An AI chat answers in complete, confident sentences whether it is right or wrong. That is what makes it pleasant to use and risky to trust. The skill worth learning is not how to ask, it is how to check. This guide gives you a routine short enough to use every time the answer matters.
Why a good model still gets things wrong
A language model writes by predicting which words are likely to come next, based on the text it was trained on. It does not look facts up unless it has been given a search tool, and it has no built-in sense of “I do not know”. Three consequences follow:
- It fills gaps. Asked for something it has not seen, it produces something that looks like the right kind of answer: a plausible number, a plausible name, a plausible reference.
- It is out of date. Its knowledge stops at the time it was trained. Prices, laws, office holders and software versions may have changed since.
- It follows your lead. If your question assumes something false, many models will build on the assumption instead of correcting it.
Smaller models, which is what most free tiers give you, make these mistakes more often than large ones.
Answers that deserve the most doubt
- Numbers. Statistics, prices, rates, dates and any arithmetic done in the reply.
- References. Book titles, court cases, studies, web addresses and quotations. Invented references are a well-known failure.
- Local rules. Anything that differs by state or country, such as stamp duty, tax limits or licensing rules. The model may give another place’s rule, or last year’s.
- Recent events. Anything from the last year or two.
- Health, law and money. Not because models are always wrong here, but because being wrong costs the most.
By contrast, explaining a well-known idea, rewording your own text, suggesting a structure or brainstorming options carry little risk, because you can judge the result yourself.
A six-step routine
- Ask where it comes from. “What is the source for that figure?” Then open the source. If the model cannot name one, or the source does not say what the model claims, treat the fact as unverified.
- Redo the arithmetic. Use a calculator for anything that is added, multiplied or converted. It takes seconds.
- Ask it to argue against itself. “What are the weakest parts of your answer? What would a specialist disagree with?” Models are often better at finding faults than at avoiding them.
- Get a second opinion from a different model. Different models are trained differently and tend to make different mistakes. Give the second one the first one’s answer and ask it to list anything wrong or unsupported.
- Test what can be tested. Run the code. Try the formula on a case where you already know the result. Follow the instructions on a copy, not the original.
- Take decisions to a person. For a contract, a diagnosis or a large payment, the AI’s answer is preparation for a conversation with someone qualified, not a replacement for it.
Doing step four in AI MultiTool
AI MultiTool is built for the second opinion. Every AI in a session can read the whole conversation, including what the others wrote. After one AI has answered, tap a different AI in the “Who answers” row under the message box and send something like: “Check the answer above. List any claim that is wrong, out of date or unsupported, and say how sure you are.” You do not need to paste anything.
Two cautions. Agreement is not proof: two models trained on similar text can share the same mistake, so a second opinion narrows the risk without removing it. And disagreement is useful: where two models differ is exactly where you should look something up.
Signs that an answer was made up
- A very exact figure with no source, such as “73.4% of buyers”.
- A reference you cannot find when you search for its exact title.
- Detail that grows when you push. Ask the same factual question twice in new sessions: stable facts come back the same, inventions tend to change.
- An answer that agrees with everything in your question, including the parts you were unsure of.
Ask in a way that makes checking easier
Tell the model what you will do with the answer and ask it to separate what it knows from what it is guessing: “Mark anything you are not sure of.” Ask for the reasoning in steps, so that you can see which step went wrong. Give it the document and ask it to quote the line it relied on. Each of these turns one opaque answer into several small claims you can test.
This article is general information. An AI model’s output is not professional advice, and neither is this guide.