All guides / Checking answers
How do we stop AI making things up?
You cannot rely on a prompt to stop every wrong answer. Give AI a small set of checked facts. Ask it to flag gaps. Then compare its draft with the original sources before you use it. Keep a person responsible for that check.
Start with a task where someone on your team knows the right answer. The prompt here helps you plan a test that can show where the draft goes wrong.
Try the promptSupply the facts. Test the gaps. Check the answer.
Give it facts you can check
Choose a few current files with a clear owner and date. Say which question the tool should answer. Tell it to leave a gap when the facts are missing.
More on this step
OpenAI’s terms say output may be inaccurate and needs review before use or sharing. A fluent answer is no reason to skip that step. Keep the source files beside the draft so the checker can find the right passage. Mark any point they cannot confirm as unresolved.
Read the source: OpenAI: accuracy and review of output
Test a question it cannot answer
Add a question whose answer is absent from the files. Keep the expected response on a separate sheet. The useful response is to show the gap and ask for the missing fact.
More on this step
Try two notes that disagree and a file with an old date. Does the draft notice the problem? Record the exact error when it does not. Change your brief or narrow the task, then repeat the same test. Keep failed examples for the next round of checks.
Open the source yourself
Follow each link and read the part meant to support the claim. Check names, dates, figures and the meaning of the words. A source can exist and still fail to support a sentence.
More on this step
The SRA has warned firms about made-up cases and other false AI content. That is one reason to check the actual source outside the chat. Asking the tool if it is sure leaves the checking task with the same system that wrote the answer.
Read the source: SRA: false information and AI misuse
Not sure where to start?
Build a test for wrong answers
Use this with made-up material or public files you have checked. The aim is to plan a test, with a person holding the correct answers.
Try this in your AI app
ChatGPT
Claude
Gemini
CodexUsing Gemini? The button copies the prompt. Paste it into the new chat.
Help me test an AI drafting task for made-up facts.
Ask me about the job first. What should the draft achieve, which sources should it use and who knows the correct answer? Ask one question at a time.
Help me choose a small source pack. Ask how we will check its date and who owns each file. Do not add facts from memory.
Plan tests with a missing fact, an old date and two notes that disagree. Include a question outside the pack. State what a good response should do in each case.
Keep the answer sheet separate from the inputs we will give the drafting tool. Suggest how our reviewer can compare the draft with the source.
End with a test table, the errors we should record and a clear point to pause. Do not promise that a new prompt will remove all mistakes. Treat claims without a checked source as unresolved.
Watch Warren's story
We checked Warren’s answers against the file
In Warren’s coaching, the hotel-standards document gave us a source to check. We added it to his Claude Project with instructions for his role. He learnt how to use that context in his work and built a dashboard to help prepare reports.
Read the source: Warren’s Leader Lab story
Testimonial · 0:24
"I’ve learned more than I have playing around with it in the last couple of months."
Warren Lavender · Luxury hospitality
Worked example
Catch a false finding or tender promise
A fire safety report has a note saying an item could not be reached. The draft turns it into a passed check. The reviewer finds the change by comparing the draft with the source note. The test fails. The team saves that case and checks it again after changing the brief.
A tender team tests an answer from a checked facts sheet. The sheet has no agreed response time. If the draft adds a promise to attend within an hour, the bid owner rejects it. A fluent sentence cannot fill that gap. The team needs a fact it can prove before the promise goes into a bid.
Made-up example.
Source notes
Keep the original finding or bid facts.
Draft check
Spot added passes and promises.
Reviewer’s decision
Reject claims with no source.
A few more things to know
Make the check part of the job
Decide who can release a draft and what that person must see. Keep the question, input files, output and review notes together. Label drafts so staff can tell them from checked copies.
If a claim cannot be traced, hold it back. Give the checker time to ask the source owner. Use the same release rule when a deadline is close. A quick first draft helps only when the whole job, including review and fixes, is worth doing that way.
Repeat tests after changes
A test result belongs to the files, task and tool you tested. When you change one of them, rerun the examples that exposed errors. Record what changed and what the checker found.
Compare the time spent finding sources, writing and reviewing. Count corrections as work. Ask the person using the final output if it helps them do their job. If review is too hard, choose a simpler task with clearer facts. Keep the scope small enough for your team to check well.
Will a better prompt fix this?
It may improve a particular draft, but you still need a check. Try known mistakes again and compare the answer with its source. Let the results guide what you use the tool for.
What if the answer has lots of links?
Open them. Check that each source exists and supports the nearby claim. Keep a list of any facts the reviewer could not verify. Links alone are not a completed check.