The verification drill

Tell, reliably and quickly, when the thing you're learning from is wrong, and build that into a habit rather than a worry.

When to use this: When checking yourself
Learning impact: Strong change
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 just got a confident explanation and need to know which claims are checkable.
An essay draft cites papers you have not opened yet.
You keep feeling vaguely suspicious of AI answers, then believing them anyway.

What to ask AI

Type something like: “Answer, then mark each non-obvious claim [high]/[medium]/[low] confidence.”

Ask it to flag what it is least sure about, then you check those first against a real source.

Keep the hard part on your side: you verify outside the chat. Asking “are you sure?” is not a check.

How it helps

Everyone knows AI makes things up.

That knowledge is almost useless on its own, because the errors that matter do not look wrong.

They are specific, confident, and nestled inside mostly correct paragraphs.

A fabricated citation names a real author and a plausible year.

A wrong constant sits in the right units.

Vague scepticism either paralyses you or wears off, because most of what you get back is fine.

What works is a drill: classify the claim, ask for confidence markers, check the checkable at the right depth, and push back once on something you know is correct to see if it folds.

Without that last calibration step, its agreement tells you almost nothing.

How to run it

Classify each claim: structural (check by reasoning), factual-specific (look it up), or consensus (check currency and disagreement).
Require confidence markers on non-obvious claims, then spend your checking time on the lows.
Sanity-check magnitudes and units in ten seconds; open an independent source for anything that will enter assessed work.
Push back once on something you believe is correct. If it folds immediately, treat later agreement as unconfirmed.
Practise with planted errors on topics you already know, so you learn the shape of a confident mistake.
Log the errors you catch: class, tell, and how you found them. The log trains the habit.

Example exchange

You

Can you explain the Treaty of Westphalia in about 200 words? For anything that isn’t obvious, mark it [high], [medium], or [low] confidence, and for the lows tell me where I could check.

The war ended in 1648 [high]. Many textbooks say the treaty “established the modern system of state sovereignty” [medium]. Historians still argue how much that story overstates what 1648 actually did [low, a recent historiography survey would help].

You

I’ll open a survey chapter before I use the sovereignty line. Quick pushback test: I claim historians treat 1648 as uncontested. Don’t agree just to be nice.

That claim is too strong. The date is solid; the sovereignty story is contested. If you need the dispute for your essay, verify it in a source, not in this chat.

Copyable prompt

Answer, then mark each non-obvious claim [high]/[medium]/[low] confidence.
For every [low], say what you'd need to be sure and where I could check.

The Tell

Here is how you know this method has flipped: you have never caught an error.

Either you are not checking, or you are only checking claims you already knew were safe.

Run a planted-error drill on a topic you know well and find out which.

Principle evidence

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

The underlying learning idea is rated moderate. Source evaluation and critical information literacy are long-standing educational constructs: learners who check claims against independent evidence outperform those who trust fluency alone. Applied to AI output, that means spot-checking dates, numbers, and citations rather than vague scepticism. Verification-as-literacy is moderate; controlled trials of AI-specific verification drills as pedagogy remain thin.

AI delivery evidence

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

A moderate principle does not prove chatbots make you a better verifier. Treat AI delivery as speculative unless a tutoring study of this drill is named. None is claimed here. The value is the procedure you run outside the model, not the model’s confidence tags alone.