Asking what it doesn't know

Require a confidence level on every non-obvious claim, then spend your checking on the low ones.

When to use this: When checking yourself
Learning impact: Clear help, modest gain
Learning evidence: Thin or mixed evidence
Where you are with the topic: Checking readiness
How often to use it: Worth repeating often

When to use it

You cannot check everything, and need a triage rule for what to verify first.
The answer sounds evenly confident from start to finish.
You have been marking claims [high]/[medium]/[low] but never following up on the lows.

What to ask AI

Type something like: “Answer my question, then mark every non-obvious claim [high], [medium] or [low] confidence.”

For each [low], require what it is unsure about and where you could check.

Keep the hard part on your side: you still verify. Tags only prioritise effort.

How it helps

Checking everything is impossible, so people check nothing.

Asking what it doesn’t know converts an impossible task into a short list.

Low-confidence flags tend to cluster around specific figures, exact dates, minor names, citations, and niche subfields.

That does not make the output reliable. A confidently stated error will still be marked [high].

It does make a habit survivable: check these four things, not the whole page.

Uncertainty forcing (once listed separately) is the same move: require a level, then verify the lows.

How to run it

Ask for confidence marks on every non-obvious claim before you start verifying.
Refuse softened lows. If it hedges a [low] into a [medium] under pressure, reset the rule.
Check the lows first, outside the chat, at a depth matching the stakes.
Spot-check one [high] claim you care about. Calibration is imperfect.
When a low flag was wrong (it was actually fine) or a high flag was wrong (it failed), note the pattern.

Example exchange

You

What dose of ibuprofen is typical for an adult headache? Mark every non-obvious claim [high], [medium], or [low], and for each [low] say where I should check. Don’t soften the lows.

Over-the-counter adult dosing is often discussed in the 200-400 mg range per dose [medium, confirm on a current label or formulary]. Daily limits and contraindications matter more than any single number [high as a caution]. Exact limits vary by product and country [low, check the pack or a trusted formulary; don’t act on chat].

You

I won’t use any dose from this chat. I’m checking the pack now.

Good. For medical dosing, primary-source checking is mandatory. My tags only show where I sounded unsure, not what is safe.

Copyable prompt

Answer my question, then mark every non-obvious claim [high], [medium] or [low]
confidence. For each [low], say specifically what you're unsure about and where
I could check it. Don't soften a [low] into a [medium] to seem more useful.

The Tell

Here is how you know this method has flipped: you have been tagging claims for weeks and never checked a flagged one.

Then the tags are decoration, and you added a ritual rather than a habit.

Open one [low] claim today against a real source before you ask another question.

Principle evidence

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

The underlying learning idea is rated weak-to-moderate. Metacognitive monitoring and confidence calibration are real learning constructs: knowing what you don’t know guides study. Mapping that onto model-reported confidence is weaker, because LLM certainty language is not calibrated human judgment. Triage-by-uncertainty as a checking strategy is useful hygiene; treating AI confidence tags as reliable remains speculative.

AI delivery evidence

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

No tutoring trial is claimed here for forcing AI confidence tags. Delivery is speculative. Use the tags only to prioritise your own checks, never as proof.