Updated November 14, 2026
AI is wrong in a particular way. It’s not wrong like a tired colleague, with typos and obvious gaps. It’s wrong fluently, with tidy formatting and a confident tone, which makes mistakes harder to spot than ever.
If you use AI for anything that leaves your desk, you need a fast, repeatable way to check AI output. This guide gives you one: the kinds of errors to expect, a checklist that takes a few minutes, and prompts that make the AI help check its own work.
Key takeaways
- AI is wrong fluently, which makes errors easy to miss.
- To check AI output, mark every number, name, date, quote and source first.
- Open sources and recompute numbers for anything that matters.
- Match how carefully you check AI output to the risk of the task.
Why you can’t skip the check
Three pieces of research and one court case explain the stakes.
In a Harvard and Boston Consulting Group study, consultants using GPT-4 did better on many tasks, but on a task outside AI’s strengths they were 19 percentage points less likely to reach the right answer than consultants without it. The AI made them worse because they trusted it.
In a 2025 study by METR, experienced software developers using AI tools took 19% longer to finish real tasks, yet afterwards still believed AI had sped them up by about 20%. Feeling productive and being productive aren’t the same.
And in 2023, two lawyers in New York were fined $5,000 after filing a brief that cited court cases ChatGPT had invented. The real problem, the judge said, was that they kept standing by the fake cases instead of checking them.
The five kinds of AI errors
- Invented facts: sources, quotes, statistics or cases that don’t exist.
- Wrong details: real facts with the wrong date, number, name or version.
- Out-of-date information: things that were true when the model was trained but have since changed, like prices and job titles.
- Missing context: an answer that’s technically right but leaves out the exception that matters to you.
- Instruction drift: the AI quietly ignores part of your request, like a word limit or “don’t change the figures”.
A five-minute checklist to check AI output
- Mark the claims that matter. Highlight every number, date, name, quote and source. Those are your check list.
- Open the sources. For anything important, click through and confirm the page exists and says what the AI claims.
- Recompute numbers. Redo totals, percentages and conversions yourself or in a spreadsheet.
- Compare against your input. If you gave it a document, make sure it didn’t change figures or add things that weren’t there.
- Read it as the recipient. Would someone who knows the subject spot anything off?
For most work tasks this takes a few minutes. For legal, financial, medical or public content, have a second person check as well.
Prompts that help you check AI output
AI can’t fully verify itself, but it can point you to weak spots:
- “List every factual claim in your answer and rate how confident you are in each one.”
- “Which parts of this answer depend on information that may be out of date?”
- “Compare your summary with the original document and list anything you added that isn’t in it.”
- “What would someone who disagrees with this answer say?”
Treat the replies as a to-do list for your own checks, not as proof.
Set up AI to make fewer mistakes
Prevention helps. Give the AI the source material rather than asking it to recall facts. Tell it to say “I don’t know” or “not stated” instead of guessing. Ask for citations with page or section numbers. Use search-enabled modes for anything current. These habits are built into the prompts in our list of AI prompts for work.
Where checking matters most
| Lower risk | Higher risk |
|---|---|
| Brainstorming ideas | Numbers in reports and decks |
| Rewording your own draft | Legal, financial or medical statements |
| Internal meeting summaries | Anything sent to clients or published |
| First drafts you’ll rewrite | Code that touches data or money |
Match your effort to the risk. A brainstorm needs a glance. A client proposal needs the full checklist.
Our guides on summarizing documents with AI, AI research at work and AI for Excel include checks specific to those tasks.
This guide is part of our complete guide to using AI at work.
Quick FAQ
What is an AI hallucination?
When an AI produces information that sounds right but is false or made up, such as fake sources, wrong numbers or events that never happened.
How often does AI make mistakes?
It varies a lot by task and model. Errors are more likely with exact numbers, recent events, niche topics and long documents. Always check the claims your work depends on.
Can I ask AI to check its own work?
It helps find weak spots, but it isn’t proof. Use its self-review as a list of things to verify yourself.
