How to check an AI answer before you use it

Artificial intelligence can help you research, summarise and organise information. It can also give you a wrong answer with impressive confidence.

The problem is not that it makes mistakes. The problem is that a wrong answer can look just as tidy, clear and convincing as a right one. That is why, before you use it to publish, make a decision or hand something to a client, you need to check what matters.

This guide shows you how to do it without turning every question into a three-day investigation.

First: decide what needs checking

Not every answer needs the same level of review.

If you ask an AI for ten ideas to name a folder, you can pick the one you like and get on with your life. If you ask it to summarise a contract, work out a budget or cite a study, the margin for error matters much more.

Start by splitting the answer into three types of content:

1. Ideas and proposals: suggestions you can judge by usefulness, tone or creativity.
2. Checkable claims: dates, figures, names, quotes, prices or product features.
3. Risky decisions: recommendations that can affect money, health, privacy, security or legal commitments.

Ideas are reviewed with judgement. Claims are checked against sources. Risky decisions also need review by the right person or professional.

Step 1. Ask it to separate facts from assumptions

An answer can mix verified information with guesses without telling you. Before you look for sources, ask the tool to identify what it knows, what it is interpreting and what information it is missing.

You can use this request:

Review your previous answer. Separate the checkable claims, the inferences and the recommendations. Say which information you cannot confirm and which primary source would help verify each important claim. Do not invent links or quotes.

This does not automatically make the answer true. It gives you a clearer list of things to check.

Step 2. Find the original source

For each important claim, ask yourself: who knows this first-hand?

• A price or a feature: the product's official page.
• A regulation: the text published by the relevant authority.
• A study: the original research and its methodology.
• A company figure: its official report or announcement.
• A quote: the interview, document or publication where it appears.

If an AI gives you a link, open it. Check that it exists, that it is about the topic and that it really supports what the answer says. A real source can be misread; a convincing link can also lead to a page that says nothing of the sort.

When the fact can change (prices, plan limits or features, for example), also check the date and the conditions that apply in your country.

Step 3. Check the numbers separately

Numbers deserve their own review. AI can misread a table, mix up units or get a calculation wrong.

For example, if a tool costs 20 dollars a month, a simple estimate for three users over twelve months would be:

20 × 3 × 12 = 720 dollars

But that calculation only holds if the price really is per user and per month. You would still need to check taxes, discounts, limits and any extra costs.

The calculator can get the multiplication right and still answer a question built on wrong assumptions. Check both things: the operation and the data that went into it.

Step 4. Look at what is missing

Sometimes an answer is not false: it is incomplete in a way that changes the decision.

If an AI recommends a free tool, for example, ask yourself:

• Does it have usage limits?
• Is the feature I need included?
• Can I export my files?
• What happens to the data I upload?
• Does the price change when more than one person uses it?

Ask it to lay out the conditions and exceptions. Then check the ones that matter most against the official source.

Step 5. Test the answer on a small scale

When AI proposes a process, a formula, some code or an automation, reading it is not always enough. You have to test it.

Start with a small example and data you can safely share:

1. Define the result you expect.
2. Run the proposal on a simple case.
3. Compare the result with what you expected.
4. Try a tricky case: an empty field, an ambiguous question or a file in a different format.
5. Fix it before you use it in a real process.

For example, if an AI helps you sort customer messages, test it first with a few made-up messages. Check whether it sorts the clear ones correctly and what it does with the ambiguous ones. A useful automation also needs to know when it is not sure and should ask for review.

A quick checklist before you publish or decide

Before you accept an answer, check:

• Do I know which parts are facts and which are suggestions?
• Did I open the important sources, and do they support what is claimed?
• Are the prices, dates and features still current?
• Did I check the calculations and their assumptions?
• Does the answer mention the relevant limits and exceptions?
• Can I test the result with a small example?
• Does the type of decision need specialist human review?

You do not need the same level of detail for every question. But if an answer is going to leave your chat and affect other people, it deserves a review in proportion to its consequences.

Prompt to review an answer

I want to check the following answer before I use it.

1. Identify the claims that can be verified.
2. Separate facts from inferences and recommendations.
3. Flag figures, dates, quotes or prices that need checking.
4. Point out information that could be incomplete or out of date.
5. For each important claim, suggest a primary source where it can be checked.
6. If you cannot confirm something, say so clearly.
7. Do not invent links, quotes or test results.

Answer to review:
[PASTE THE ANSWER HERE]

After using it, do the part the AI cannot do for you: open the sources, review the context and decide whether the answer works for your case.

Related resource: Loki's template for fact-checking claims complements this guide.

Keep reading

• The end of Claude's "amnesia": how to teach it your tone, format and rules with Skills, connectors and plugins so you stop repeating them in every chat.
• Skills for coding agents: how to stop repeating your project rules to your agent, with a real skill that reviews drafts.
• Reverse engineering with AI and REA: how to understand a feature in software you are entitled to analyse and build your own version, step by verified step.