Quick answer
Ask the AI to show how it got its answer in a form you can test: sources with links, step-by-step working, formulas instead of typed-in numbers, and a list of anything it assumed or couldn’t confirm. Then spot-check one or two of those items. It takes a minute to ask and a couple of minutes to check, and it catches most confident-sounding mistakes.
Key takeaways
- Checkable beats confident. An answer you can test is worth more than one that just sounds right.
- Four requests do most of the work: sources, working, assumptions and a confidence note.
- Spot-check, don’t re-do. Checking one or two claims is usually enough to judge the rest.
- Go further for money, health, legal or public claims.
📘 Main article: This guide follows our news round-up Claude Helps Prove the Riemann Hypothesis: What It Means, Plus the GLM 5.5 Teaser and This Week’s AI News. The Riemann Hypothesis proof was trusted because a computer could check every step; this guide applies the same idea to everyday AI use.
When a mathematician used Claude to help prove the Riemann Hypothesis, nobody was asked to simply trust the AI. The proof was written so computers could check every step. You won’t be writing proofs, but the principle scales down perfectly: get AI to hand you work you can check, not just a conclusion.
Why ask for checkable answers?
- AI tools can state wrong facts, figures and quotes in a perfectly confident tone.
- A bare answer gives you nothing to test, so you either trust it or start again.
- Asking for checkable output often makes the AI more careful, because it has to show its route.
What should you ask AI to include?
Add this to the end of any factual, research or number-based request:
💡 Tip: “Show your working so I can check it: 1) link the source for each key fact, 2) show any calculations step by step, 3) list any assumptions you made, and 4) mark anything you’re not sure about or couldn’t confirm.”
| Request | What it lets you check |
|---|---|
| Sources with links | Whether the fact exists where the AI says it does |
| Step-by-step working | Whether the logic or maths holds up |
| Assumptions | Whether it guessed something you know is wrong |
| Uncertainty notes | Where to focus your checking |
What does a checkable answer look like for different tasks?
| Task | Ask for | Quick check |
|---|---|---|
| Research or facts | Links next to each claim, with quotes | Open two links and find the quoted text |
| Spreadsheets and numbers | Formulas, not typed totals | Change one input and see if totals update |
| Pricing or plan comparisons | Official page links and the date checked | Check one price on the official page |
| Code or automations | A test that proves it works | Run the test yourself |
| Summaries of a document | Page or section references | Read two referenced sections |
For more detail on these, see How to Fact-Check an AI Research Report Before You Rely on It and How to Check an AI-Made Spreadsheet Before You Trust the Numbers. If you use Claude Code, you can make it test its own work with a verify skill.
How do you check an answer in two minutes?
- Pick the claim that matters most, the one your decision depends on.
- Open its source and find the exact sentence or number.
- Check one calculation by hand or with a formula.
- Read the assumptions list for anything that doesn’t fit your situation.
- If any check fails, tell the AI what was wrong and ask it to recheck everything else on the same basis.
⚠️ Warning: A link isn’t proof. AI tools sometimes cite real pages that don’t say what they claim, or links that don’t work. Always open the source.
When should you check more carefully?
- Anything involving money, contracts, tax or legal obligations.
- Health or safety information, which should go to a qualified professional.
- Facts you will publish, send to customers or put your name to.
- Anything that surprises you. Surprising claims are where errors hide.
Your next step: save the “show your working” prompt above as a text snippet, and use it on your next research or number task.
Frequently asked questions
Can AI check its own answers?
It can help, for example by re-reading sources or running tests, but you should still spot-check the key claim yourself.
What if the AI can’t find a source?
Treat that claim as unconfirmed. Ask it to say so plainly rather than guess.
Does asking for working slow the AI down?
A little, but the answer is usually more careful and much faster for you to check.
Which AI tools can show sources?
Most chat assistants with web search can include links. Always open them to confirm they say what the AI claims.
The bottom line
You don’t need to trust AI or distrust it. You need to be able to check it. Ask for sources, working and assumptions, spot-check what matters, and you’ll get the speed of AI without betting your decisions on a guess.
Related posts
- Claude Helps Prove the Riemann Hypothesis: What It Means, Plus the GLM 5.5 Teaser and This Week’s AI News
- How to Fact-Check an AI Research Report Before You Rely on It
- How to Check an AI-Made Spreadsheet Before You Trust the Numbers
