AI at Work

How to Fact-Check ChatGPT

Fact-check ChatGPT by tracing claims to sources, running numbers in a real tool, and refusing to ship anything you cannot point to.

Keywordshow to fact-check chatgptBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity

Related searcheshow to fact check chatgptverify chatgpt outputchatgpt wrong answers at workfact checking chatgpt at workchatgpt accuracy for professionalshow to fact-check chatgpt

Fact-checking ChatGPT is a source map, not a reread

People fact-check ChatGPT by reading the draft again and asking whether it sounds right. That is how wrong answers survive. A fluent paragraph about competitor headcount, a 2024 regulation, or a margin movement will pass a reread because it is internally tidy. Work does not care about tidy. Work cares whether you can point to the filing, the workbook, or the clause. The method is slower to describe than to do: extract the claims, attach a source or a gap, run every number in a real tool, and refuse to ship the line you cannot point to. You are not marking style. You are deciding what is allowed to leave the building under your name.

The typical miss is a number that is almost the appendix figure, plus an according to that names no document. ChatGPT will also turn a consultation paper into law, or a pilot into a published result. Your scepticism should land on those upgrades. Ask what changed in the wording: proposed versus required, trial versus rolled out, management view versus audited. Then open the thing. If you cannot open it, you do not have a fact. You have a sentence. Sentences are cheap. Facts are what a partner, a regulator, or a customer can test while you are in the room.

This essay is the workplace habit, not a research seminar. You will use it on emails, commentaries, slide titles, and answers you are about to paste into a client thread. It assumes you already have an approved place to work and a pack you are allowed to use. It does not assume the model is trying to deceive you. It assumes the model is trying to complete a pattern. Completion is not verification. A named human still has to do the unglamorous pointing at the file, the cell, and the clause before anyone else reads the page.

If you cannot say where a claim lives in a file you hold, it is not checked. It is still a draft, however confident the tone.

Three tests: source, number, and what was upgraded

Build a three-column map on every draft that might be forwarded: claim, pointer, status. Pointer means a page, a cell, a ticket, or a URL you opened today. Status is sourced, rerun, or unverified. Unverified does not ship. Numbers get a fourth action: rerun in Excel, the warehouse, or the finance system, not in the chat. The upgrade test is separate. Highlight verbs and time words. If the source says considering and the draft says will, you have not found a typo. You have found the model finishing a story. Put the original verb back. That single habit catches a large share of workplace error without any special tool.

Grounded tools do not retire the map. A copilot that can see the document still summarises, and summaries drop exceptions. Ask for pointers, then click them. If the pointer is wrong, the rest of the page is suspect. Do not outsource the map to the same model that wrote the draft. You may ask it to list claims and suspected inferences. You still open the file. A self-critique that says looks consistent is worthless. Consistency is what fluent error already has. You need correspondence with the world, which in practice means correspondence with the pack.

Claim typeHow it failsThe check
Number or percentageAlmost the appendix, wrong periodRerun in Excel or the source system
Named sourceTitle, section, or finding invertedOpen it. Read the line.
Status wordProposed becomes requiredCompare the verb with the file
Person or organisationA role that confirmed itCheck the mail, the minutes, or cut it

A fact-check pass you can finish before you send

Work on a copy of the draft. Highlight every number, name, date, and rule. For each, write a pointer or [unverified]. Run the numbers in the real tool. Open the cited page. Search the pack for the name. Then delete or qualify every unverified line. Do not negotiate with a sentence because the paragraph will look uneven. Uneven and true is professional. Smooth and false is not. If the draft collapses once the unverified lines are gone, you did not have a document. You had scaffolding. Write the short true version yourself, or regenerate from a brief that only contains sourced facts.

Time-box the pass so it actually happens. Ten minutes on a one-page note. Twenty on a client letter. If the artefact needs an hour of checking, the prompt was doing research you have not resourced. Stop and get the pack, or send a human to the source. Keep a two-line log: what was wrong, what you changed. After a dozen cycles you will know your local failure modes: dates, legal status, and figures from the wrong tab. Share that list. Fact-checking ChatGPT becomes a team skill when the misses are named, not when everyone rereads for tone.

Role: You are an analyst preparing a draft I will fact-check and sign.
Task: List every factual claim in the text I paste.
Context: The artefact will be sent to the reader I name.
Constraints:
- Do not defend the draft. Do not add new facts.
- For each claim, say whether I provided a source in the paste.
- If I did not provide a source, mark [unverified]. Do not invent a citation.
Output: A table of claim, source in paste or [unverified], and any status word that looks upgraded.
Quality checks: Which three claims would embarrass us if a reader opened the file.

Use that list as a checklist, not as clearance. Open the file for the three claims that would embarrass you. Rerun the numbers. Put the original verb back wherever the model upgraded a status. Delete the unverified remainder. If the page becomes too thin to send, that is the fact-check working. Write a shorter note from the pack. Do not ask the model to fill the holes it just helped you see.

Rereading, circular citations, and the almost-right figure

Rereading rewards fluency. Circular citations reward laziness: the model cites a page, you skim the title, and you ship. Open the paragraph. Almost-right figures are the worst of the three. A margin of 12.4 percent when the pack says 12.1 will not trip your ear. It will trip the person who owns the workbook. Always rerun. If you will speak the number, take it from the cell in the same hour, not from a chat you ran yesterday. Files move. Chats do not update themselves. Yesterday's fluent total is how commentary drifts from the close.

Do not fact-check in the same window you use to generate, then paste the blessed text into a client system without a second look at names. Copy-paste reintroduces an earlier error you thought you cut. Work in the artefact that will be sent: the email, the slide, the memo. Watch personal data. A fact-check that requires pasting a customer file into a consumer chat is the wrong procedure. Use the approved tool or check the file with your eyes and a calculator. The method is the pointing, not the brand of assistant.

MistakeWhat it looks likeWhat to do instead
Sounds rightReread for toneClaim map and a pointer
Citation as proofA title in a listOpen the line and read it
Almost the number12.4 instead of 12.1Rerun in the real tool
Unverified fillerKept so the page looks fullCut or mark as unknown

A draft with three unsourced claims is not three-quarters done. It is not ready. Remove the claims or get the file. Do not send the mixture.

Related reading on StudyGrid

Read next: Where AI Falls Short How to Catch AI Hallucinations How to Cite AI Sources. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.

What to do this week

Pick one artefact you will send this week. Build a claim map with a pointer or an unverified mark on every number, name, and rule. Rerun every number in the real tool. Open every cited page. Cut what you cannot point to. Send the thinner true version. Keep a two-line log of what the model got wrong. That log is how fact-checking ChatGPT becomes faster than rereading for a feeling of quality.

FAQ: How to Fact-Check ChatGPT

Common questions about this page.

How do I fact-check ChatGPT at work?

List every claim. Point each one to a source you hold, a number you can rerun, or mark it as unverified. Do not ship the unverified line. Fluency is not a source.

Why does ChatGPT give wrong answers at work?

It predicts plausible text. It does not query your ledger or your filing cabinet unless you grounded it in an approved tool. Gaps get filled with something that sounds like a fact.

Can I trust ChatGPT if it cites a source?

Only after you open that source. Invented titles, wrong section numbers, and real papers with the finding reversed all occur. The citation is a lead. It is not a check.

Is this StudyGrid essay free?

Yes. The full blog on StudyGrid (studygrid.in) is free. Open Blog in the header, or follow Previous and Next at the bottom of each essay.

Where should I start the StudyGrid blog?

Start at The AI Opportunity if you want the series in order. Open a single essay if you searched for a specific workplace task such as email, Excel, policy, or prompting.