ChatGPT is accurate enough to draft, brainstorm and explain, but not accurate enough to trust unchecked. It can state wrong facts, names and figures with total confidence. Use it for first drafts and thinking, then check anything you'll publish, send or act on, especially facts, quotes, figures and anything a customer will rely on.
The short verdict
People ask this as a yes or no question, and the honest answer is neither. ChatGPT is very good at sounding right. Most of the time it's right on general, well-known topics. But it doesn't know when it's wrong, and it won't always tell you. It produces the most likely-sounding answer, which is a different thing from the true one.
That's the whole problem in one line. The mistakes don't look like mistakes. A made-up statistic reads exactly like a real one, and a quote that nobody ever said sits in the text as calmly as a real one. If you want the full explanation of why this happens, read my piece on what AI hallucination is.
So treat it like a fast, well-read assistant who's sometimes confidently wrong. You'd let that person draft. You wouldn't let them send anything with your name on it without a look.
How much checking each task needs
Not everything needs the same scrutiny. Match your checking to what happens if it's wrong.
| Task | How much to check | Why |
|---|---|---|
| Brainstorming ideas, headlines, angles | Light | You're choosing from options, not relying on facts. |
| Rewriting or tidying your own text | Light to medium | It can quietly change your meaning. Read it back against the original. |
| Summarising a document you gave it | Medium | It can drop the important caveat or blur two points together. |
| Explaining a concept you'll repeat to others | Medium to heavy | Check the explanation against a source you trust before you teach it. |
| Facts, figures, dates, names, quotes | Heavy, every time | This is where invented detail turns up most and does most damage. |
| Anything legal, medical, financial or safety related | Don't rely on it | Go to a qualified professional. Use the AI only to help you write your questions. |
A five-minute check before you publish or send
For anything going out under your name, run this routine. It's quick, and it catches the errors that hurt.
- Highlight every checkable claim. Any fact, figure, name, date, quote or statement about what a company does. Those are your targets.
- Find each one at the source. Not in another AI answer, and not on a page that copied the claim. The company's own site, the original report, the official guidance.
- Cut what you can't confirm. If you can't find it in a couple of minutes, delete it or rewrite the sentence without it. A vaguer true sentence beats a specific false one.
- Check it still says what you meant. Read it once as your reader would. Does it promise anything you can't deliver?
- Read it out loud. You'll hear the bits that sound like a machine wrote them, and those are often the bits where accuracy slipped too.
For advice that'll shape a business decision, go further. I've written a separate guide on how to check AI business advice before you act on it.
Prompts that cut mistakes
You can't make it perfect, but how you ask makes a real difference. These are the habits I'd build in.
- Ask it to say when it's unsure. Add a line such as: "If you're not confident about a fact, say so and tell me what I should check." It won't catch everything, but it flags more.
- Give it your text to work from. Paste the document, notes or page and say: "Answer only from this. If the answer isn't here, say so." Working from your material is far safer than working from memory.
- Ask for its reasoning. "Walk me through how you got there." Weak steps are much easier to spot in the reasoning than in a polished answer.
- Ask what would change its answer. "What would make this advice wrong for my business?" That question surfaces the assumptions it made on your behalf.
- Never ask it to invent evidence. If you ask for a statistic to support your point, it'll often produce one. Find the figure first, then ask it to help you write around it.
Understanding what's going on under the bonnet helps here too. My explainer on what a large language model is shows why it predicts words rather than looking facts up.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
When to keep it out of the job
Some jobs just aren't a good fit. I wouldn't use ChatGPT as the final word on anything where being wrong has serious consequences: legal positions, health questions, money decisions, safety instructions or contracts. Use it to understand the vocabulary and prepare your questions, then take those questions to a person who's accountable for the answer.
I'd also keep it away from anything confidential unless you're clear about how your account handles what you type. Accuracy is one risk; privacy is another, and I've covered that in is ChatGPT private.
Used well, it saves hours. Used as an oracle, it'll eventually embarrass you. The difference is a few minutes of checking.
Quick answers
Why does ChatGPT make things up?
It predicts the most likely-sounding words based on patterns it learned, rather than looking up verified facts. When it lacks the answer, a plausible guess can come out sounding certain.
Is ChatGPT reliable for research?
It's useful for getting oriented and finding what to look for. Treat every fact, figure and quote as a lead to verify at the original source, not as the finding itself.
Is ChatGPT always right?
No. It's often right on general topics and wrong on specifics, and it rarely signals which. Check anything you'll publish, send or act on.
