I probably catch an AI error about once a week.
Usually it happens when I’m asking about something I happen to know a lot about. Something in the answer doesn’t sound right, so I push back: “I don’t think that’s correct. Go back and check the actual source.”
And often the answer changes. The model acknowledges the mistake, corrects it, or tells me it can’t verify the information and asks me to provide the underlying data.
That doesn’t particularly bother me. I know AI gets things wrong.
What bothers me more are the subjects I don’t know enough about to catch it.
That’s where I’ve changed my own behavior. A citation isn’t proof that I’ve verified something. It’s a trail back to where the verification needs to happen.
A citation is a trail, not a guarantee
AI research products don’t all retrieve information the same way. Depending on the tool and question, an answer may rely on model knowledge, live web retrieval, search indexes, accessible webpages, licensed material or some combination.
And the information universe underneath the answer is uneven. Reuters Institute research found that many major news sites block AI crawlers; in its 2023 sample, 79% of top U.S. news sites blocked OpenAI’s crawler. (source) Tow Center testing has also found serious citation and source-identification errors in AI search tools, including fabricated or broken links. (source)
So what do you do when nobody on your team knows enough to catch the mistake?
Create a little friction before you trust the answer
1. Make the AI audit itself. Ask it to identify the factual claims that materially affect its conclusion, show the original source for each, and tell you what it is least confident about. Tell it not to defend its first answer — look for reasons it could be wrong.
2. Use another model as the skeptic. Give the answer to a second model and tell it to attack the facts, not improve the writing. Ask what is unsupported, overstated, outdated or contradicted by the original sources.
3. Go to the original source when it matters. Especially when the models disagree, the information affects an important decision, or the claim is going outside the company.
Two AIs agreeing still isn’t verification. But two AIs disagreeing is a very good reason to stop and look.
Decide what “checked” means
For a founder-led company, I don’t think the answer is a 40-page AI governance policy. Most small businesses need something much simpler: a shared definition of what “checked” means.
Low-stakes brainstorming or drafting? Use AI freely. Important internal decisions? Inspect the sources and look for contradictory evidence. External factual claims? Open the original source and verify the specific claim. High-stakes legal, financial, regulatory, safety or employment decisions? AI can assist the research; it shouldn’t be the final verification layer.
If five people in your company are using AI every day, they may have five different definitions of verification.
That’s not really a technology problem. It’s an operating-system problem.
“AI found it” and “we verified it” are two different jobs.