Why a wrong answer sounds exactly like a right one
A chat AI writes its answer one word at a time, choosing words that are likely to follow the ones before. Unless it has a search tool and uses it, it is not looking anything up as it writes. A number that merely sounds likely comes out in the same steady voice as a correct one, so nothing in the tone or the polish tells you which you got. A false statement made this way is called a hallucination.
That is why reading an answer more closely does not help much. A person who is unsure hesitates or says so. An AI answer stays just as fluent when it is unreliable. The guide to reducing AI hallucinations explains the cause in more detail.
It happens often enough to plan for. In a study coordinated by the European Broadcasting Union and led by the BBC, published on October 22, 2025, journalists at 22 public service media organizations in 18 countries judged more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity on how they handled the news. The EBU’s summary reports that “45% of all AI answers had at least one significant issue” and that 20% “contained major accuracy issues, including hallucinated details and outdated information”. The study looked only at news, so the figures do not carry over to every kind of question.
Seven signs worth a second look
None of these signs proves an answer is wrong. Each marks a kind of detail that an AI can write without having it right, and that you cannot judge by reading. When one appears in an answer you plan to use, that detail goes on your list to check.
- A precise number with no source. A percentage or a price that arrives without saying where it comes from. An exact-looking figure is easy for an AI to write and impossible to judge by eye.
- A quote. Words in quotation marks credited to a person or a document. The wording may have been changed, or the line may never have been said.
- A reference or a link. A study, a book, a court case or a web address. An AI can write a reference with every part in place that points to nothing, as the next section shows.
- A date. The day something happened, or the year a rule took effect. A date looks exact, and one wrong digit changes what it means.
- A sweeping claim with nobody behind it. “Studies show” and “experts agree” name no study and no expert, so there is nothing to open. “The first” and “the largest” can be true of only one thing, and the AI may have picked the wrong one.
- Anything recent. An AI learns from text collected up to a certain date. Unless it searches the web, it may answer from old information as if it were current. Prices, laws, job titles, product details and news all go out of date.
- An answer that agrees with what you suggested. If you asked “Isn’t it true that…?” and it said yes, you have learned less than it seems, because these tools lean toward agreeing with the person they are talking to. Ask again without your own view in the question.
Try the signs on a short answer. Everything in the example below is invented for this guide.
“Morning is usually the best time to water tomatoes.”
Not quite. This is general advice. If it turns out to be wrong, the cost is small and you would see it in your own garden.
“Water at the base of the plant so the leaves stay dry.”
Not quite. Advice like this is easy to try and easy to judge for yourself. Nothing in it claims to come from a source.
“A 2021 study by the Acme Garden Institute found that morning watering raised yields by 23%.”
Yes. A named study with an exact percentage: two of the seven signs in one sentence. Search for the study before you repeat the figure. If you cannot find it, leave the sentence out.
“In very hot weather, check the soil again in the late afternoon.”
Not quite. A habit you can test with a finger in the soil. It makes no claim that you would need a source for.
An invented example. There is no Acme Garden Institute, and the 23% figure was made up for this quiz.
A made-up reference looks like a real one
References deserve their own warning, because the invented ones are so well formed. In a paper published in JMIR Mental Health on November 12, 2025, researchers asked GPT-4o for six literature reviews and checked every citation it gave. The paper reports that “GPT-4o generated 176 citations; 35 (19.9%) were fabricated” and that among the 141 real citations, 64 (45.4%) contained errors (Linardon and colleagues, 2025).
The cost can be high. Damien Charlotin’s AI Hallucination Cases database tracks legal decisions in cases where generative AI produced hallucinated content, typically fake citations. It listed 2,292 cases when we opened it on November 13, 2026.
Here are two references of the kind an AI might give. Both are invented for this guide, and neither points to a real publication.
Lee, D., & Ortiz, P. (2019). Morning routines and sustained attention in office teams. Journal of Example Studies, 12(3), 45-61. https://doi.org/10.0000/example.2019.045
Acme Institute. (2022). Workplace focus report.
The authors’ names
Not quite. Plausible names are easy for an AI to write. A name on its own shows nothing.
The journal name and the page numbers
Not quite. These give a reference its familiar shape, which an AI reproduces well. They look the same on an invented reference as on a real one.
The DOI link at the end
Not quite. A DOI is the ID number a publisher gives an article. An AI can write one that is well formed and leads nowhere, so you learn nothing until you open it.
None of them. Only searching for the title and opening the source can tell you.
Yes. Nothing on the page separates an invented reference from a real one. Search for the exact title in quotation marks, and if the source exists, find the passage that supports the claim.
Reference B looks vaguer, so many readers would doubt it first. Both are equally invented, and the detail in reference A is no evidence.
What to do when a reference has already failed is covered in AI made up a source?, which walks through checking every citation in a draft.
The five-minute check
Checking is the habit people skip. Anthropic studied 9,830 conversations with its Claude assistant from one week in January 2026 and found that 8.7% showed the behavior it labels “checks facts and claims that matter”, the least common of the 11 behaviors it measured (AI Fluency Index, February 23, 2026). The report notes that people may also check in ways a conversation does not show. The check below takes about five minutes for a short answer.
- 0–1 minPick the details that matterOnly what you will repeat or act on.
- 1–3 minLook each one upAt a source that would know.
