How do you check an AI answer before you use it?
Run a check sized to the stakes: open the sources, test the numbers and names, and never treat 'are you sure?' as a check.
Part of Check the workTime about 20 minutes, with the try-itChecked 16 Oct 2026Done saved in this browser
Three things to take away
- 1
An AI answer arrives looking finished. That look is not evidence: likely isn't the same as true.
- 2
Size the check to the stakes: who will see it, can it be undone, and does it carry names, numbers, quotes, law or money?
- 3
Open the sources yourself and check outside the chat. Asking it "are you sure?" isn't checking.
In your assistant
Same in all three
- Each can search the web and, often, show the sources an answer came from.
- Each has a deep research mode that writes a longer report with its sources: Research in Claude, deep research in ChatGPT, Deep Research in Gemini. A longer report has more claims, so it needs more checking, not less.
- Each vendor warns that its assistant can be wrong. The check is the same in all three: open the sources and confirm the key claims before you share.
Claude
- Web answers: when Claude searches the web, "Every response includes citations, so you can easily verify sources yourself", with direct citations, source links and relevant quotes. Source: Anthropic: Enable and use web search
- Research: turn on web search, then choose "+" and Research (Pro, Max, Team and Enterprise). It searches the web plus the connectors switched on in that chat, such as Gmail, Calendar and Docs; on the phone, the web only. Switch on only the connectors it should read. It produces a longer report "with proper citations". Open them as in move 2, and check which of your own files it used as well. Sources: Anthropic: Using the research feature; Anthropic: When should I use web search extended thinking and research; Anthropic: Claude Cowork and chat are one Claude
- Where Claude stands out: Claude for Excel (Pro and up) answers with clickable cell references, so you can jump to the cell a number came from. It finds the cause of errors, changes values while keeping formulas intact, and warns before overwriting. Dashboards (beta, Pro, Max, Team and Enterprise) show the query behind each chart and when it last refreshed, so you can see where a number came from. Sources: Anthropic: Excel; Anthropic: Dashboards and motion
- Anthropic's own limit for documents: it does not recommend Claude for Word for "Final client deliverables or counterparty sends without human review". Source: Anthropic: Word
ChatGPT
- OpenAI's own advice: "Use ChatGPT as a first draft, not a final source. Please check important information." and "Always verify quotes, data, technical information or references to external documents." The same page lists "Fabricated quotes, studies, citations or references to non-existent sources" among the things it can get wrong. Source: OpenAI: Does ChatGPT tell the truth
- Web answers: OpenAI: "Open a cited source to check that it supports the answer, review when it was published or updated, and use an authoritative source when accuracy matters." Source: OpenAI: ChatGPT search
- Deep research: type /Deepresearch, or choose "+", then Deep research, or use the sidebar. It searches the web, your files and your connected apps (read only). You review its plan first, and you can limit it to sites you name. The report comes with "A sources used section for reference checking". Limited on Free, with more on the other plans. Source: OpenAI: Deep research
- Where ChatGPT stands out: the most control over sources. Edit the plan before it starts, and name the sites it may use, so you know where every claim should come from.
Gemini
- Google's own warning: "Gemini can hallucinate and present inaccurate information as factual", and "Gemini Apps may provide inaccurate or inappropriate responses about people, so double-check its responses." Google also says not to rely on Gemini's answers as medical, legal, financial or other professional advice. Source: Google: Gemini Help
- Not every answer has sources. Google: "Not all responses include related links or sources." When an answer shows none, check it some other way, outside the chat. Source: Google: Gemini Help
- Deep Research: in the text box, choose Add Files, then Deep Research. It uses Google Search by default, plus Gmail or Drive if you connected them. Edit the plan, then choose "Start research"; it takes about 5 to 10 minutes. Free, with higher limits on Google AI Pro and Ultra. Source: Google: Gemini Help
- Where Gemini stands out: when the answer must come only from your own files, use a notebook in Gemini Notebook (the new name for NotebookLM since 16 Jul 2026). It is built for answers drawn only from the sources you add. Source: Google: NotebookLM Gemini notebook
Which one for checking?
| The job | Claude | ChatGPT | Google Gemini | My pick |
|---|---|---|---|---|
| A research report | Research that mixes the web with your connected work tools | Most control: edit the plan, limit it to named sites | Free Deep Research | ChatGPT, when you must control the sources |
| An answer from your own files only | A project with your files | A project with your files | Gemini Notebook, built for it | Gemini Notebook |
| Checking numbers in a spreadsheet | Claude for Excel: cell references, formula-safe edits, overwrite warnings | Excel and Google Sheets add-ins | Gemini in Sheets, for Google Workspace users | Claude |
My recommendation: Claude first, ChatGPT second, Gemini third.
