The answer on your screen reads clean and sure of itself. It might be a price range for a customer, a reply to a supplier, a rule about overtime or a paragraph for your website. And a small voice asks what happens if one detail in it is wrong and you are the one who sent it.
How to check an AI answer before you act on it is a short routine, and it fits in about ten minutes. You ask for the sources, open them, test one fact you already know, and redo every number yourself. Anything that touches law, tax, health, safety or a promise to a customer goes to the person who is responsible for it.
You do not have to stop using the tool to do this. You only have to put a short, repeatable step between the answer and the send button. Try it on the next answer, or on one you used this week.
Key Takeaways
Request each source, its date and the exact supporting passage, then open every link yourself and confirm it supports the exact wording.
Test one date, name, price or rule you already know. If it is wrong, stop trusting the answer until the rest is verified.
A citation, a confident tone, a green highlight or a second chatbot agreeing with the first is a lead to verify, not a confirmation.
Send legal, tax, medical, safety and customer-commitment claims to the responsible professional, and keep the output, the sources, the reviewer and the approved version.
Finish your work with a ten-minute check you can run alone
Pick one answer you plan to use. The routine works for a quote, an email, a calculation or a research summary, and none of it needs special software. The steps run in the order you would naturally do them.
- Mark the claims that matter: Highlight every statement that could change a decision, such as a date, a price, a rule, a name or a promise.
- Ask for the evidence: Ask the AI for the source, the publication date and the exact passage that supports each important claim.
- Open every source yourself: Confirm the page exists, is about the same subject and supports the wording in the answer, not just the topic.
- Test one fact you already know: If the date, name, statute, price or calculation is wrong, stop trusting the rest until each part is checked.
- Recompute every number: Use a calculator or spreadsheet for totals, percentages, discounts, taxes and time estimates.
- Check anything time-sensitive against the current official page: Fees, deadlines and rules change, and a summary can be out of date.
- Confirm quotes word for word: Match the wording, the speaker and the date to the original.
- Label each claim: Verified, needs correction, or unresolved. Do not send or act on an unresolved claim that carries risk.
OpenAI's help page tells users to "verify important information from reliable sources," and steps 2 and 3 are how you do that in practice. Reject links that are broken, undated, unrelated, or secondhand when the original is available.
| Minutes | What to do | What it catches |
|---|---|---|
| 1 | Copy the answer and highlight every statement that affects a decision | Claims you would otherwise skim past |
| 2 to 4 | Open every citation and reject any that is broken, unrelated, undated or unable to support the exact claim | Invented or mismatched references |
| 5 to 6 | Check one date, name, rule or other fact you can confirm independently | A wrong answer hiding behind a confident tone |
| 7 to 8 | Recalculate every number, percentage, total, tax, discount or time estimate | Arithmetic and unit errors |
| 9 | Compare high-risk claims with the regulator, agency, contract, professional body or original study | Rules that are wrong or out of date |
| 10 | Label each claim verified, needs correction or unresolved | Anything unresolved that should not go out |
If you have the time to do only one thing, open the sources. A claim that does not appear in the page it cites is the clearest sign that the answer needs rework.
Get real confirmation from what each AI tool tells you to do
The companies behind these tools tell you to check, and they say it in their own help pages. Knowing exactly what each one offers saves you from trusting a feature that was never meant to settle the question.
OpenAI says ChatGPT can produce incorrect facts, fabricated quotes, studies, citations and references, and that citations can be "incomplete, outdated, or incorrect." Its guidance is to open the cited sources directly, and to use ChatGPT search or Deep Research when the question calls for it. The same page is OpenAI's explanation of whether ChatGPT tells the truth.
Google's Gemini has a Sources panel and a "Double-check response" feature. Google says green highlights mean Search found content likely similar to a statement, and orange highlights mean it found conflicting or insufficient support, and that related links do not guarantee correctness.
Anthropic says Claude's web search responses include citations and source links so you can verify them. Microsoft says Copilot responses can sound convincing while being incomplete or inaccurate, and recommends reviewing citations and double-checking facts before you act.
All four point to the same step: the links are there for you to open.
Stop a confident tone from standing in for proof
A polished answer feels finished. Four things look like confirmation and are not.
A citation tells you where to look, and the page it names may be irrelevant, outdated, broken or read wrongly. A confident tone says nothing about accuracy. A green highlight in Gemini means Search found similar content, not that the statement is correct legally, scientifically or commercially. And a second chatbot that agrees with the first can repeat the same underlying error.
Treat each of these as a lead and take it to the original: the regulator, the contract, the statute, the manufacturer's page, the study itself. That is a check you can finish, and it gives you something you can point to if the claim is ever questioned.
Hand risky claims to the person who can stand behind them
Some answers can cost real money or hurt someone if they are wrong, and for those the final reviewer should not be you or the chatbot. Match the amount of review to the damage a mistake could do.
These uses call for a responsible professional or the official authority before anything is sent, published or acted on:
- Legal questions: Contract wording, employment rules and regulatory obligations.
- Tax and money: Calculations, filings, deductions, payroll and financial recommendations.
- Health and safety: Medical, health, safety or product-risk advice.
