
Treat an AI tool's answer as a draft. Before you use it, check every fact, figure, name, date, quote, and link in it against a source you trust, and have a person review anything that reaches a customer.
That habit is how to check AI output. The sections below cover why it is needed, what to check, how to ask the tool to help, and where a person has to stay in the loop.
Why AI Tools Sometimes Get Things Wrong
The National Institute of Standards and Technology (NIST) released a Generative AI Profile in July 2024, and its news release lists 12 risks that the profile centers on. One of them is generative AI systems "confabulating" or "hallucinating" output. The release describes the profile's actions as ones developers can take, so it is not a rulebook for small businesses. What matters here is that the risk is named there.
Anthropic, the company behind the Claude models, says the same about its own product in its documentation: "Even the most advanced language models, such as Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context." It calls this phenomenon "hallucination." That page is written about Claude, so we do not claim every tool behaves the same way. Find out how yours behaves by testing it.
Neither source gives a figure for how often this happens, and we do not either. Our own point, which is advice and not a finding from either source: a wrong answer can read just like a right one, so the check has to be a habit, not a hunch about which answers look off.
What to Check Every Time
This checklist is our advice. Run it on any answer you plan to use:
- Numbers. Find where each figure comes from. If the tool cannot point to a source you can open, treat the number as unverified. - Names. Confirm the person, business, product, or place exists and is the right one, and that the spelling is right. - Dates. Check each date against the original document or calendar, including the year. - Quotes. Find the original wording. If you cannot find it, do not publish it as a quote. - Links. Open every link. Check that the page exists and that it says what the answer says it says. - Claims about rules or policies. Read the rule itself, on the page that publishes it.
An AI output review checklist like this is short on purpose. If the list is long, people skip it.
Ask the Tool to Show Where Each Claim Comes From
Anthropic's documentation lists several techniques for reducing hallucinations in Claude, and they translate into plain habits:
- Allow the tool to say it does not know. The documentation lists this as a technique, so tell the tool that "I don't know" is an acceptable answer. - Ask it to quote. For long documents, the documentation recommends asking the tool to pull word-for-word quotes first, before it does the task. - Verify with citations. The documentation describes having the tool cite quotes and sources for each claim, which makes the response auditable. You can also have it look for a supporting quote after it answers. If it cannot find one, it must retract the claim. - Run the prompt more than once. The documentation describes running the same prompt multiple times and comparing the outputs. Inconsistencies across outputs could indicate hallucinations. - Limit it to your documents. The documentation describes explicitly telling the tool to use only information from the documents you provide, not its general knowledge.
Whether your tool supports the same behavior is something to test. And a citation is a place to look, not proof. Our advice: open the source the tool names and read the passage yourself.
Giving a tool your own documents to answer from is the idea behind a knowledge base. Our article on knowledge base automation covers building the internal source an AI answers from.
When a Person Must Review It First
Our rule is simple: a person reads anything an AI tool drafts before a customer sees it. That covers emails, replies, quotes, web copy, and anything that states a fact about your business. The reader in the loop checks the list above and judges tone and fit.
Some situations go to a person from the start. Sensitive or complex situations belong with a person, and AI does not give legal, medical, accounting, or lending advice. Use the tool to prepare, and let a person decide.
Anthropic's documentation adds a caution worth keeping near the screen. Although these techniques significantly reduce hallucinations, "they don't eliminate them entirely. Always validate critical information, especially for high-stakes decisions."
Checking works best when there is something to check against. Our article on keeping the banquet event order current makes the case for one place where the current version of a detail lives, and the same idea helps here: know where the source of truth is before you compare an answer to it.
Where to Start
Pick one task where you already use an AI tool. Write down the checks that apply to its output and who runs them before it goes anywhere. Put the checklist where the person using the tool will see it.
Checking an answer is separate from the mistakes in how a tool gets adopted, such as buying before naming the problem. Our article on common AI mistakes covers those.
To work out where a person should review AI output in your business, get your free AI assessment.
Frequently Asked Questions
Why does AI sometimes give wrong answers?
Anthropic's documentation says even the most advanced language models can sometimes generate text that is factually incorrect or inconsistent with the given context. It is written about Claude, so test your own tool.
What is an AI hallucination?
It is the term Anthropic's documentation uses for text from a language model that is factually incorrect or inconsistent with the given context. NIST's news release also names generative AI systems "confabulating" or "hallucinating" output.
How do I check whether an AI answer is correct?
Check each number, name, date, quote, and link against a source you trust. This is our advice, and it works whatever tool produced the answer.
Should a person review AI output before a customer sees it?
Yes. That is our rule. A person reads and approves it first, and sensitive or complex situations go to a person.
Can I ask the AI tool to show its sources?
Yes, and Anthropic's documentation describes having Claude cite quotes and sources for each claim. Then open the source and read it yourself, because the citation does not prove the claim.

