Checking AI output before you use it

What it is

AI assistants such as Claude, ChatGPT, Microsoft Copilot and Gemini write fluent, confident text. Fluent is not the same as correct. They can invent facts, misread a document, get a sum wrong or miss something important, and the writing gives no hint of it. Checking means reviewing the output before anyone relies on it.

Think of AI output like a draft from a capable new colleague in their first week. You would not send their report to a customer without reading it. You would check the numbers, the names and whether they understood the brief. The same habit applies here, and the person who uses the output remains responsible for it.

How much checking you need depends on the stakes. A brainstorm list needs a quick read. A figure in a board paper, a contract clause or a reply to a complaint needs careful checking against the source.

Why it matters

• Errors in AI output look just as polished as correct text, so they are easy to miss.
• A wrong number or invented reference can damage trust with customers and colleagues.
• Some mistakes have legal, financial or safety consequences.
• Clear checking habits let teams use AI with confidence instead of avoiding it.
• Responsibility stays with people: "the AI wrote it" is not an excuse.

How to do it

1. Decide the stakes: who will see this, and what happens if it is wrong?
2. Verify facts and numbers against the original source, not against the AI's summary.
3. Ask for sources: "Which part of the document supports each point?" Then open the source and confirm it says that.
4. Spot-check longer work: pick several items at random and check them fully; if any fail, check everything.
5. Run a second-opinion prompt: in a fresh chat, ask the assistant to find errors, gaps and weak claims in the draft.
6. Look for red flags (see below) and investigate each one.
7. Get sign-off from the named person who owns the decision before it is sent, published or acted on.

Example

An assistant has summarised a 20-page supplier contract into a one-page brief for a manager.

Second-opinion prompt (in a fresh chat):

Below is a contract and a summary of it. Act as a careful
reviewer. List:
1. Any statement in the summary that the contract does not support
2. Any important term in the contract the summary leaves out
   (dates, amounts, notice periods, penalties, renewal)
3. Any wording that is vaguer or stronger than the original
Quote the contract section for each point.

CONTRACT: [paste approved contract text]
SUMMARY: [paste summary]

It finds things to check; it does not replace checking the source.

Then the human check: open the contract and confirm each quoted section. Check the notice period and renewal date yourself, because these matter most. The contract owner reads the final brief and signs it off.

Red flags to watch for:

• Exact numbers, dates or percentages with no source.
• References, quotes or reports you cannot find.
• Very confident wording on a topic that is genuinely uncertain.
• Answers that agree a little too neatly with what you hoped.
• Content that changes when you ask the same question again.

Common mistakes

• Checking the AI against itself. Fix: always go back to the original document or a trusted source.
• Reading for style, not substance. Fix: check facts, numbers and names first, wording last.
• Trusting links and references without opening them. Fix: open every one you plan to use.
• No clear owner. Fix: name who signs off before work starts.
• Pasting sensitive documents into an unapproved tool to check them. Fix: use only approved tools and follow your data rules.

Try it this week

1. Take a piece of AI output you used recently, or create one with made-up content.
2. Highlight every fact, number, name and date in it.
3. Check each highlighted item against a source and mark it right or wrong.
4. Run the second-opinion prompt above on the same text.
5. Write down what you would have missed without checking.

Going further

• Write a short checklist for each type of document your team produces (for example "amounts match the invoice; dates match the contract").
• For numbers, ask the assistant to show its working, then recalculate in a spreadsheet.
• Compare answers from two different approved assistants on high-stakes questions; disagreement tells you where to look.
• Keep a simple log of errors you find; patterns show where AI needs more context or should not be used.
• For repeat work, build the checking step into the process, with a named reviewer, rather than leaving it to memory.

Key words

• Verify: confirm something is true by checking an independent source.
• Spot-check: check a random sample in full to judge overall quality.
• Second-opinion prompt: a request asking an assistant to critique a draft.
• Red flag: a warning sign that something may be wrong.
• Sign-off: approval by the person responsible before something is used.

Examples and figures in this guide are illustrative. AI tools change often: check what your organisation's approved tools offer today. Use approved tools only, and a person checks and decides.

Prefer to print or share it? Download this guide as a PDF (249 KB), or the complete AI Field Guide (1.7 MB).