Reality Lab · 10 MIN · Judge

Match the check to the risk

You can’t fact-check everything to the same depth. Given five AI outputs, decide how hard to verify each — so the effort lands where being wrong actually costs something.

THE SITUATION

In one afternoon someone uses an assistant for five things: brainstorming names for a team offsite, drafting a personal reminder to themselves, summarising a customer contract, writing a dosage note for a family member’s medication, and generating a snippet of code that will touch live payments. Treating all five with the same suspicion is exhausting; trusting all five equally is dangerous.

Your task: Sort the five outputs by the cost of being wrong, and decide the level of checking each one warrants — from ‘glance and go’ to ‘independent expert verification.’

THE FIVE OUTPUTS, AS GENERATED

1. A list of 20 punny names for the team offsite.
2. A note-to-self: “remember to water the plants and call the dentist.”
3. A one-page summary of a 40-page supplier contract.
4. “Take two tablets every 4 hours” — a paraphrase of a medication label.
5. A code snippet that calculates and applies a refund to a customer’s card.

YOUR WORKSPACEIn the box below (local to your browser), rank the five from lowest to highest stakes and write the check each one gets. Be specific: what would you verify, against what, and who (if anyone) other than you needs to look before it’s acted on?

This box is local to your browser. Nothing you type is sent anywhere or saved after you leave.

Check your own work

There's no grade here — checking against these is the exercise. Compare each point to what you wrote.

The stakes ranking reflects reversibility and harm, not how confident the output sounds.All five read fluently. Fluency is uniform; consequences aren’t. The medication note and the refund code can hurt someone or move real money — the offsite names can’t.
The low-stakes items get a light touch, on purpose.Checking punny names or a note-to-self as hard as a contract wastes the attention you need for the dangerous ones. Right-sizing down is as important as right-sizing up.
The contract summary is checked against the source document, not trusted as a summary.Summarisation quietly drops or distorts terms. For anything you’ll rely on, you verify the summary against the clauses that matter — indemnity, termination, price.
The medication note and the payment code escalate beyond ‘I checked it’ to a qualified human or a test.These are the ‘stop and involve an expert’ cases: a health detail and irreversible movement of money. A pharmacist confirms the dose; the refund code is reviewed and tested before it touches a live card.

Reveal a worked approach

A sound triage: (1) offsite names — glance and pick, wrong is free. (2) note-to-self — no check, only you bear it. (3) contract summary — read it against the actual contract, spot-check the high-value clauses, don’t sign off blind. (4) medication note — do not rely on the paraphrase; confirm against the label and a pharmacist, because a wrong dose is irreversible harm. (5) refund code — treat as untrusted; review it, test it on a sandbox, and never let it run against live payments unreviewed.

Notice the spread: two got almost no checking, one got a source comparison, and two got escalated out of your hands entirely. That range is the whole point — a single verification habit applied to everything is either paranoid or reckless.

The habit to carry: before trusting any AI output, ask “if this is wrong, what happens, and can I undo it?” The answer sets how hard you look — and whether looking is even enough.

LANDMARKS

CONCEPTS