How to fact-check AI before you share it

How to fact-check AI at work comes down to a short routine you can finish in the time it takes to read your document once more. Most guides are written for students checking essay sources, but work drafts usually fail in simpler places: a total that skipped rows in your spreadsheet, a broken link, an invented quote, or an internal detail that should not leave your team. Running a focused check catches those mistakes while the draft is still yours.

Why does AI get facts wrong?

Every major AI vendor says in its own help pages that its tool can state wrong things confidently. OpenAI's research defines hallucinations as "plausible but false statements generated by language models," and OpenAI warns that ChatGPT "can produce incorrect or misleading outputs," might sound confident even when it is wrong, and can fabricate quotes, studies, citations or references to non-existent sources. Anthropic likewise states that Claude can produce incorrect or misleading responses, may lack up-to-date information past its model cutoff (for example, June 2026 for Claude Opus 5.5), and can generate broken links or authoritative-sounding quotes that are not grounded in fact.

Google says that Gemini can hallucinate, present inaccurate information as factual, and misdescribe how Gemini works, advising users to double-check responses and statements about people. Microsoft says Copilot in Word cannot understand meaning or evaluate accuracy.

How do you fact-check AI, step by step?

You can fact-check an AI report, summary or brief in seven steps that take about as long as reading the draft once more. Check the claims that carry the most risk first: the numbers, the quotes, the links, and the names of people. Check the numbers and links first, before the smooth prose lulls you into skimming.

1. Mark every specific claim. Highlight every number, date, named person, direct quote and linked source in the draft so you have a clear list to verify.
2. Compare every number against your source file. Open the spreadsheet or document you uploaded and confirm that each figure matches the exact sheet, column and rows.
3. Open every link in a browser tab. Click each citation because ChatGPT can reference non-existent sources and Claude has a documented issue with links that do not work.
4. Read the cited passage on the source page. OpenAI advises checking that a cited source actually supports the answer, and Anthropic warns that the original page may contain context Claude's summary left out.
5. Verify every quote and named person. OpenAI and Anthropic warn that their tools can fabricate authoritative-looking quotes, and Google specifically advises double-checking what Gemini says about people.
6. Check dates against the model's cutoff. Without tools or web search, ChatGPT and Claude do not know about events after their training data cutoffs, so confirm time-sensitive claims on live pages.
7. Remove details you did not mean to share. AI output can repeat names, numbers or details from your prompts or files that should not go to a boss or client.

When a claim comes from the web rather than your own files, use the SIFT method created by Mike Caulfield, which university libraries teach for checking information online. Its four moves are Stop, Investigate the source, Find better coverage, and Trace claims, quotes and media back to the original context.

How do you check numbers and charts from AI?

To check numbers and charts from ChatGPT or Claude, review the calculation steps the tool shows and test the totals against your original file. When analyzing data, both tools write and run Python scripts on your file, and both let you inspect what they did. If a chart comes from a screenshot or a scanned PDF, spot-check every row yourself.

For data work, ChatGPT writes and runs Python code on your file, and OpenAI says to review the generated code, outputs and assumptions before relying on the result. OpenAI warns that ChatGPT may not reliably extract exact values from image-based tables or scanned files, may not analyze the whole file (if a result looks incomplete, ask it to inspect specific sheets, rows or columns), and may pick an analysis method or chart type that does not match what you meant.

Claude's code execution and file creation feature (available on all plans; Free, Pro and Max users turn it on in Settings, then Capabilities) also writes Python scripts, gives summaries of its actions, and lets you review what it did, though Anthropic says you may need to review and refine its outputs.

How do you use each AI's built-in source checks?

ChatGPT, Claude, Gemini and Microsoft Copilot all include citation features that show which web pages or workspace files informed an answer, though none of them replace opening the source yourself. Each tool places its links slightly differently and carries specific limits.

ChatGPT search is on by default for all models across every plan (Free, Go, Plus, Pro, Business, Enterprise, Edu) and adds inline citations plus a Sources view when available, while ChatGPT deep research includes citations and a sources-used section in every report. Claude web search adds citations to every search-backed response so you can cross-reference authoritative sources.

In the Gemini app, a Sources button appears at the bottom of a response or in line for web and Google Workspace sources (older guides mention a green-and-orange "double-check response" button, but Google's help page for that feature is now titled "View related sources from Gemini Apps" and no longer describes double-checking). Gemini Notebook (formerly NotebookLM) answers from the sources you add, with citations you can hover over or select to jump to the passage, and Microsoft Copilot Chat provides inline citations and a bottom Sources button.

Can AI fact-check its own work?

A self-check by the same AI is a pointer, not proof, because a model can sound confident when it is wrong. Asking your AI to list where each number came from in your file or to run web search for links helps you find what to inspect, but OpenAI, Anthropic, Google and Microsoft all advise verifying the primary sources yourself.

Who should review AI work before you send it?

Before sending a report, brief or dashboard to a boss or client, ask a colleague who knows the underlying work to read it alongside your source files. Because you prompted the draft, your eyes easily skip over smooth sentences where the AI filled in a gap. A second reader catches what you stopped seeing, especially when their comments sit directly on the sentence or number in question. Pick a reader who knows the numbers, not just the writing.

If your work is already in Google Docs or Word, their built-in comments do that job well for a document. For an AI-generated brief, research summary, guide or lightweight dashboard, you can publish the draft as a page on SecondPage, keep general access set to Restricted for the reviewers you invite, and let them comment or suggest an anchored edit on the exact sentence. Mentioning Turner (SecondPage's built-in AI assistant) or using your own connected AI applies the fixes while keeping the same link, and version history shows what changed. (Disclosure: SecondPage is our product, and you are reading a SecondPage page.) For more ways to hand off drafts, see our guide on how to share a ChatGPT or Claude project with your team.

Before you send it

Run through these five checks before your draft goes to a boss or client.

• Numbers and totals. Every figure matches the exact sheet, column and rows in your source file, and any values from scanned or image-based tables were checked by hand.
• Links and context. Every cited link opens a working page that actually supports the claim without leaving out key context.
• Quotes and people. Direct quotes and statements about named people have been verified in the original source.
• Unintended details. Names, numbers or internal details you did not mean to share have been removed.
• Second reader. A colleague who knows the work has reviewed the draft on the document or page itself.