For everyday checking, Claude makes the trail easiest to follow: Claude for Excel points you to the cell each figure came from, and its Research mixes the web with your own connected files. But checking is one job where I send people elsewhere too: when you must control which sources a report uses, use ChatGPT deep research, and when the answer must come only from your own files, use Gemini Notebook.
In every tool, the check is the same and it is yours. Use the tool your company approves for company data.
Try it (10 minutes)
Use a real question from your work, but paste nothing private.
- Pick a factual question you would normally look up, on a public topic: a product's limits, an industry figure, the date of a public event. Ask your assistant with web search or sources on.
- Underline every claim you could check: each number, name, date and quote.
- Open two of the cited sources. For each, find the exact line the answer relies on. Does it say the same thing, with the same hedges? Is it current?
- Redo one number yourself.
- Note what you found in one line: what was right, what was wrong, what you could not confirm.
If the topic touches a law or a rule, read the official text, not a summary; and this lesson is not legal advice.
Check yourself
-
1You ask the AI "are you sure?" and it says yes. Has the answer been checked?
The answer: A "Are you sure?" doesn't check the answer; it tells the AI which way you're leaning. Check against the source. -
2The answer quotes a report and gives a working link. What do you do before you use the quote?
The answer: C Quotations and links can be invented too. OpenAI's own page says citations can be incomplete, outdated or incorrect, and to open the source and check that it supports the answer. -
3You will recommend a supplier to your director, based on an AI comparison. Which move fits a decision?
The answer: B For a decision, ask for the case against. Telling it what you prefer is the opposite of keeping your opinion out of the question.
Read the whole lessonThe idea in full, the worked examples, common mistakes and what to check
The idea
An AI answer arrives looking finished: clean sentences, confident numbers, a tidy list of sources. That look is not evidence. These tools write a likely answer, and likely isn't the same as true. All three big vendors warn on their own help pages that their assistant can be wrong, and each one tells you to check (details below).
Checking is the skill people now rate highest. In Microsoft's 2026 Work Trend Index, a survey of 20,000 AI users in 10 countries (February to April 2026), the human skill rated most important as AI takes on more work was "quality control of AI output", named by 50 percent, ahead of critical thinking at 46 percent. Yet in the University of Melbourne and KPMG global study (more than 48,000 people in 47 countries, November 2024 to January 2025), the findings on AI at work say 66 percent rely on AI output without evaluating its accuracy, and 56 percent have made mistakes in their work because of AI.
The cost lands on someone else. BetterUp and the Stanford Social Media Lab use the word "workslop" for AI-made work that looks polished but is unhelpful, low quality or off the mark. In their September 2025 survey of 1,004 full-time US desk workers, 40 percent believed they had received some in the last month.
The good news: a real check is short, if you size it to the stakes and know where to look.
First, size the check to the stakes
Not every answer needs the same check. Ask three things: who will see it, can it be undone, and does it carry names, numbers, quotes, law or money?
| The answer is... | Check |
|---|---|
| A draft only you will see, easy to change | Move 1, and a careful read |
| Going to your team or your manager | Moves 1 to 4, on every claim that matters |
| Leaving the team (a client, the public, the board), carrying names, numbers, quotes, law or money, or hard to undo | All five moves, on every claim you can check |
The five moves
- Is it the job you asked for? Before you check a single fact, check the answer against your request: the right question, the right reader, the right format, nothing left out. A correct answer to the wrong question still fails.
- Open the sources yourself. Quotations and links can be invented too, so open the source yourself. A link that works is not enough: find the line it is cited for, check that it says what the answer claims, and check its date. OpenAI's own page says search results and citations "can be incomplete, outdated, or incorrect" (as of Oct 2026). If there is no source, ask for one: have it quote the line or link the page, then check the ones that matter yourself, outside the chat.
- Test the numbers, names and dates. Redo any sum or percentage from the figures in your own file. Check every name and job title against the source, and every date against a calendar or the document. OpenAI's page for ChatGPT for Word names the details to check: "Check important facts, figures, citations, and edits before using or sharing a document" (as of Oct 2026). Watch for lost hedges: if the source says "about" or "in one survey", the answer should too.
