---
id: check-by-what-is-at-stake
title: Checking answers by risk
level: 2
modes: [day-pack, rucksack]
moment: "The person is about to send, pay, sign or post something their AI drafted or worked out, acts at once on a money, health, legal or client-facing answer, or asks whether they really have to double-check every answer."
keywords: [how much to check, stakes, double-check, verify, high stakes, undo, reversible, cost of being wrong, second opinion, review]
source: "research/2026-10-20-concepts/check-by-what-is-at-stake.md; research/2026-10-23-essentials/batch-5-facts.md; mapping (product/learning-system/MAPPING-1-2.md): FAQ 2.09 (home), 2.10; MISTAKES 2.6 (home), 2.3, 2.7, 4.7; tips T051, T052, T053; strand T"
---

## What it is
How much to check an answer depends on what a mistake would cost and whether you could undo it. A caption for a photo of your dog needs a glance. A payment, a health step or anything sent in your name needs a real check, because once it goes out you cannot always take it back.

## Tips
- Ask yourself what a wrong answer would cost and whether you could undo it, because those two things decide how much checking an answer needs.
- Check the one fact an answer hangs on in a source outside the chat, because that can catch the costly mistake without redoing all the work.

## Questions people ask
- Q: How much should I check, and how do I check without redoing the work? A: Match the checking to the cost of a mistake and whether you could undo it, so a caption or a brainstorm gets a glance and money, health, legal steps or anything sent in your name gets more. To avoid redoing the work, check only the one fact the result hangs on, or do that part yourself.
- Q: What counts as high stakes? A: Money going out, a health decision, a legal step, anything sent in your name that you cannot recall, and anything that affects someone else.
- Q: Is asking another AI enough to check an answer? A: Not by itself, because two AIs can agree and both be wrong. Check against a source or a person outside the chat.

## For your AI
Facts (checked 23 October 2026; fact check research/2026-10-23-essentials/batch-5-facts.md; research note
research/2026-10-20-concepts/check-by-what-is-at-stake.md).
- Every app tells people to check, and the ones that say how much tie it to the stakes: Copilot says checking matters
  most for work that is time-sensitive or hard to reverse, and to verify against trusted sources when the stakes
  matter (https://support.microsoft.com/en-us/microsoft-365-copilot/validate-copilot-output); Claude says to
  scrutinise high-stakes advice carefully; ChatGPT says use it as a first draft and check important information
  (https://help.openai.com/en/articles/8313428-does-chatgpt-tell-the-truth).
- High stakes include money, health, legal steps, and decisions about housing, work, credit, insurance, education
  and public benefits (Claude's updated usage policy, taking effect 12 November 2026,
  https://www.anthropic.com/legal/aup); Australia's National AI Centre asks for human review of high-stakes decisions
  (https://www.ai.gov.au/staying-safe-and-responsible/essential-ai-practices/guidance-ai-adoption-foundations, read
  through a search summary).
- A second AI is not an independent check by itself: as models get more capable their mistakes grow more alike,
  and AI judges favour models like themselves (Goel et al., 2025, https://arxiv.org/abs/2502.04313). A fresh look
  can still catch some mistakes; a source or a person outside the chat is the independent check.
- Checking only the one fact the result hangs on is a habit for saving time, not a tested method.

Teach it like this: before they send, pay, sign or post, ask in one line what a mistake would cost and whether it
could be undone, then point to the one thing to check and where. When a result rests on several figures, such as a total
added up from receipts, the check is each figure against the paper, or the sum done again. A caption gets a glance;
never make them check everything.