// Lesson 2 of 3 · ~20 min · Beginner

Why models make things up (and what to do about it)

After this lesson you can recognise a likely hallucination and set up a prompt that makes one less likely and easier to catch.

You will be able to

Recognise when an answer is not grounded in anything you provided
Apply “I don’t know”, quotes-first and citation checks
Know that these techniques reduce, not eliminate, errors

Warm-up · Activity 1 of 7
Recall from the last lesson: which of these counts toward the context window?
Only the system promptYour messages, but not the model’s repliesOnly your latest messageEverything in the request, including the reply being written
Predict · Activity 2 of 7
You ask an assistant: “On which page of our annual report is the CEO quoted about hiring?” You did not attach the report. It replies, confidently: “Page 47.” What most likely happened?
It must be right, because it answered without hedgingIt produced a plausible answer that is not grounded in anything: a hallucinationIt found the report online and checkedIt refused to answer
Practice · Activity 3 of 7
Which of these are techniques Anthropic’s guide lists for reducing hallucinations? Pick all that apply.
Select all that apply.
Ask for a longer answerAllow the model to say “I don’t know”Warn the model that it will be punished for mistakesAsk for direct quotes first, then answer based on themRequire citations, and retract claims that have no supporting quote
Practice · Activity 4 of 7
Put this grounding workflow for a long document into the right order.
1. 1.Provide the document in the prompt
2. 2.Check that each claim is backed by a quote
3. 3.Ask the model to extract word-for-word quotes relevant to the question
4. 4.Have it answer using only those quotes
5. 5.Remove any claim that has no supporting quote
Practice · Activity 5 of 7
Complete the instruction that gives the model permission to admit uncertainty.
If you are not sure of the answer, say “Your answer”.
Brain teaser · Activity 6 of 7
Brain teaser. You ask the same factual question five times in fresh chats. Four answers agree; one gives a completely different date. What does the disagreement tell you?
Inconsistency is a warning sign that the answer may not be groundedThe odd one out is the correct answerThe model is broken and should not be usedNothing: the majority is always right
Apply · Activity 7 of 7
Mini-task. Rewrite this prompt so a hallucination is less likely and easier to catch: “Summarise the attached quarterly report and include the key figures.”
Check your work against this list
- It allows the model to say it does not know
- It asks for quotes (or exact figures with their location) before the summary
- It restricts the answer to the attached report
- It asks for a source for every figure

Exit ticket

5 questions, no hints. Score 80% or more to complete the lesson.

Finish every activity above to unlock the exit ticket.

Key ideas

Fluent is not the same as true

Language models are trained to predict the next word from the text that comes before it, as Anthropic’s glossary puts it. When the answer is not in its context, a plausible-sounding guess can come out in exactly the same confident tone as a fact. Tone tells you nothing about accuracy.

Give it a way out, and a way to show its work

Explicitly allowing “I don’t know” reduces invented answers. For long documents, ask for word-for-word quotes first and an answer based only on those quotes. Then check that every claim has a supporting quote, and retract any that do not.

Reduce, then verify

Anthropic’s guide is explicit that these techniques reduce hallucinations but do not eliminate them. Other checks it lists include comparing several answers to the same question (best-of-N) and restricting the model to the documents you provided. For anything that matters, a human still checks the output.

Sources

Last reviewed September 28, 2026

Why models make things up (and what to do about it)

Recall from the last lesson: which of these counts toward the context window?

You ask an assistant: “On which page of our annual report is the CEO quoted about hiring?” You did not attach the report. It replies, confidently: “Page 47.” What most likely happened?

Which of these are techniques Anthropic’s guide lists for reducing hallucinations? Pick all that apply.

Put this grounding workflow for a long document into the right order.

Complete the instruction that gives the model permission to admit uncertainty.

Brain teaser. You ask the same factual question five times in fresh chats. Four answers agree; one gives a completely different date. What does the disagreement tell you?

Mini-task. Rewrite this prompt so a hallucination is less likely and easier to catch: “Summarise the attached quarterly report and include the key figures.”