Understand LLMs · Lesson 04

Why can an LLM hallucinate?

Why can a linguistically convincing answer still be false or unsupported?

Level
Beginner
Core time (estimate)
13 min
Updated
11 September 2026
Prerequisites
Progress
Not started

What you will be able to explain

Explain why fluent generation is not fact-checking, distinguish supported, contradicted and insufficiently evidenced claims, and choose a verification step proportionate to the stakes.

The essential path is the explanation, the five-step activity and the quiz. The presentation and technical notes are optional. You can mark the lesson complete at any time.

Start with the mechanism

A token is a piece of text defined by a tokenizer, not necessarily a word. A distribution assigns probabilities across the possible next tokens. The context is the input available for this step.

Three steps to keep in mind

  1. 1

    Split the answer into claims

    A polished paragraph may contain several independently checkable statements.

  2. 2

    Find suitable evidence

    Prefer primary, current, accessible sources that directly address each claim.

  3. 3

    Classify support

    Evidence can support, contradict, or remain insufficient for a conclusion.

Try it yourself

Follow these five steps in order. Everything in the lab runs locally; no real AI model is called.

1 · Predict

The first claim says that 42 people attended a workshop. Would a schedule giving the opening time prove that number? Predict the verdict before selecting evidence.

2 · Manipulate

  1. Select Case 1 and start with neither source checked. Keep the claim fixed. Select only Workshop schedule and read the verdict; the schedule describes sessions, not attendance.
  2. Leave the schedule selected and add only Workshop report, Attendance section. Compare the verdict before and after adding this direct passage. Uncheck that report to check that the verdict returns.
  3. As a separate case, select Case 2; changing cases clears the selected evidence. Read the claimed year, then select only Official invitation. Compare the exact dates, not just the fact that both mention June 18.
  4. Select Case 3 and read both excerpts before checking either. Add them one at a time. Ask whether registration growth during a poster campaign proves that the poster caused it.

Connect a claim to evidence

Verification workshop

Choose a case, read the prepared sources, and select only those that help assess the claim.

Local fictional corpus

Claim

The report says that 42 people attended the workshop.

Does the report support the announced number?

Evidence to consider

Teaching verdict

Insufficient evidence

The selected material does not directly confirm the claim.

Limits of the workshop
  • Claims and documents are fictional and prepared locally.
  • The lab does not search the internet or run an LLM.
  • In a real verification, a relevant source must still be authenticated, dated, and put in context.

3 · Observe

The schedule alone leaves ‘Insufficient evidence’. Adding the attendance report changes the prepared verdict to ‘Claim supported’. In Case 2 the invitation contradicts the claimed year. Case 3 stays insufficient even with both passages: a correlation and a survey do not establish the claimed causal link.

4 · Explain

In your own words, explain why adding one passage changed the first verdict but adding more passages did not establish causation in the third case. Name the next check you would make before publishing an important claim. Use speech or paper; nothing is collected.

After your own explanationCompare with one possible explanation

Evidence must address the exact claim. The attendance passage directly addresses the number, while the schedule does not. More text is not automatically stronger evidence: neither poster passage rules out alternative causes. In a real check I would authenticate and date the original source, inspect its scope and methods, and keep the conclusion ‘not verified’ when the necessary evidence is missing.

5 · Qualify

Open “Check a claim”

Check the model in your head

Six short questions, each with an explanation. You can retry or skip the quiz; your best score stays on this device and is shared between languages.

No tricks, just explanations

Check your understanding

Nothing is locked by this quiz. Use mistakes to refine your explanation.

Question 1 of 6
A very fluent answer must be factually correct.

Keep these three ideas

Go deeper when you need it

The essential path is complete. Open only the resources you need; none are required to finish.

Technical detailWhat the short version leaves out

Retrieval and tools can ground an answer, but they can also retrieve irrelevant material or be interpreted incorrectly.

Useful systems expose uncertainty, separate facts from hypotheses, and make sources easy to inspect.

Training data can contain mistakes and gaps; a request can be ambiguous or based on a false premise. Missing evidence, poor retrieval or pressure to answer can contribute. ‘Hallucination’ does not by itself diagnose which cause produced an output.

Prioritize claims that affect a decision. Check source authenticity, date, scope and the precise passage. For an important published date or number, open the original; for high-stakes decisions, use authoritative sources and a qualified professional. Missing evidence means ‘not verified’, not permission to present a guess as fact.

Reusable teaching resourceOpen the eight-slide presentation

Use this as a recap or a teaching outline. Download the complete Markdown file for reuse without a network request.

When the presentation has focus, use the left and right arrows to change slides.

The question

Why can a fluent answer be wrong?

Token prediction learns patterns of plausible continuation. It is not automatically a procedure for verifying current facts.

Slide 1: Why can a fluent answer be wrong?

1 / 8

Sources and further reading

References checked for this English lesson on . Publication years are listed separately. These explain the mechanisms, not a current ranking of products; real systems and documentation evolve.

Primary sources

Your local progress

Status: Not started

Progress stays on this device.