- Check outside the chat, never with "are you sure?" In 2023, a lawyer asked ChatGPT for court cases. It gave him six that looked real, with names, dates and quotes from the judges. They went to a federal judge, and not one of them existed. When he asked ChatGPT whether they were real, it said yes. Asking it "are you sure?" isn't checking. "Are you sure?" doesn't check the answer; it tells the AI which way you're leaning. Check against the source, a second source, or a person who knows.
- For a decision, ask for the case against. When the answer feeds a decision (a supplier, a hire, a plan), ask for the case against, so it has to argue the other side. And keep your opinion out of the question: don't say the plan is yours, and don't say what you're hoping to hear. Then weigh both sides yourself.
And when an AI tells you its work passed a check, ask what was checked.
A bigger model is not a check. Turning up the effort, or moving to a heavier model such as Opus 5.5, GPT-6 at High or Gemini's Pro, can give a more thorough answer, and it uses more of your limit. It does not replace the five moves. The check is still yours, outside the chat.
Common mistakes
- Reading for tone instead of checking claims. A likely answer reads smoothly whether or not it is true.
- Checking inside the same chat. "Are you sure?", "Double-check that" and "Is that correct?" all ask the same machine.
- Trusting a link because it is there. Open it, find the line, check the date.
- Losing the hedge. "About", "on average" and "in one study" are part of the fact. Keep them.
- Checking everything the same amount. Size it to the stakes, then check the claims that matter.
- Trusting a deep research report because it is long. More pages means more claims to check. Open the sources for the ones you will use.
- Switching to a bigger model instead of checking. A heavier model can be more thorough. It is still not a check.
- Forwarding it without saying what was checked. "The AI did it" is never the answer to "who checked this?"
What to check before you trust it
Before you trust your own check:
- You checked outside the chat: the source itself, a second source, or a person who knows.
- You read the source, not a summary of it, and it says what the answer says, hedges included.
- The source is current and from someone in a position to know (the vendor's own page, the official text, the original study).
- Anything surprising has two sources.
- You can say what you checked, what you could not confirm, and when it should be checked again. For a fact that changes, like a product feature, write the date beside it.
Checked against · read 16 Oct 2026Every claim, its source and the date read
As of: product steps as they were on the vendors' own pages on 16 Oct 2026.
- Neosaipien's platform reference (checked on the vendors' own pages, 16 Oct 2026): model names, research modes, the comparison and the recommendation
- OpenAI, Does ChatGPT tell the truth?
- OpenAI, ChatGPT search
- OpenAI, Deep research
- OpenAI, ChatGPT for Word
- OpenAI, ChatGPT plans (Excel and Google Sheets add-ins)
- Anthropic, Enable and use web search
- Anthropic, Use research on Claude
- Anthropic, Claude Cowork and chat are one Claude (help page; Research in the new single Claude)
- Anthropic, When should I use web search, extended thinking, and research?
- Anthropic, Use Claude for Excel
- Anthropic, Use Claude for Word
- Anthropic, Build live dashboards and animate explainers with Claude (8 Oct 2026)
- Anthropic, Change the model, effort, and thinking settings
- Google, Learn about responses from Gemini Apps
- Google, View related sources from Gemini Apps
- Google, Use Deep Research in Gemini Apps
- Google, NotebookLM is now Gemini Notebook (16 Jul 2026)
- Google, Get started with Google Workspace with Gemini (Gemini in Sheets)
- Microsoft, 2026 Work Trend Index (5 May 2026)
- KPMG and the University of Melbourne, Trust, attitudes and use of AI: a global study 2025
- University of Melbourne, media release on the 2025 global study
- BetterUp, The hidden cost of AI "workslop" (29 Sep 2025)
Where next
When should work live in a project, and how do you set one up?
Tell when a job belongs in a project rather than a single chat, set one up from start to end in Claude or ChatGPT (or a notebook in Gemini) with its files and instructions, and know who can see it and what it remembers (as of Oct 2026).
Open the next lessonThe AI Fact-Check Kit: a five-step checking routine (Sort, List, Trace, Hedge, Log), eight review cards by type of work, and a claims log.
The Team workshop. Workshop in a Box: Check What AI Writes teaches the same routine to a team in 60 minutes, or I run it live.
Book a 20-minute call