Skip to content
unmagic.ai

Concept 04 · 7 min

It doesn’t lie. It doesn’t know it’s wrong.

Scroll to start

Scene 1

Three questions, one tone

Sarah asks the assistant three quick questions. No file attached: it answers from what it learned.

Assistant · no file
  1. Sarah

    What’s the capital of Australia?

  2. Sarah

    What’s Mr. Miller’s current overdraft limit?

  3. Sarah

    What year was Miller Joinery founded?

Will it say it doesn’t know, this time too?

Same company, same lack of facts: it declined once, then made up a year, in the same calm tone. An answer invented with confidence is called a hallucination: plausible text with nothing true behind it. Nothing in the tone sets it apart from a right answer.

Scene 2

Spot the hallucination

This time, Sarah gives it Mr. Miller’s client file and asks for a summary for the credit file, with revenue, founding year and headcount. Here is the file, then the summary.

The file it was given

  • Client since 2019.
  • Contact: Mr. Miller, owner.
  • Accounts: business current account, savings account, equipment loan.
  • Overdraft limit: €10,000.
  • Mr. Miller has asked to raise his overdraft limit.
  • One invoice is 12 days overdue.

Select the assistant’s 4 hallucinations.

Credit file summary: Miller Joinery

The summary reads perfectly, yet half of it isn’t in the file: four hallucinations, with the file in hand.

Scene 3

Make it honest

Sarah needs the credit file form filled in. Two settings: give it the file, and allow it to write “Not in the file”. Find the combination with no hallucination.

Challenge

  • Zero hallucinations, and every fact the file has.
  • Bonus: catch it hallucinating with the file in hand.

Choose what Sarah gives it

Ask for the form

Credit file form

Client since
…
Founded
…
Revenue last year
…
Employees
…
Overdraft limit
…
Overdue invoices
…

Giving it the facts is called grounding the answer in a source. Hallucinations become rarer but never impossible. Always check the important details.

Scene 4

When to trust it

Five things you might ask it for. Would you skim the result, or check it against the source?

Drag each card into the right column: “Skim it” or “Check against the source”.0/5 sorted

No mouse? Tap a card, then a column.

      Rule of thumb: the more it costs to be wrong, the closer you check.

      Hype vs reality

      • What the hype says

        “Hallucinations are a bug that will soon be fixed.”

        What actually happens

        They come from how these models work: always produce a plausible next word. They get rarer, not impossible.

      • What the hype says

        “If it gives a source, you can trust it.”

        What actually happens

        It can invent sources too. Open the source and check the sentence is really there.

      • What the hype says

        “It sounds sure of itself, so it’s probably right.”

        What actually happens

        Its tone is the same when it’s right and when it’s wrong. Confidence is style, not evidence.

      Under the hood

      In 3 sentences

      1. A model always writes the most plausible answer; when it has nothing to go on, that’s a hallucination.
      2. It can’t tell plausible from true, so its confidence proves nothing.
      3. Give it the facts, allow it to say “not in the file”, and check what matters.

      Did this make sense?

      Next step

      How do I make it use our internal documents?

      05RAGStart

      The open-book exam

      Use this journey with your teams

      Recorded with Ministral 3 (3B), Oct 9, 2026.

      1. LLM
      2. Context
      3. Prompt
      4. Hallucinations
      5. RAG
      6. Tools
      7. Agent
      8. Skills
      9. Limits & safety
      Home: the journey
      For teams
      Lire en français
      Theme: light
      Theme: dark
      Theme: follow the system
      ↑↓ to move · ↵ to open · esc to close