Avoiding Fake Expertise and Overconfident Claims

Unit ID: PAW-M04-U04 Estimated active time: 20-25 minutes

Fluent and certain are not evidence

An AI tool writes with the same confidence whether it is right or wrong. There is no wobble in the prose when it is guessing.

That is the core reason review cannot be skipped, and it is why role instructions that add authority are a risk rather than a feature.

Where false confidence appears

It clusters in predictable places:

• numbers and percentages that were not supplied;
• dates, versions, and prices;
• citations, standards, and clause numbers;
• claims about what "research shows" or "best practice is";
• statements about what a named organisation does.

Ask for the hedge to be visible

Instructions that make uncertainty explicit are more useful than instructions to be careful.

> After each claim, mark it as [FROM MY NOTES], [GENERAL KNOWLEDGE], or [UNCERTAIN]. Do not mark anything as certain that you cannot trace to the notes I gave you.

The output becomes reviewable. You go straight to the [UNCERTAIN] lines instead of reading everything with equal suspicion.

The claims audit

CLAIMS AUDIT

For each factual sentence in the output:

Claim              (quote it)
Source             my notes / general / unknown
Verifiable         yes -- how / no
Consequence if wrong   low / medium / high
Action             accept / verify / cut / rewrite as a question

Run it on the high-consequence claims only. Auditing every sentence of a routine email is not a good use of anyone's time.

Prefer questions to confident gaps

A useful standing instruction:

> Where you do not know something, write a question for me instead of an estimate.

This converts the failure mode from a false fact into a to-do item, which is a much better place for it.

Mini practice

Take an AI output you have used.

Run the claims audit on its five most consequential sentences.

Count how many you could actually trace. That count is your real evidence base.

Mark complete