Checking an answer before you rely on it
A quick, repeatable way to check AI answers, so you can trust what matters.
Before you rely on an AI answer, ask: what happens if this is wrong?
- Not much (a birthday message, recipe ideas)? A quick read is enough.
- Something matters (a deadline, a price, a health or money question)? Check it.
A 3-step check:
- Ask for sources. "Where does that come from? Give me links."
- Open them. Make sure the page exists and says the same thing. Prefer official sites like GOV.UK, the NHS, or the company's own website.
- Ask it to doubt itself. "What in your answer might be wrong or out of date?"
Red flags:
- very exact numbers or dates you didn't give it
- quotes from people or articles you can't find
- links that don't work
- advice that doesn't match where you live (for example US rules when you're in the UK)
Tip: assistants with web search switched on can show their sources next to the answer, which makes checking much quicker.
Start with the stakes
Not every answer needs checking. Ask yourself one question: what's the cost if this is wrong?
| Stakes | Examples | What to do |
|---|---|---|
| Low | Gift ideas, rewording, a recipe | Read it; use your judgement |
| Medium | General explanations, summaries | Sense-check against what you know; spot-check one or two points |
| High | Deadlines, laws, prices, health, money, anything you'll act on | Verify with an official or primary source |
The checking method
1. Turn on search, or ask for sources. Many assistants can search the web and show where each point came from. If yours can't, ask "Which sources would confirm this?"
2. Open the sources. Click through. Does the page exist? Is it a trustworthy site? Does it actually say what the assistant claimed? It's surprisingly common for the source to be real but say something slightly different.
3. Go to the primary source for anything important. That's the original: GOV.UK for benefits and tax, the NHS for health, your energy supplier's own price page, the actual text of your contract.
4. Ask for uncertainty. "Which parts of this are you least sure about?" Assistants are quite good at flagging their weak spots when asked.
5. Check the date and place. Rules and prices change, and differ between England, Scotland, Wales, Northern Ireland and other countries. Ask "Is this current, and does it apply in [your nation]?"
Quick sense-checks
- Does the answer make sense with what you already know?
- Are the numbers in the right ballpark?
- Is it suspiciously neat or specific?
- Did it answer the question you actually asked?
Use AI to help you check
You can ask a second assistant: "Here's an answer I was given. Check it for errors and tell me which claims need verifying." Different assistants make different mistakes, so a second opinion catches a lot.
Why checking is a skill worth having
AI assistants are right most of the time, which is exactly why errors are dangerous: we stop expecting them. Independent testing continues to find that even top models produce false statements on a meaningful share of factual questions, especially about niche topics, recent events and precise figures. Good checking habits let you use AI confidently rather than nervously.
Understanding search-backed answers
When an assistant searches the web, it runs queries, reads a handful of pages and writes an answer drawing on them, usually with citations. This is better than relying on memory alone, but new failure modes appear:
- Weak sources: it may cite a forum post or a content farm instead of an authority.
- Misreading: the summary may overstate or simplify what the source says.
- Mismatched citations: a citation may be attached to a sentence it doesn't support.
- Out-of-region results: US pages answering a UK question are common.
Clicking through to check the key citations deals with all four.
The SIFT approach
Fact-checkers and librarians teach a simple method, often called SIFT, that works well for AI answers:
- Stop. Notice when a claim matters and resist accepting it straight away.
- Investigate the source. Who published it? Are they an authority on this?
- Find better coverage. Search for the same claim from other trusted sources.
- Trace claims to the original. Find the original study, law or official announcement rather than a summary of it.
Primary sources in the UK
| Topic | Go-to source |
|---|---|
| Benefits, tax, driving, passports | GOV.UK |
| Health conditions and treatment | NHS website, your GP or pharmacist, NHS 111 |
| Consumer rights, housing, debt | Citizens Advice |
| Money, investments, scams | FCA register and warning list, MoneyHelper |
| Laws | legislation.gov.uk |
| Company terms and prices | The company's own website |
Calibrating trust in the assistant itself
You can ask models to express uncertainty ("rate each claim 1 to 5"). This is helpful but not perfect: models can be confidently wrong. Treat the rating as a pointer to where to check first, not a guarantee.
Cross-model checking is effective because different models are trained differently and make different mistakes. If two assistants disagree, you've found something worth checking. If they agree, you've gained some confidence, though not proof, since both may have learned the same wrong information.
Numbers deserve special care
For calculations, ask the assistant to show its working or to use its calculation tool. Check the inputs (did it read the right figures?) as well as the arithmetic. For anything financial, re-do the key sum yourself or in a spreadsheet.
Building it into habit
The aim isn't to check everything. It's to automatically pause on high-stakes claims. Over time you'll learn where your assistant is reliable for your kinds of questions, and where it isn't.
Try this prompt
For your last answer, list each factual claim, rate your confidence from 1 to 5, and give me an official source I can use to check the most important ones.
Replace the [brackets] with your own details
On haystak
Quick check · 1 of 3
Which answer most needs checking?
Next step
Speaking and showing: voice mode and the camera
Use your voice and your phone camera to get help hands-free, from cooking to fixing things.