- Customer commitments: Prices, discounts, delivery dates, warranties, service limits and quotes.
- Advertising claims: Testimonials, comparisons and performance promises.
- Security and privacy: Instructions about protecting or handling data.
The FTC says advertising claims need a reasonable basis before they run, and that health or safety claims generally require competent and reliable scientific evidence, in its advertising FAQs for small business. NIST's Generative AI Profile (July 2024) recommends managing AI risk according to the intended use and the potential impact. Researchers who tested legal research tools found that even specialized legal AI tools can hallucinate citations, so a legal-sounding answer is still a draft for a lawyer.
Low-risk work needs far less. Brainstorming, tone changes, formatting and rough outlines usually get a quick read. A short factual email may need only a date check.
Keep a staff rule and a record that protect you later
If anyone else on your team uses these tools, the check should not live only in your head. A staff rule makes the step the same for everyone, and a record shows what was verified if a customer, a client or a regulator ever asks.
A rule that covers it in a few lines: AI may help draft, summarize, calculate or research, but no employee may publish, send, quote, price, approve or act on an AI-generated claim until the responsible person verifies the important facts against the original source. Every legal, tax, medical, safety, financial, pricing and customer-commitment claim gets verified. If a source cannot be opened or an answer cannot be confirmed, stop and escalate.
For higher-risk work, keep four things: the AI output, the sources checked, the reviewer's name and the final approved version. NIST's AI Risk Management Framework supports documenting intended use, knowledge limits, human oversight and verification. If you want the wider written policy, the post on an AI use policy for a small business covers what to put in it.
Know when the check needs more than one careful owner
You can run this routine alone, on the occasional answer, and it works. No consultant can promise that a general-purpose AI model is never wrong, and anyone who offers that is selling something nobody can deliver.
Help becomes worth paying for when the checking has to happen the same way across employees, approved tools, privacy rules, escalation paths and customer channels. That calls for a written workflow, named reviewers and records, not a one-time fix to one answer. Sensitive data, connected tools and customer-facing replies all raise the stakes, because a mistake there repeats at scale.
Two neighboring problems are different jobs. If the AI is answering your customers directly, read about what to do when a chatbot gives customers wrong answers. If the wrong answer is Google's own summary of your business, see when Google's AI Overview is wrong about your business.
Can you use this AI answer yet?
Pick an answer to begin.
1. An AI answer gives you a link as its source. What should you do first?
2. You test one fact in an AI answer you already know, and it is wrong. What follows?
3. The answer includes a customer quote with a price and a delivery date. Who should confirm it?
Frequently Asked Questions About Checking AI Answers
How can you verify an AI answer?
Ask for the source, date and supporting passage, open each source, and confirm it supports the exact claim. Test one fact you already know, recompute every number, and compare time-sensitive or high-risk claims with the current official source.
How do you fact check ChatGPT?
Use the same routine. OpenAI says ChatGPT can produce incorrect facts and fabricated citations, so open every cited source yourself and check what the page says against the answer.
Can ChatGPT make up sources?
Yes. OpenAI says ChatGPT can fabricate citations, studies, references and quotes, and that citations can be incomplete, outdated or incorrect. A link you have not opened is not yet evidence.
Should I ask a second chatbot to check the first one's answer?
It can give you another lead, but agreement between chatbots is not independent proof, because both can repeat the same error. Confirm with the original source.
Which AI tool is best for fact checking?
Choose by how easily you can open and check what it cites. Gemini's Double-check, ChatGPT search, Claude's web search and Copilot's citations each give you links to follow, and each still needs your review before you act.
Is AI a good tool for fact checking?
It is useful for finding leads, such as possible sources and related facts. The final check belongs to you or the responsible professional, using the original source.
Moving Forward
Checking an AI answer takes a short, repeatable routine: ask for sources, dates and passages, open every source, test one fact you know, recompute the numbers, and confirm time-sensitive claims against the current official page. A citation, a confident tone, a green highlight or a second chatbot is a lead, not proof. Legal, tax, medical, safety and customer-commitment claims go to the person responsible for them.
Do this a few times and it becomes a habit that takes minutes. You send fewer corrected emails, give fewer wrong quotes, and have a record showing who checked what when someone asks.
If you want this to work across your whole team and your tools, Web Leveling can build it with you: which uses are low risk, which need approval, which sources count, and how reviews get recorded. Our AI consulting work covers that written workflow in plain terms. We work with small and medium businesses across the country and overseas. Tell us how your team uses AI today, and we will help you set up the check.
Terms
Words used in this post
Tap a term to see what it means.
Citation. A link or reference an AI tool gives for a claim; it is a lead to open and read, not proof.
Fabricated citation. A source, study, quote or reference that the AI produced but that does not exist or does not say what the answer claims.
Double-check response. A Gemini feature that uses Google Search to flag statements as similar to or different from content Search found.
Verification. Matching a claim to the original, authoritative source and confirming it says what the answer says.
Time-sensitive claim. A fact that can change, such as a fee, deadline, rule or price, and needs the current official page.
Reviewer. The named person responsible for approving an AI-assisted claim before it is sent, published or acted on.