- 3–4 minAsk the AI for its sourceThen open it and find the sentence.
- 4–5 minAsk again in a new chatCompare the two answers.
Timed for a short answer with a few details that matter. A longer answer takes longer, and the order holds.
1. Pick the details that matter
You do not need to check a whole answer. Go through it once and mark the details you will repeat to someone or act on, using the seven signs as your guide. The Fact Check List does the marking for you: paste the answer and it pulls out every number, quote, source and date, riskiest first, with a search link for each. Our longer guide on how to fact-check an AI answer covers this step in depth.
2. Look each one up at a source that would know
Go to wherever the fact lives, not to another summary of it.
- A price: the seller’s own page.
- A rule or a deadline: the office that sets it.
- A study: the study itself, not an article about it.
- Who holds a job: the organization’s own website.
A second AI answer is not a source. If you cannot find the fact in a couple of minutes, treat it as unconfirmed and leave it out.
3. Ask the AI for its source, and open it
Ask: “Where does that figure come from? Name the source and give me the link.” Then open the link yourself and find the sentence that says it. A real page that does not contain the claim counts as no source. If the AI cannot name one, the detail stays unconfirmed.
4. Ask again in a new chat and compare
Start a fresh conversation and ask the same question in neutral words, without mentioning the first answer. If the two answers disagree on a detail, that detail is unreliable until you have looked it up. If they agree, it can still be wrong, because the same AI can repeat the same mistake. Why AI answers change explains where the variation comes from.
Two more checks compare an answer with something you already have. Summary Check compares an AI summary with the original text and marks any number, date, name or quote that is not in the original. Instruction Check compares an answer with your prompt and lists the details the answer added that you never gave it.
A prompt that does the looking up
If your AI can search the web, it can do the searching for several claims at once. Copy this prompt and paste your claims where the blank is.
Check the claims listed below, one at a time, by searching the web. I plan to repeat them to other people, so I need to see where each one comes from.
Claims:
"""
{{claims_to_check}}
"""
For each claim, give me:
1. The claim, word for word.
2. One verdict from this list: Supported, Partly supported, Contradicted, Could not confirm.
3. The source you opened, with its date and the link.
4. The sentence in that source that supports or contradicts the claim, in quotation marks.
Use only pages you opened in this search. Prefer the original source, such as the study itself or the organization's own page, over an article about it. Do not use your memory as a source, and do not add new claims.
If you cannot find a page that states the claim, write "Could not confirm" and say what you searched for. That answer is more useful to me than a guess.
Format: a numbered list, one claim per number, with no introduction and no summary at the end. Keep each entry to 5 lines or fewer, and clear enough for me to check against the source in a minute.
Then open every source it names. The prompt saves you the searching. The reading is still yours, because a search tool can attach a claim to a page that does not say it.
When to stop checking and ask a person
Checking has a limit. Some answers should not rest on your own five minutes, however careful: anything about your health, your money, your legal position or someone’s safety.
For those, use the AI to prepare. It can explain a term in a letter or help you list the questions to ask. The answer you act on should come from a person who can look at your situation and is accountable for what they tell you: a doctor, a pharmacist, a lawyer, an accountant, or the office handling your case.
If you are new to these tools, How to use ChatGPT for beginners covers the first 20 minutes, including what not to paste.
Key takeaways
- An AI answer sounds equally sure whether it is right or wrong, so tone tells you nothing.
- Look twice at numbers without a source, quotes, references and links, dates, sweeping claims, anything recent, and answers that agree with you.
- A made-up reference looks exactly like a real one. Search for the title and open the source.
- The five-minute check: pick the details that matter, look each one up, ask for the source and open it, then ask again in a new chat.
- A second AI answer is a hint, never a source.
- For health, money, legal questions or safety, ask a person who is accountable.
Questions
How often is ChatGPT wrong?
No single figure applies, because it depends on the question. In a study coordinated by the European Broadcasting Union and led by the BBC, published in October 2025, 45% of AI answers had at least one significant issue. That study covered only news. Whatever the topic, check the details you plan to rely on.
Does ChatGPT know when it is wrong?
You cannot count on it. The answer is written the same way whether its details are right or wrong, and it comes with no reliable signal of doubt. Asking “are you sure?” gets you another answer written the same way, which settles nothing. Ask where a detail comes from, then open that source yourself.
Can I ask ChatGPT to check its own answer?
It can help, as long as you do the opening and the reading. If it can search the web, ask it to find a source for each claim and give the link, or to say it could not confirm the claim. Then open every link. Without search, it is checking its own memory against itself, which catches little.
Is Claude or Gemini more accurate than ChatGPT?
No assistant is right every time, and rankings shift with each release. The EBU and BBC study of October 2025 “identified multiple systemic issues across four leading AI tools”: ChatGPT, Copilot, Gemini and Perplexity. Pick the assistant that suits your tasks, then check its answers the same way you would check any other.
How do I check a reference that ChatGPT gave me?
Search for the exact title in quotation marks instead of clicking the link the AI gave. If nothing matches, treat the reference as invented. If it exists, check that the authors and the year match, then find the passage that supports the claim. A real source that does not say what the answer claims is no better than an invented one.
Prompts to use with this
- Write a team chat message that gets a quick, clear answerYou need a quick answer from a busy colleague on Slack or Teams and want a message they can reply to in a few seconds from their phone.