AI Guides On
Sign in

"AI responses may include mistakes": what the warning means and what to do about it

It is a standing warning that any AI answer can be wrong, because these tools predict likely words and nothing checks the result against the facts. Treat it as a prompt to check anything that matters; for everyday questions a one-minute, three-step check is enough.

By Jack Stovell · Published

Short answer: the line is a standing caution attached to every AI answer, right or wrong. It tells you mistakes are possible, gives no clue where they are, and leaves the checking to you.

Where you see “AI responses may include mistakes”

In Google Search, the line comes with AI Overviews, the AI-written summary Google shows for some searches, and with AI Mode, where you can keep asking follow-up questions. Google’s help pages for both features carry the same words: “AI responses may include mistakes.” On the AI Overviews help page, the line links to a page about generative AI and its limitations.

Other AI chat tools say the same thing in their own words. OpenAI’s help page says ChatGPT “can produce incorrect or misleading outputs”. Google’s Gemini Apps help says Gemini “may sometimes produce inaccurate, offensive, or inappropriate information”. The line is small and easy to scroll past, but each of these tools has its own version.

What the warning is telling you

Google is direct about it. Its AI Overviews help page says they “can and will make mistakes” and asks you to “always check important info in more than one place”.

What the line leaves out matters just as much:

  • It does not say which part of an answer might be wrong.
  • It gives no score, and no sign that one answer is safer than another.
  • It does not mean a person has checked the answer.
  • It reads the same whether you asked about bin day or a tax bill.

The reason lies in how these tools work. Google’s page on generative AI says large language models “don’t gather information at all” the way Search does. Instead they “predict which words come next”. OpenAI says ChatGPT answers “based on patterns in data it was trained on”. So an answer reads smoothly because smooth text is what the prediction produces. Whether it is true is a separate question.

For the bigger picture of what these tools are and how they produce an answer, the lesson What AI chat tools are covers it in a few minutes.

The kinds of mistake to expect

Three kinds come up again and again.

Made-up details. When a tool invents an answer, it is called a hallucination. Google and OpenAI both use the word. OpenAI’s examples include “fabricated quotes, studies, citations or references to non-existent sources”. A name, a book title or a source can sound right and not exist.

Out-of-date facts. OpenAI says its models are trained on data up to a certain point and know nothing about later events unless tools such as search are used. Prices, rules, opening times and who runs what may all have changed since then.

Confident summaries of sources. Ask about an article or a report and you may get something tidy that drops a caveat or bends the meaning. OpenAI warns that the model may “oversimplify complex or nuanced issues” or “present a single perspective as absolute truth”.

News puts all three to the test. In October 2025 the European Broadcasting Union (EBU) published a study, led by the BBC, in which journalists checked more than 3,000 answers about the news from ChatGPT, Copilot, Gemini and Perplexity. The EBU reports that 45% of all the answers had at least one significant issue, 31% had serious sourcing problems and 20% contained major accuracy issues, “including hallucinated details and outdated information”.

Why a confident tone proves nothing

A wrong answer and a right answer sound the same. The tool does not hesitate when it is guessing. OpenAI puts it plainly: “Confidence isn’t reliability”, and the model “may express high confidence even in incorrect answers”.

When a person explains something clearly, that is often a fair sign they understand it. With AI tools the link breaks. A well-organised paragraph shows that the tool put likely words in a sensible order. It says nothing about whether the facts in it are right. Once you stop treating tone as evidence, checking starts to feel normal.

A one-minute check for everyday questions

For everyday questions, such as a recipe swap, what a word means or when a shop opens, three steps are enough.

  1. Pick the claim that matters. Find the one or two things that would cause trouble if they were wrong: a date, a name, a figure, a “you must”.
  2. Check it with whoever owns the answer. Go to the official page: the shop for its opening times, the government for a rule, the original article for a quote. Another AI answer does not count.
  3. Open the links. If the answer gives sources, open them and see whether they say what the answer claims. For AI Overviews, Google’s own advice is to click the links to supporting information and try other search results too. If there is nothing to open and no official source, treat the answer as a guess.

Checkpoint: you can point to one place outside the AI that agrees with the claim you care about.

The lesson Check what it tells you walks you through this with a real question and a real official page, so you practise it properly once.

If you are going to act on an answer and a mistake would cost you, slow down. Health, money and legal questions come first, along with anything you will send to someone else or put your name to.

Google’s Gemini help is clear: “Don’t rely on responses from Gemini Apps as medical, legal, financial or other professional advice.” Use the AI to understand words you do not know and to plan the questions you want to ask. Then take those questions to the official source or a qualified person, and let them have the last word. The tool will sound just as sure about a medicine dose as about bin day.

What to do next

If you are new to AI chat tools, the free course AI for beginners: start here takes you from what the tools are to your first conversation and on to checking answers.

If your answers come back vague, the prompting course can help. Start with why answers come back vague, then use fix an answer to turn a weak reply into a useful one.

Sources

Questions people ask

Is ChatGPT always right?

No. OpenAI's own help page says ChatGPT “can produce incorrect or misleading outputs” and may sound confident even when it is wrong. Check anything that matters with an official source.

Is AI prone to making mistakes?

Yes. Google says AI Overviews “can and will make mistakes”, and Gemini Apps help says Gemini may produce inaccurate information. That is why the warning line is there.

What is it called when an AI makes a mistake?

When an AI tool makes up an answer, it is usually called a hallucination. Google and OpenAI both use that word. Plain errors, such as an out-of-date fact, are usually just called mistakes.

Why does AI give wrong answers?

Google explains that large language models predict which words come next and do not gather information the way Search does. The answer reads smoothly whether or not it is true, and it can be out of date or summarise a source badly.

Should I trust Google's AI Overview?

Use it as a starting point. Google's own help page says AI Overviews can make mistakes and advises you to check important information in more than one place, click the supporting links and try other search results.