ChatGPT and Claude make things up because they generate likely-sounding text from patterns, and when your request is missing the facts, they fill the gap with a guess. The guess comes out in the same confident voice as the correct answers, which is why it's so hard to spot.
You can't switch it off completely. You can cut it down a lot, in about fifteen minutes, by handing it the source, giving it permission to say "I don't know" and checking quotes and links yourself.
Why does ChatGPT make things up?
OpenAI's own help page is blunt about it. It says ChatGPT can sound confident even when it's wrong, and calls this a hallucination. Its examples include incorrect definitions, dates or facts, fabricated quotes, studies and citations, references to sources that don't exist, and overconfident answers to ambiguous questions.
It also lists the limits that cause the gaps. There's a knowledge cutoff, so it doesn't know about events after its training data unless it's using a tool. It may not be able to read a website because of technical issues, paywalls or robots.txt settings. And, in OpenAI's words, confidence isn't reliability.
Claude is no different. Anthropic's guide to reducing hallucinations says even the most advanced models, Claude included, can generate text that's factually incorrect or inconsistent with the context you gave.
So treat the output as a first draft from a fast, eager assistant who never says "no idea". Everything below is about making it easier for the assistant to be honest.
How do I stop it guessing?
Most bad answers come from missing context. Claude isn't dumb, it's guessing. Asking again doesn't add context. Giving it the context does.
1. Give it the source
Paste the document, upload the file or connect the app, then tell it to use only that. Anthropic lists this as "external knowledge restriction": instruct Claude to only use information from the documents you provide, not its general knowledge.
I do this when I build case study decks from client testimonials. I put the testimonials in NotebookLM, untick every source I don't want it touching, and it works from what's left. More on that in the stack I use every day.
2. Give it permission to say "I don't know"
By default, the model wants to be helpful, and a made-up answer looks more helpful than a gap. Anthropic says explicitly allowing "I don't know" can drastically reduce false information.
Add this to the end of any factual request:
If you're not sure, or the information isn't in what I've given you, say "I don't know" instead of guessing.
3. Ask for quotes first
For long documents, ask for word-for-word quotes before the analysis. Anthropic recommends this for documents over about 20,000 tokens. I'd try it on shorter ones too. The answer then rests on the text, not on the model's memory of what contracts usually say.
4. Make it check itself
After the first answer, ask it to find a direct quote from your source supporting each claim, and to remove any claim it can't back up. That's straight from Anthropic's guide, and it turns a wall of confident text into something you can audit.
5. Turn on search, then click the links
OpenAI says search lets ChatGPT look up current information and give cited answers, and deep research gives multi-source cited answers. Its own tips are to verify quotes, data and references, and to visit the links directly when accuracy matters.
Check that search is actually on. OpenAI says to look for the search or deep research icons in the input bar, or for web sources cited in the reply. If the answer has no citations, it came from memory.
6. Run it twice
Anthropic calls this best-of-N verification: ask the same question a few times and compare. If the numbers or names change between runs, don't trust any of them.
What should I check by hand?
Anything with a number, a quote, a name, a link or a legal or money consequence. If your name goes on it, you check it.
OpenAI's page also warns that ChatGPT can't reliably tell whether text was written by AI, and may make up an answer if you ask. Don't use it as a plagiarism checker.
Can I get ChatGPT to forget something false?
If you mean something wrong that you told it, yes. OpenAI's memory page says you can tell ChatGPT the remembered information is wrong or out of date, and edit or delete it from the Memory section of your Personalization settings. Deleting a saved memory doesn't remove mentions of it from past conversations, so delete the chat too.
I haven't found OpenAI documentation on correcting what the model says about a real person or business, so I won't guess at that process.
Do this in 15 minutes
- Pick one task you already do, like summarising a client call or answering a customer question.
- Open a fresh chat and paste in the real source, such as the transcript or policy.
- Add: "Use only what I've pasted. If it isn't there, say 'I don't know'. Start with word-for-word quotes."
- Ask it to match each claim to a quote and delete any it can't support.
- Click every link before it goes anywhere near a client.
If you keep pasting the same background about your business, put it in a Claude Project once instead of retyping it. How I use Claude Projects walks through the setup. And if you want it answering from your real inbox and calendar rather than guessing, connectors are the way.
Quick answers
Is Claude more accurate than ChatGPT?
I wouldn't pick one on that basis. Both OpenAI and Anthropic say their models can state things that are wrong, and both recommend verifying important claims. The fixes above work in either.
Does search mean I can stop checking?
No. Search gives it something to cite, and OpenAI still says to verify and to visit links directly. A cited answer can still misread its source.
Can I use ChatGPT to check whether something was written by AI?
No. OpenAI says it can't reliably tell, and may make up an answer if you ask.
Want someone to pressure-test your setup?
I run the Wright Mode membership, so I have an obvious interest here. Building a prompt that stops the guessing is quick. Building one for your own client work, then testing it on your real documents, goes faster with someone to ask. Inside Wright Mode you get the monthly live Build with Brooke, fortnightly Co-Work and Q&A, and a private community where you can bring the answer that went wrong and work out why.
