in a completely confident voice. You know that tone is not evidence, and what actually is. You have a short checklist you can run on any AI answer that matters, and you know the three things that genuinely reduce the problem: grounding, asking for and checking sources, and cross-checking.

Keep this workbook. It is yours to keep, print and write in. Come back to the checklist chapter whenever an AI answer really matters, and run it properly rather than trusting the tone.

https://trust-agent.ai/exam/why-ai-makes-things-up

This is free educational content published by Mickai LTD, not professional, legal, financial or compliance advice, and Mickai accepts no liability for how it's used.

Trust Agent publishes workbooks like this one for free because knowing how to catch an AI's confident mistake is one of the most useful skills anyone can build right now, and it sits at the centre of why trust and human verification matter alongside AI's real capability.