---
title: "Can you rely on ChatGPT for important information?"
url: https://canyouuseai.com/questions/chatgpt-can-make-mistakes-check-important-info/
verdict: yes-with-care
verdict_label: "Yes, with care"
category: "Everyday life"
published: 2026-10-22
updated: 2026-10-22
author: "Jack Stovell"
publisher: "Adapt Progress Evolve Limited"
language: en-GB
---

# Can you rely on ChatGPT for important information?

**Verdict: Yes, with care.** Yes, with care. ChatGPT can give confident answers that are wrong, so check anything you will act on that could cost you money, health, marks or your legal position against the original source.

Author: Jack Stovell · X https://x.com/StovBuilds · Instagram https://www.instagram.com/stovbuilds · LinkedIn https://www.linkedin.com/in/jackstovell/

Short answer: treat any answer you plan to act on as a draft until you have found the same thing in the source that actually sets the rule. You can judge a recipe idea or a reworded email by eye. You cannot do that with a tax deadline, a legal right or a question about medicines.

### What the warning under the ChatGPT box actually says

The notice reads "ChatGPT can make mistakes. Check important info." It makes two points. Mistakes happen, and checking the ones that matter is your job.

OpenAI's help article, "Does ChatGPT tell the truth?", explains the first half. It says ChatGPT can produce incorrect or misleading outputs, and that it might sound confident even when it is wrong. OpenAI calls this a hallucination and gives three kinds of example:

- incorrect definitions, dates or facts
- made-up quotes, studies, citations or references to sources that do not exist
- overconfident answers to ambiguous or complex questions

The same article covers the second half. OpenAI says to use ChatGPT as a first draft and to check important information. It also says to always verify quotes, data, technical information and references to external documents.

### Why ChatGPT gets things wrong

OpenAI gives these reasons in the same article:

- **It works from patterns.** ChatGPT answers from patterns in the data it was trained on.
- **It has a knowledge cutoff.** Without tools such as search, it knows nothing about events after its training data ends.
- **Confidence is not reliability.** It may sound very sure of an answer that is wrong.
- **It cannot always reach the source.** Technical problems, paywalls or a site's own settings can stop it reading a page.
- **It can oversimplify.** It may present one view as the absolute truth, or misrepresent how settled a scientific or social debate is.

In practice, this means the tone of an answer tells you nothing. A wrong answer reads exactly like a right one. Search helps because ChatGPT can then cite web pages, and OpenAI's advice is to open those links yourself. Treat a citation as somewhere to start looking, because made-up references are on OpenAI's own list of errors.

### Which answers count as important

An answer is important if you are going to act on it and acting on it could cost you something. That covers money, your legal position, your health, a mark, a job or a deadline. It means anything where a mistake leads to a fine, a missed entitlement or a decision you cannot easily undo.

A birthday card rhyme or a polite note to a neighbour about the bins is not important. If it is wrong, you will notice, or it will not matter.

A quick test: would you be embarrassed or out of pocket if you repeated this answer to someone official and it turned out to be wrong? If so, check it.

One category catches people out. You ask for a quick summary, then quote it in a report or forward it to a colleague. Once an answer feeds into someone else's decision, it is important, however casually you asked.

## How to do it well

### A five-minute method for checking an answer

1. **Pull out the claim.** Find the one sentence you would act on, such as "You must do X by this date", "This is allowed" or "These two medicines are fine together".
2. **Ask ChatGPT where the claim comes from, and treat the reply as a lead.** Act names, section numbers and web addresses can be wrong too. Do not copy them anywhere until you have opened the real page.
3. **Go to the primary source.** That is the body that makes or enforces the rule. A blog summarising the rule does not count. The next section lists where to look.
4. **Find the matching line.** Read the actual wording. Does it say what the answer said? Does it apply to your situation, or a different one? Dates, thresholds and exceptions are where answers tend to fall apart.
5. **Decide.** If it matches, carry on. If it does not match, or you cannot find it, do not act on the answer. If the stakes are high, ask a person.

Five minutes is a target. Nobody has measured it. Some checks take one minute. If a check drags on much longer, the question was probably too complex to rely on a chatbot's answer anyway.

### Checking by task: law, tax, health, study and work

The method stays the same. Only the place you check changes.

**Law.** legislation.gov.uk publishes UK legislation. Find the Act or regulation and read the section itself. A "Changes to Legislation" box at the top of the page lists changes that have not yet been written into the text, so read it before you rely on a section. If ChatGPT gave you a section number that does not exist, treat the whole answer with suspicion. The [legal advice page](/questions/get-legal-advice/) covers where AI fits and where you need a person.

**Tax.** Use HMRC's guidance on gov.uk. It covers things like Self Assessment, tax codes and refunds. Tax answers often depend on your own circumstances, and ChatGPT only knows those as well as you described them. The [taxes page](/questions/do-my-taxes/) goes further.

**Health.** The NHS Conditions A to Z covers symptoms and how conditions are treated. In England, if you think you need medical help right now, the NHS says to use 111 online, use the NHS App or call 111. Elsewhere in the UK, the same NHS page points to NHS 24 111 in Scotland, NHS 111 Wales in Wales and nidirect in Northern Ireland. In an emergency, call 999 or go to A&E. Never use a chatbot to decide what a symptom means.

