Skip to content

Guides

When AI is wrong: how to notice and what to do next

Every AI model makes mistakes, and it rarely sounds unsure while making them. Knowing the common kinds of errors and a few warning signs lets you keep the useful answers and catch the bad ones early.

5 min readUpdated

Language models are good enough that it is easy to forget they make mistakes. GPT writes in a calm, confident voice whether it is right or not, and a wrong answer often looks exactly like a right one. That does not make it useless; it means you need a sense of where errors are likely and a few quick ways to check. This guide covers the common kinds of AI mistakes, the warning signs in an answer, prompts that help you check, and the cases where you should not rely on an AI chat at all. It applies to GBT Chat and to every other AI assistant.

How a smooth answer ends up wrong

A language model does not look facts up. It writes an answer piece by piece, choosing what is most likely to come next given your question and everything it learned in training. Most of the time, the likely continuation is also the correct one. But when the model has seen little about a topic, or the question is unusual, the most likely-sounding answer can be invented. The style stays just as smooth either way, which is why tone tells you nothing about accuracy.

Our chat adds two more limits to keep in mind. It cannot search the internet or open links, and its knowledge stops at a training cut-off. According to OpenAI, GPT-6 Luna learned from data up to 18 May 2026, and GPT-6.1 Terra up to 1 June 2026. Anything after those dates is unknown to the models unless you paste it in.

The five kinds of mistakes to expect

  • Invented specifics. Names of studies, book titles, quotes, court cases, statistics and web addresses that sound real but are not. This is the most common serious error.
  • Out-of-date facts. Prices, laws, software versions, who holds a job, what a company sells. These were true once and may have changed after the cut-off.
  • Calculation slips. Long sums, percentages, unit conversions and dates, especially when several steps are chained.
  • Misread questions. The answer is correct for a slightly different question than the one you meant, because your prompt was ambiguous.
  • Overconfident advice. A general rule stated as if it applies to your situation, when in fact your country, your contract or your health makes it different.

Warning signs in an answer

You do not need to check every sentence. Look harder at these:

  • Exact numbers, dates and percentages, especially when you did not provide them.
  • A named source, author or quote you have never heard of.
  • Anything that sounds recent: "currently", "the latest", "this year".
  • A neat answer to a question you know is contested or complicated.
  • Legal, medical or financial statements phrased as certainties.
  • Details that contradict something you said earlier in the chat.

Lines with none of these, such as a rewording of your own email or an explanation of a well-established idea, are much safer. Spend your checking time where the risk is.

Prompts that make checking easier

You can ask the model to help you check its own work. It will not catch everything, but it points you to the weak spots:

Weak:
Are you sure?

Strong:
List every factual claim in your last answer that I should verify
before relying on it. For each one, say how confident you are
(high, medium, low) and what kind of source would confirm it.

"Are you sure?" often makes a model either apologise and change a correct answer, or repeat a wrong one with more conviction. The strong version asks for something concrete: a list of claims and how to check them. Open this checking prompt after any answer you plan to use.

Other useful lines:

  • "Show your working step by step," for any calculation.
  • "If you don't know, say so instead of guessing," added to the original question.
  • "What would change this answer if I lived in [country]?"
  • "Give me the strongest argument against what you just said."

Use Think for logic and numbers

Many maths and logic errors come from answering too fast. Pressing Think in the composer hands the message to GPT-6.1 Terra, a model that reasons for a moment before writing anything. That helps with multi-step calculations, scheduling puzzles and questions with several conditions. It does not help with invented facts or out-of-date information, because the problem there is missing knowledge, not hasty reasoning. You can also regenerate an answer and compare versions: if two attempts disagree on a number, at least one is wrong, and that is your cue to check. The Luna vs Terra comparison explains when the extra thinking pays off.

A quick routine for answers you will act on

  1. Mark the risky lines. Names, numbers, dates, quotes, anything recent.
  2. Ask the model to list its claims with the prompt above.
  3. Check two or three of the most important ones in a source you trust: an official website, a recognised reference, the original document.
  4. Correct the chat. Tell the model what was wrong and ask it to revise only that part.
  5. Keep the source, not the AI, if you need to cite something.

For a casual question this takes a minute. For anything that affects money, health, legal matters, work you hand in or something you will publish, it is the minimum.

Give the model the facts instead of asking for them. If you have the real document, price list or rule, paste it into the chat and ask the model to work from that text only. Errors drop sharply when the model is reorganising your material rather than recalling its own.

When not to rely on AI at all

Some situations need a person or an official source from the start:

  • Emergencies and urgent health symptoms. Call your local emergency number or a doctor.
  • Medical, legal and financial decisions. Our chat gives only general information. Use it to understand terms and prepare questions for a professional.
  • Breaking news and anything after the cut-off. The models cannot see it.
  • Exact quotes and citations for anything you publish or submit.
  • Safety-critical instructions such as electrical work, medication doses or chemicals.

Teaching others about this is worthwhile too. Our guide for parents suggests a simple exercise to show children how AI can slip, and the reading and research guide covers how to avoid invented sources in study work. When you need a summary of a text you trust, the summarizer works from what you paste rather than from memory, which is the safer way round.

FAQ

Why does the AI make things up?

Because it generates likely text rather than retrieving stored facts. When it lacks good information, a plausible invention can be the most likely-sounding continuation. Newer models do this less often, but none are free of it.

Is GPT-6.1 Terra more accurate than GPT-6 Luna?

It is better at reasoning through multi-step problems, so it makes fewer logic and calculation slips. Both can still invent specifics or miss recent changes. Read more on the GPT-6.1 Terra page.

Can the chat check facts online for me?

No. It does not have web access, so it cannot verify anything against current sources. Check important facts yourself on trusted websites.

What should I do if an answer could cause harm?

Do not act on it. Check with a qualified person. If you think an answer from our chat is dangerous, tell us through the contact form.

Your question, answered now

No account, no card, nothing to download. Type what you want to know and the first words of the answer appear in a second or two.

Start a chat