ChatGPT invents facts, citations, and quotes while sounding authoritative, so you can’t take its output at face value for anything that matters. This five-step routine catches most mistakes fast: verify sources exist, check the specifics against a primary source, prompt it to critique itself, cross-check with search or a second model, and use a 30-second risk test to decide how hard to verify.
Why ChatGPT sounds right even when it’s wrong
In 2023, a New York lawyer used ChatGPT to help write a legal brief. It cited half a dozen past cases, complete with names, quotes, and citations. There was just one problem: several of the cases did not exist. ChatGPT had invented them. The judge was not amused, and the lawyers were fined $5,000. [1]
What makes this story useful isn’t the schadenfreude. It’s the lesson underneath it. The fake cases looked completely real. They had the right format, plausible names, confident language. That’s the whole problem with AI mistakes: they don’t come with a warning label.
Here is what no one explains clearly enough. ChatGPT does not “know” facts the way a database does. It predicts the most likely next words based on patterns in its training data. Most of the time those patterns line up with reality. Sometimes they don’t, and you get a fluent, confident, totally fabricated answer. The industry word for this is “hallucination,” which is a polite way of saying the model made something up.
You can’t switch this off. But you can build a habit that catches it. That habit is what this guide is about. (For the deeper toolkit on reducing these errors in the first place, see our anti-hallucination toolkit.)
Step 1: Make it show its sources, then check they exist
The fastest way to catch a fabricated claim is to ask where it came from. After any factual answer, send a simple follow-up: “List the specific sources for each of those claims, with links.” Then, and this is the part people skip, actually click the links.
You’re looking for two failure modes. The first is dead or invented URLs: links that go nowhere, or to a page that has nothing to do with the claim. The second is the “real source, wrong fact” trick, where ChatGPT cites a genuine report but attributes a number to it that the report never mentions.
Be aware of one quirk. If ChatGPT isn’t actively browsing the web, it can invent citations from memory, and they’ll look perfect. If your version has web browsing or search built in, turn it on for anything factual, because a model that can check a live page is far less likely to guess.
If a citation has no link, or the link is to a homepage rather than a specific page, treat the claim as unverified until you’ve confirmed it yourself. A real source points to a real page.
Step 2: Verify the specifics against a primary source
Fabrications hide in the details. The shape of an answer is usually fine; it’s the exact name, number, date, or quote that goes wrong. So those are exactly what you check.
Pull out every specific claim and confirm it at the original source. A statistic should trace back to the organisation that published it (Gartner, McKinsey, a government body, a company’s own blog), not to a third blog quoting a second blog quoting nobody. A quote should appear, word for word, somewhere you can actually find it. A date should match the record.
This sounds tedious. In practice it takes a couple of minutes, because you’re not re-checking everything. You’re checking the three or four hard facts the answer rests on. If those hold up, your confidence in the rest goes up. If even one is invented, that’s your signal to slow right down.
Step 3: Ask it to find its own mistakes
This feels almost too simple, but it works more often than you’d expect. Once ChatGPT has given you an answer, push back on it directly. The single most useful line: “Review your answer and flag anything you’re not fully certain is accurate, and explain why.”
Giving the model explicit permission to admit uncertainty changes its behaviour. Instead of defending its first draft, it’ll often surface the exact spots where it was guessing: “I’m confident about the general approach, but I’m not certain the 2026 figure is correct.” That’s gold. It tells you precisely where to aim your verification.
A close cousin of this is the “what would make this wrong” prompt. Ask: “What assumptions did you make, and under what conditions would this answer be wrong?” You’re using the AI’s fluency against its own overconfidence. It’s a trick I teach in every workshop, and it pairs well with the structure in our guide to writing better AI prompts.
Review your previous answer. List every factual claim, rate your confidence in each as high, medium, or low, and flag anything I should independently verify before relying on it.
Step 4: Cross-check with a second tool
One model’s blind spot is often another’s strength. When an answer really matters, run the same question through a second source and look for where they disagree. That disagreement is a map of exactly what to investigate.
You’ve got a few cheap options. Run it through a plain web search and see if the top reputable results agree. Ask a different AI (if you used ChatGPT, try Claude or Gemini) the same question and compare. Or use a tool with live search built in so it’s citing current pages, not its training memory.
When two independent sources land on the same answer, your confidence should rise. When they split, you’ve found the soft spot. Don’t average them and move on. Go and find out which one is right. This is also a handy skill in reverse, when you’re trying to tell whether a piece of text was AI-generated in the first place, which we cover in how to spot AI-generated writing.