**Money.** Before you act on anything about a financial firm, use the FCA Firm Checker. It shows whether the firm is authorised and has permission for the service it is offering you. The [investing page](/questions/invest-in-stocks/) covers the rest.

**Study.** Check facts and quotes against the actual reading, and check your institution's rules on AI before you use it. See the [homework page](/questions/do-my-homework/). If you asked for a summary of a document, check every figure and quote against the original. The [PDF summary page](/questions/summarise-a-pdf/) explains why.

**Work.** It is easy to mistake a confident answer about a procedure for company policy. Check it against the real policy document, or ask whoever owns the policy.

## Rules and risks

### What the research says about warnings like this one

Two published studies use this warning in their titles.

**Kıyak and colleagues (Medical Education, 2026)** ran a randomised controlled trial with 186 fourth-year medical students. Each student worked through three clinical cases, each written so that two diagnoses were equally plausible. Each student chose a diagnosis, then got feedback labelled as coming from ChatGPT that argued for the other one. Half the students saw the warning "ChatGPT can make mistakes. Check important info" displayed prominently with the feedback. The other half did not.

The warning made no measurable difference. 15.3% of the no-warning group changed their diagnosis, against 15.9% of the warning group. The students already gave the AI's advice little weight, and the authors suggest their trust was already so low that the warning could not lower it further. They conclude that simple warnings may not be enough to make people trust AI the right amount.

**Urban and colleagues (British Journal of Educational Technology, 2025)** asked 98 university students to write using articles that contained both correct and incorrect information. Experts had written some of the articles and ChatGPT had written the others. 49 students wrote alone and 49 wrote with help from ChatGPT-4o. The students with ChatGPT included more correct information from both kinds of article. ChatGPT's help did not change how much incorrect information they took from the ChatGPT-written article. Students who were worse at judging their own knowledge included more of the incorrect ChatGPT content, and so did students with certain beliefs about knowledge. The authors conclude that these skills remain essential for avoiding AI errors. This study tested ChatGPT's help. It did not test the warning itself.

Taken together, these studies suggest the line on screen will not do the checking for you.

### What the sources do not settle

**How often ChatGPT is wrong.** OpenAI's article gives no error rate, and neither do we. Be wary of anyone quoting a neat percentage without a source you can open.

**Whether warnings work for everyone.** Both studies used students doing set tasks. In one, medical students judged clinical cases. In the other, university students wrote from articles. Neither tested members of the public asking about their own council tax or tenancy. The Kıyak trial found the warning did nothing for people who were already sceptical. It does not show what a warning does for someone who trusts ChatGPT completely.

**How much search improves things.** OpenAI says tools such as search and deep research can help ChatGPT give more up-to-date and verifiable answers, but the article gives no figures.

**Whether this method catches every error.** It will not. You can misread a primary source just as a chatbot can misstate one. Checking the source is still safer than trusting the answer alone, and when the stakes are high, the final step is to ask a qualified person.

## Sources

- [OpenAI Help Center: Does ChatGPT tell the truth?](https://help.openai.com/en/articles/8313428-does-chatgpt-tell-the-truth)
- [Kıyak, Coşkun and Budakoğlu (2026), 'ChatGPT can make mistakes' warnings fail: a randomized controlled trial, Medical Education](https://onlinelibrary.wiley.com/doi/10.1111/medu.70056)
- [Urban et al. (2025), "ChatGPT can make mistakes. Check important info." Epistemic beliefs and metacognitive accuracy in students' integration of ChatGPT content into academic writing, British Journal of Educational Technology](https://bera-journals.onlinelibrary.wiley.com/doi/10.1111/bjet.13591)
- [legislation.gov.uk: Help and guidance](https://www.legislation.gov.uk/help)
- [GOV.UK: HM Revenue & Customs](https://www.gov.uk/government/organisations/hm-revenue-customs)
- [NHS: Conditions A to Z](https://www.nhs.uk/conditions/)
- [NHS: When to use NHS 111 online or call 111](https://www.nhs.uk/nhs-services/urgent-and-emergency-care-services/when-to-use-111/)
- [FCA: FCA Firm Checker](https://www.fca.org.uk/consumers/fca-firm-checker)

## Questions people ask

### Can ChatGPT make mistakes?

Yes. OpenAI's own help article says ChatGPT can produce incorrect or misleading outputs and may sound confident even when it is wrong. Its examples include wrong dates and facts, and quotes, studies or references that do not exist.

### What does 'ChatGPT can make mistakes. Check important info' mean?

It says two things. The answer may be wrong however sure it sounds, and checking anything that matters is your job. OpenAI's help article tells users to treat ChatGPT as a first draft and to verify important information from reliable sources.

### Does ChatGPT give incorrect information, and why?

It can. OpenAI gives the reasons. It works from patterns in its training data. It knows nothing about events after its knowledge cutoff unless a tool such as search is used. It cannot always get into a website. It can oversimplify or present one view as the whole truth.

### Which answers from ChatGPT should I always check?

Check anything you will act on where a mistake could cost you money, your health, your legal position, marks or a job. Also check quotes, figures and references, and anything you are about to pass on to someone else.

### Does the warning actually stop people trusting wrong answers?

The evidence is thin. A randomised trial with 186 medical students found the warning made no measurable difference to whether they changed their diagnosis after advice labelled as coming from ChatGPT. That was one group doing one task, so it does not settle the question for everyone.
