Methods · Working with the tool itself · Short guide

Hallucination hunting

Find planted errors in AI output: training the verification reflex before you depend on the tool for real work.

When to use this: While you practise
Learning impact: Reliable gain
Learning evidence: Credible with caveats
Where you are with the topic: Checking readiness
How often to use it: Worth repeating often

When to use it

You want to train error-spotting before an essay week that depends on AI help.
You ace explanations in chat, then miss subtle wrong numbers in real drafts.
You need a timed drill that makes confident mistakes feel visible.

What to ask AI

Type something like: “Write ~200 words on [TOPIC] containing exactly [N] factual or reasoning errors. Do not mark them.”

Hunt first. Only then ask it to reveal and classify what you missed.

Keep the hard part on your side: finding the errors is the learning event.

How it helps

Verification in the wild is hard because you do not know where the landmines are.

Hallucination hunting plants a known number of errors so you can practise the reflex safely.

You learn the shape of confident interpolation: the wrong date, the tidy fake mechanism, the too-convenient citation.

Transfer to live use is not automatic. Pair this drill with the verification habit on real work.

How to run it

Pick a topic you already know well enough to mark yourself.
Ask for a short passage with a fixed number of unmarked errors.
Find them under time pressure before you ask for the key.
Compare your list to the reveal. Classify misses: fabrication, outdated consensus, bad mechanism, arithmetic.
Repeat weekly until misses cluster on one class, then target that class.

Example exchange

You

Write about 200 words on mitosis for a first-year bio student, and hide three factual or reasoning errors in it. Don’t mark them. I’ll try to find them, then you can score me.

Here’s a short paragraph on mitosis. Go ahead and hunt. I’ll wait until you list what you think is wrong.

You

I found two: chromosomes “disappear” in interphase, and cytokinesis listed as a mitotic stage. I’m stuck on a third.

You caught two. The third was a confident wrong chromosome count for humans, slipped into an otherwise correct sentence. That kind, specific, unmarked, surrounded by good prose, is what usually costs marks.

Copyable prompt

Write ~200 words on [TOPIC] containing exactly [N] factual/reasoning errors.
Do not mark them. I will hunt. Then reveal and classify what I missed.

The Tell

Here is how you know this method has flipped: you ace the drills, then paste unverified answers into essays.

The hunt trained recognition, not the habit of checking live work.

Move one planted-error skill into every real draft: open at least one claim you would otherwise trust.

Principle evidence

Strength of the underlying learning idea, not a claim about AI products.

The underlying learning idea is rated moderate. Error-detection practice and worked examples that include mistakes help learners notice flaws rather than only copy correct solutions. This method is that drill applied to planted AI failures. Principle moderate for error-spotting; transfer from games to live AI use is plausible but not settled.

AI delivery evidence

Whether an AI tutor delivers this method well is a separate question.

Whether an AI plants useful, realistic errors on demand is a delivery question, and it is speculative here. The learning happens in your hunt and classification, not in the model’s cleverness at hiding mistakes.