AI hallucinations: recognize a claim that needs checking

By Jeremy Kenerson, Founder at BreadHead.

Last updated October 7, 2026

An AI hallucination is an answer from an AI tool that sounds right but is wrong, off topic or made up. That's how IBM defines it. So check an AI draft against records you trust before you use it.

What's an AI hallucination?

IBM says an AI hallucination is an output that sounds right but is wrong, off topic or made up. It also says hallucinations can be reduced but not fully removed.

Ask whether the draft has invented a detail, such as a service you've never offered. Check that detail against your own records before you approve the wording.

Start with the claims you'd need to defend to a customer: prices, service details and promises about your work. Mark names, dates and quotes for checking too.

Use this check:

• Where did this detail come from?
• Can I open the source and find the words that support it?
• Does the source cover this service or this use?
• What limits must stay beside the claim?
• Who can answer if the source leaves a gap?

Choose a small task before asking for a draft. Use the one-page task inventory to list the work you'd like help with and pick where to start.

If you want an agent to do that work, write down the task and the decisions you want to approve. Use the agent job brief to put those instructions in one place.

Why does AI make stuff up?

IBM says text-based large language models are trained to predict the next piece of text from patterns in a huge amount of writing. When you read the draft, ask what supports the words, even if they sound right.

Give the draft a narrow job. Supply the source text you want it to use, and ask for gaps to be left open.

After you get the draft, open those notes yourself. Compare each claimed source line with the actual text. If the wording adds a promise, remove it or ask the person who can approve that promise.

How do you identify hallucinations in AI?

Check each factual claim against a source you can read. If you can't find support, leave the claim out until you've checked it.

Use this check on a draft you're ready to approve:

1. Copy the exact claim into your notes without changing its wording.
2. Open the named source. If there's no source, ask for one or find the record yourself.
3. Find the passage that backs the claim. Read the text around it for limits or exceptions.
4. Compare the draft with that passage. Check that the draft hasn't changed who, what or when the source covers.
5. Record your decision: keep, revise or hold. For a hold, name the person who will resolve it.
6. Read the revised draft before you approve it. Check any new claims added during the edit.

Example

Suppose a draft claims your service includes weekend support. Before using that line, make a note like this:

• Claim to check: The service includes weekend support.
• Source: The approved service terms.
• Open question: Where do the terms confirm weekend support?
• Decision: Hold the claim until the service owner confirms the wording.
• Person to ask: The person who approves support terms.

Keep a blank version for your next draft:

• Claim to check: ______
• Source and exact passage: ______
• Limits to keep: ______
• Decision and reason: ______
• Person to ask: ______

Name who will check the work before it reaches a customer. With BreadHead, you review and approve the work before it goes public or into use. For a trial, require each factual claim to match a source you can read. Record that check on the pilot acceptance sheet.

Comparing platforms? See the best AI workflow automation tools.

What changed

• October 7, 2026: First version. Facts checked.