PromptForge
Student Hub

The Fabrication Hunt

ChallengeIntermediate25 minAny

Deliberately induce a confident, wrong answer, then find the prompt wording that prevents it. The most useful twenty minutes you can spend on calibrating how much to trust a model.

Why this one

You cannot calibrate trust from articles about hallucination. You calibrate it by making a model confidently invent something in front of you, in a domain you know well enough to catch it.

0 / 5 steps

Steps

  1. 01Ask about something obscure but checkable

    A minor detail of a niche topic in your own field. Obscure enough to be thinly represented in training data, checkable enough that you can verify the answer in five minutes.

  2. 02Push for specifics

    Ask for dates, names, numbers and sources. Specificity is where fabrication surfaces — a model that is vague is often being appropriately uncertain, and pressing for precision converts that into invention you can see.

  3. 03Verify every specific

    Check each claim against a real source. Note which were right, which were subtly wrong, and which were invented outright. The subtly-wrong category is the dangerous one and the one people never test for.

  4. 04Find the wording that prevents it

    Re-ask with an explicit uncertainty instruction and compare.

    Prompt
    Answer the question below. Before each claim, mark it [CERTAIN], [LIKELY] or [UNSURE]. If you do not know a specific detail, write "I do not know" — do not produce a plausible-sounding value. At the end, list what you would need to look up to answer fully.

QUESTION: [your obscure question]

  5. 05Write down what you learned about this model

    One line: which kinds of question this model answers reliably and which it does not. That line is worth more than any general advice about AI accuracy.

You should end up with

A documented fabrication, the prompt that prevented it, and a one-line trust rule.

Done when

  • You produced at least one confidently stated false specific
  • The uncertainty instruction measurably changed the output
  • You can state where this model is and is not reliable for your work

If you want to go further

  • Repeat with a different model and compare where each one breaks

Try next