1.What is hallucination?
Hallucination refers to an AI stating things that are not true as if they were facts. It produces, in smooth sentences, book titles and authors that don't exist, legal provisions that don't exist, wrong dates and quotations that differ from the real ones. The term is borrowed from the word for seeing things that aren't there.
Hallucination is dangerous because wrong answers don't look wrong. They come in the same confident tone and the same neat format as correct answers. A sentence sounding natural is no proof that its content is correct.
2.Why does it happen?
In Lesson 4 we saw that a language model picks the most plausible token to come next. A model is not a device that checks "Is this sentence true?" but one that calculates "Is this sentence plausible?" Widely known facts appear many times in the training data, so plausibility and fact almost always match. But for content that appears rarely in the data — rare facts, things that happened after training, exact numbers — the two can diverge.
For example, if you ask for the author of a little-known study, the model may answer by combining a "plausible name" and a "plausible journal title" for that field. For a model, continuing with a plausible answer can sometimes come more naturally than saying it doesn't know. The people who build models and services are adding various safeguards to reduce this, but it is safest to assume that, by its nature, it can never be eliminated completely.
3.Questions that tend to trigger hallucination
You don't need to doubt every answer equally. Knowing which kinds of questions are risky lets you save your checking effort for where it counts.
| Kind of question | Risk | Why |
|---|---|---|
| Sources, quotations, references, links | High | It can invent them by combining plausible titles and authors |
| Exact numbers, dates, statistics | High | There are many similar numbers, so they easily get mixed up |
| Recent events, current regulations, prices | High | It does not know information from after its training |
| Little-known people, small places, a particular company's circumstances | Medium to high | They appeared rarely in the training data |
| Information where mistakes cause real harm, such as law, medicine and taxes | High (by potential harm) | Even small errors have big consequences |
| Polishing text, summarizing, changing formats | Low | It works within the material you gave it, so there is little room to make things up (still check that the summary matches the original) |
4.Five habits for checking answers
Checking is not hard. Applying the habits below just to important answers is enough to filter out a large share of hallucinations.
- Check primary sources: check laws against the original text of the law, programs and regulations against the official guidance of the agency in charge, and statistics against the publishing agency's data
- Look up sources yourself: search for the titles of books, papers and articles the AI gave you to see whether they really exist and whether their content matches
- Recalculate numbers: check calculation results yourself with a calculator or spreadsheet (this is where the Math Basics course comes in)
- Ask again in a different way: rephrase the question or ask again in a new conversation to find the parts where the answer wavers
- Ask for the basis: have it mark the basis for its answer and the parts it is unsure of, and check those parts first
- First, remember that sources and quotations are a high-risk category for hallucination.
- Search for the titles and authors in an academic or library search to confirm they really exist.
- If they do, read at least the abstract of the original and compare it with the AI's summary.
- Leave any study you cannot find out of the report, and use only the confirmed ones, with their sources.
5.Verification in practice — following one number
Let's apply the five habits to an actual answer. The answer below is made up for illustration. Answers like this are easy to take at face value, because the sentences are smooth and they even come with a calculation.
Not every wrong AI answer is a hallucination. Different causes call for different ways of checking, so when you find a wrong answer, it helps to work out what kind of error it is as well.
- Calculation mistakes: the more steps there are, the easier it is to get a number wrong while carrying it over — check with a calculator or spreadsheet
- Outdated information: it states programs, prices or facts about people that changed after training as they used to be — check against official sources with dates
- Misreading the question: it answers something you didn't ask or drops one of the conditions — compare the conditions in your question with the answer one by one
- Misreading the material you gave it: it mixes content that isn't in the pasted material into a summary — have it mark which part of the original each sentence of the summary comes from
- Step 1: See whether it is a risky kind of question: it asks for an exact number, so it needs checking.
- Step 2: Recalculate it yourself: with compound interest you multiply by 1.05 each year. 1,000,000 × 1.05 = 1,050,000, × 1.05 = 1,102,500, × 1.05 = $1,157,625.
- Step 3: Find the cause of the difference: $1,150,000 is a simple-interest calculation, which adds interest only on the principal (1,000,000 × (1 + 0.05 × 3) = 1,150,000). The phrase "$50,000 each year" in the answer was already a sign of simple interest.
- Step 4: Correct it and ask again: if you say, "Not simple interest — use compound interest, and show the yearly multiplication in a table," you can compare the process line by line.
6.Prompts that reduce hallucination
Besides checking, there are ways of asking that make hallucination less likely in the first place. The most effective is to provide the source material yourself. If you paste in reliable material and say, "Answer only from this material, and if it isn't there, say so," there is far less room to make things up.
It also helps to give it permission to say it doesn't know. Writing "If you're not sure, it's fine to say you don't know" or "Mark the parts you guessed with (guess)" brings the uncertain parts of the answer to the surface. On the other hand, forcing a count, as in "You must give me exactly five," raises the risk that when it knows only three, it will make up the other two.
- Wrong: Give me five statistics related to this topic.Right: If there are statistics related to this topic that you know fairly reliably, tell me, and include the publishing organization and reference year for each. If you're not sure, don't fill up a set number; write "needs checking" instead.Not forcing a count, and having it mark sources and uncertainty, makes it clear what needs checking.
📌 Key points
- Hallucination is when an AI confidently states things that are not true as if they were facts
- It happens because the model calculates plausibility, not truth
- Always check sources, numbers, recent information, rare facts and information where mistakes cause harm
- The basic habits are checking primary sources, looking up sources yourself, recalculating numbers, asking again and asking for the basis
- Hallucination decreases when you provide source material, let it say it doesn't know and don't force a count
🤖 Try asking AI like this
Copy a prompt and replace the [ ] parts with your own situation. Don't take the answer on trust — check it against this lesson.
When you want to bring out the uncertainty before getting an important answer
Answer the question below. But at the end of each sentence, mark how sure you are with one of (certain)/(fairly certain)/(guess), and at the end, list the items I should check myself and where to check them (what kind of organization or source). Question: [your question]
When you want to check an answer you already got
In the answer you just gave me, pick the three places most likely to contain factual errors and explain why. Focus especially on numbers, dates, names and sources.
When you want answers based only on your material
Answer using only the material below. Never add anything that isn't in the material; write "Not in the material" instead. After answering, quote word for word the sentence in the material that each point came from. [material] Question: [your question]
- General fact-checking principles commonly found in guides to using generative AI
- The general definition of "hallucination" as used in the field of natural language processing
Reached every goal above? Mark the lesson complete.
Storage is unavailable in this browser, so this lasts only for this page.🤖 AI Basics
- 1What Is AI? — Rules and Learning
- 2Machine Learning Basics — Data, Model, Training, Evaluation
- 3Neural Networks, Intuitively — Small Calculations Add Up to Judgment
- 4How Do Large Language Models Produce Text?
- 5Prompt Basics — Saying Exactly What You Want
- 6Advanced Prompting — Roles, Examples, Steps, Format
- 7Hallucination and Verification — How to Doubt What Sounds Plausible
- 8Privacy, Copyright and Ethics — Using AI Responsibly
- 9Working with AI — Delegate, Check, Connect
- 10Learning Strategies for the AI Era — The Basics Shape Your Questions