Skip to content
← All free guides
Using AI well Multi-tool 7 min read · free

How to catch AI when it is wrong

AI sounds confident even when it is making things up. Here is how to spot the moments it is likely wrong, verify the things that matter, and know when not to use it at all.

The short answer

Treat AI like a brilliant, fast intern who never says "I am not sure": great for drafts, not the final word on facts. Verify anything with names, numbers, dates, quotes or legal/medical/money stakes — and for those, use AI to draft while a human (you) decides.

AI does not "know" facts the way a database does — it predicts the most likely next words. Most of the time that lands on the truth; sometimes it lands on a confident, well-written falsehood, often called a hallucination. The danger is not that it is wrong, it is that it is wrong in a fluent, believable voice. Knowing where it tends to slip turns it from risky to genuinely reliable.

Helpful for checking
  • ▸Gemini free — can ground answers in current Google results
  • ▸Perplexity free — shows sources alongside the answer
  • ▸NotebookLM free — answers only from documents you give it, with citations
Beginner

Know where it slips

You do not need to fact-check every sentence — you need to know which sentences are risky. The model is most likely to invent specifics: a statistic, a date, a person's title, a study, a quote, a citation, a price. Anything that sounds precise is exactly where to slow down and confirm.

  1. 1Scan the answer for hard specifics — numbers, names, dates, quotes, sources.
  2. 2Treat those as "check me" flags, not facts.
  3. 3Confirm the ones that matter with a primary source.
  4. 4Keep the reasoning and structure; just verify the specifics.

Copy-paste prompts

promptGo back through your last answer and flag every claim that a careful reader should fact-check. List them separately.
promptWhich parts of that answer are you most confident about, and which are you least confident about? Be honest.
promptGive me the source for that statistic. If you are not certain it is real, say so plainly.
promptRewrite that answer and clearly mark anything you are unsure about with [uncertain].
Intermediate

Make it check itself

Models are often better at evaluating an answer than producing one cold. Use that: have the AI critique its own output, or ask the same question a second way and watch for contradictions. An invented fact rarely survives being cross-examined — if two phrasings disagree, you have found something to verify.

  1. 1Ask the AI to find the weakest or least-supported claim in its own answer.
  2. 2Ask the same question again, worded differently, in a fresh chat.
  3. 3Compare — contradictions are red flags.
  4. 4Ground it: paste the real source and ask it to answer only from that.

Copy-paste prompts

promptAct as a fact-checker reviewing your own answer. Find the single weakest claim and tell me how to verify it.
promptAnswer this only using the document I am pasting — do not use outside knowledge, and if the answer is not in the document, say "not in the document": [paste].
promptI asked you this in another chat and got a different answer. Here are both. Which is right, and how would I confirm it?
promptRate your own answer for factual reliability out of 10 and explain what would raise the score.
promptList the assumptions behind your answer. Which ones, if wrong, would change the conclusion?
Advanced

Know when not to use AI

Judgment is the real power-user skill. There are tasks where a confident wrong answer does real harm — specific legal, medical or financial facts; anything you genuinely cannot verify; and decisions that need your accountability and values. For those, the move is not "trust the AI" or "avoid the AI" — it is to let it draft, gather and organise, while a qualified human makes the call.

  1. 1Ask: if this is wrong, who gets hurt and how badly?
  2. 2High stakes or unverifiable → use AI to draft and organise only.
  3. 3Always confirm legal, medical, tax and safety specifics with a qualified source.
  4. 4Keep the final decision — and the accountability — human.

Copy-paste prompts

promptHelp me prepare questions to ask a [lawyer / doctor / accountant] about [situation]. Do not give me the professional advice yourself — help me get ready for the expert.
promptSummarise this document so I understand it before my meeting, and list anything I should double-check with a professional: [paste].
promptExplain the general options around [topic] in plain English, and clearly note this is not professional advice for my specific case.
promptDraft this [contract clause / message] for a professional to review. Mark every spot where I must confirm a detail.
💡 A simple rule that keeps you safe: AI drafts, you decide. Let it do the fast, heavy lifting — then put your own eyes on anything that involves a fact, a number, or a real-world consequence.

Want the prompts that go further?

Drop your email + WhatsApp — I'll send the weekly playbook and the full prompt vault.

Prefer to ask first? Message me on WhatsApp · see plans

More guides like this

How to talk to AI so it gives you gold

The four things to put in every important prompt — a role, your real goal, an example of "good", and the exact format — so AI stops handing you generic mush.

Using AI well

Make AI your thinking partner, not a vending machine

Most people use AI like a vending machine — one question, one answer. The real unlock is letting it interview you, break big problems down, and pressure-test your thinking.

Using AI well

10 things your AI wishes you knew (from the AI)

A note from the assistant itself: the small things people do not realise that quietly make every answer better — written from the AI's side of the conversation.

Using AI well