Step 5: Apply the 30-second risk test
Let’s be honest: you’re not going to forensically fact-check every single thing ChatGPT tells you. Nor should you. The skill is knowing when it matters. Before you act on an answer, ask one question: what happens if this is wrong?
If the answer is “nothing much” (brainstorming ideas, drafting a rough email, summarising your own notes), trust it and move on. The stakes are low and you’ll catch errors naturally.
If the answer is “that could be embarrassing or expensive” (anything going to a client, anything with a statistic, legal or medical or financial claims, anything published under your name), verify hard. Run all five steps. This is the difference between using AI well and ending up like that lawyer. The tool didn’t fail him. The missing 30-second risk check did.
Bonus: prompts that prevent mistakes upfront
Fact-checking catches errors after the fact. Even better is reducing them before they happen. A few prompt habits make a real difference:
- Give permission to say “I don’t know.” Add to any prompt: “If you’re not certain, say so rather than guessing.” This one line cuts a surprising amount of invented content.
- Ask for sources up front. “Answer this and cite a specific, real source for each factual claim” makes the model lean on things it can actually point to.
- Constrain it to what you provide. For summaries, paste the source text and say “Only use the information in the text above. If it’s not there, say so.” This stops it filling gaps with guesses.
Setting these as defaults saves repeating yourself. You can bake them into your account so every chat starts safer, which is exactly what our walkthrough of ChatGPT custom instructions shows you how to do.
What to do this week
Pick one task you already use ChatGPT for and run the full routine once. Get your answer, ask it to list and rate its own claims, check the two or three hardest facts at a real source, and cross-check anything that surprised you.
Then set one default. Add “If you’re not certain, say so rather than guessing” to your custom instructions so it applies automatically. It’s the single highest-leverage change you can make.
AI is an extraordinary thinking partner. It’s just not a reliable narrator. Once you stop treating its output as finished truth and start treating it as a confident first draft that needs a quick check, you get all of the speed with almost none of the risk. That mindset, more than any single prompt, is what separates people who use AI well from people who get burned by it.
Frequently asked questions
Does ChatGPT make up facts?
Yes. ChatGPT predicts likely text rather than retrieving verified facts, so it sometimes produces fluent, confident answers that are simply wrong. These are called hallucinations, and they include invented statistics, fake citations, and misattributed quotes. You can reduce them with better prompts, but you cannot fully switch them off, which is why a quick fact-check habit matters.
How do I know if ChatGPT is telling the truth?
Verify the specifics. Ask it to list its sources and check each link actually exists and supports the claim. Confirm names, numbers, dates, and quotes against a primary source you trust. Ask the model to flag its own low-confidence claims, and cross-check anything important with a web search or a second AI tool before you rely on it.
Can ChatGPT fact-check itself?
Partly. If you ask it to review its answer and flag anything it is not certain about, it will often surface the exact claims it was guessing on, which tells you where to verify. But it cannot reliably confirm whether something is true, because it has the same blind spots reviewing its work as it did writing it. Use self-review to find weak spots, then verify them with an outside source.
Why does ChatGPT sound so confident when it’s wrong?
Because it is designed to produce fluent, natural-sounding text, and fluency reads as confidence. The model has no built-in sense of certainty, so a guess and a fact come out in the same calm, authoritative tone. That mismatch between confidence and accuracy is exactly what makes AI errors dangerous, and why you verify the specifics rather than trusting the delivery.
When should I fact-check ChatGPT?
Use a 30-second risk test: ask what happens if the answer is wrong. For low-stakes tasks like brainstorming or rough drafting, trust it and move on. For anything going to a client, anything with statistics, anything legal, medical, or financial, or anything published under your name, verify hard using the full five-step routine.
About this guide
A practical, non-technical walkthrough from the team at Future Factors AI, who have trained 2,000+ professionals to use AI with confidence. Tools and features change often, so confirm current settings and always verify specific claims before you rely on them.
Related reading
Sana is an AI educator and learning designer specialising in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.
More about Sana →- [1] Wikipedia. Mata v. Avianca, Inc. (ChatGPT fake-cases sanction). 2023.
- [2] Seyfarth Shaw LLP. Counsel who submitted fake cases are sanctioned. 2023.
- [3] OpenAI Help Center. Data controls and model behaviour. 2026.
- [4] OpenAI. How ChatGPT handles browsing and sources. 2026.
