“ChatGPT can make mistakes. Check important info.” Here’s how.
ChatGPT and Claude both warn that they can make mistakes. What “check important info” means for answers about your documents, and how to check them in minutes.
Key takeaways
- The warnings are accurate. Both companies say plainly that their assistants can give answers that are wrong but sound right.
- “Check important info” leaves four questions open: which answers to check, what to check them against, how, and how you show that you did.
- For answers about a document, the check is concrete: find the passage each answer relies on, and confirm it says what the answer says.
- A check you can’t show is hard to rely on later. Keep the page, the passage and your decision with each answer.
ChatGPT and Claude both tell you to check their answers. Neither shows you how. Here is what the warning asks of you when the answer is about your own documents, and a way to do that check in minutes.
The line under every chat
Under the message box in ChatGPT sits a short line: “ChatGPT can make mistakes. Check important info.” Claude has its own: “Claude is AI and can make mistakes.”
These aren’t legal boilerplate. Anthropic’s help article on incorrect or misleading responses calls this “hallucinating”: Claude can produce quotes that look authoritative but aren’t grounded in fact. It advises that users “should not rely on Claude as a singular source of truth”, and should carefully scrutinize high-stakes advice. OpenAI’s help center gives the same advice for ChatGPT: outputs can be incorrect, so check important information.
That’s honest. It also moves the work to you.
What the warning leaves to you
“Check important info” is easy to say. In practice it raises four questions:
- Which answers are important? In a credit memo, a lease abstract or a diligence note, most of them: figures, dates, obligations, and anything a decision rests on.
- Check against what? For an answer about an attached PDF, the only check that counts is the PDF itself. Another chatbot, or the same one asked again, isn’t a source.
- How? Find the sentence or table row the answer relies on, and confirm it says the same thing, for the same period, entity and unit.
- How do you show you checked? A week later, “I looked at it” doesn’t help a reviewer, an auditor or a client.
The hard part is that wrong answers look like right ones. They carry real-looking figures, near-exact quotes and plausible page numbers, in the same confident tone. Our guide to catching AI hallucinations covers the common patterns. Asking for sources helps, but a citation is generated too, and it can point to the wrong page.
What a proper check looks like
For every answer that matters:
- The passage. The exact wording the answer relies on, found on the page of the PDF. If it can’t be found, the answer is unsupported, however reasonable it sounds.
- A status. Supported, calculated from figures in the document, contradicted by it, or not in the document. These are different outcomes: a contradiction is an error to fix, while “not in the document” is a gap to fill from another source.
- A decision. A person approves the answer or flags it. The tool finds the evidence; the judgment stays with you.
- A record. The page, the passage, the status and the decision, kept with the work.
By hand, that means a search for every answer, page by page. With twenty questions about a 200-page filing, the check takes longer than the answer did, which is why it often doesn’t happen.
Checking inside ChatGPT and Claude
CiteLLM adds that check to the chat you already use. Attach the PDF and ask your questions, or ask it to check answers you already have:
- Every answer comes back at its page, with the passage highlighted on the original document. A quote that can’t be found in the PDF is never highlighted.
- Each answer gets a status: supported, calculated (with the formula and each input highlighted), contradicted, or not in the document.
- You approve or flag each one, and download an evidence report with the passages and your decisions.
In our demo, an assistant summed up a nonprofit’s audited financial statements: an unmodified opinion and no internal control problems. The opinion was right. The control finding was not. The auditors’ schedule of findings reports a material weakness, and the check highlighted it on the page. Another answer gave a volunteer count the audit doesn’t contain. It was marked as not in the document, so no number was guessed.
A highlight shows where a passage is; it doesn’t prove that a conclusion drawn from it is right. That’s why each answer ends with your decision, not ours.
Frequently asked questions
Why does ChatGPT say it can make mistakes?
Because it can. Language models write the most likely wording, not wording they have verified, so an answer can be fluent and wrong. OpenAI and Anthropic both say so in their help centers.
Does asking ChatGPT or Claude for sources fix it?
It helps, because a cited answer can be checked. But the citation is part of the answer and can be wrong in the same ways: the wrong page, or a quote the document doesn’t contain. You still need to find the passage.
Is CiteLLM a fact-checker for the whole internet?
No. It checks answers against the documents you give it. If the document doesn’t say something, the answer is marked not in the document, even if it’s true elsewhere.