The single biggest professional risk with generative AI in a small business isn't data leakage or copyright or even cost. It's hallucinations: the tool confidently producing something that sounds completely plausible and is just wrong. A fabricated statistic in a proposal. A misattributed quote in a board paper. A made-up case study in a sales email. A non-existent piece of legislation cited in an HR letter.
These slip through because AI hallucinations don't look like errors. They look like normal, professional, well-written content. They have the same tone, the same confidence, the same structure as the right answer. That's what makes them dangerous - and what makes them detectable, once you know what to look for.
Why AI tools hallucinate
It helps to understand the mechanism briefly, because it tells you when to be most suspicious. Large language models don't 'know' facts. They predict the most plausible next word given everything that came before. Most of the time, plausibility and truth line up - because the training data contained mostly true things. But when the model is asked about something specific (a particular court case, a specific company's 2024 revenue, the exact wording of a regulation), it will often produce something that sounds right rather than admit it doesn't know.
This is improving fast. Modern Copilot and ChatGPT are dramatically more honest about uncertainty than versions from a year ago, and grounding (where the tool quotes from your own files or live search results) hugely reduces the risk. But the failure mode still exists, and in business use the cost of one missed hallucination can be substantial.
The five places hallucinations hide
Train your team to be extra sceptical of five specific things.
- Numbers and statistics. Especially round, memorable, persuasive ones. 'Studies show 73% of customers prefer X' is a classic hallucination shape. If the source isn't named and linked, the number is more likely invented than not.
- Quotes and attributions. AI happily generates quotes from named people that those people never said. Particularly common in marketing drafts ('As Steve Jobs once said...') and in case studies.
- Specific citations. Case law, academic papers, news articles, regulatory clauses. AI will produce a perfectly formatted citation to something that doesn't exist. This has cost lawyers their jobs.
- Names of things that should be checkable. Product features that aren't in the actual product. Company achievements that didn't happen. Job titles for real people that are wrong. The AI is filling in plausible-sounding detail.
- Specific dates and historical events. Especially anything from the recent past, where the AI's training data may be incomplete or it's stitching together separate events.
If any of these appear in an AI-generated draft and matter to the outcome, they get checked. No exceptions.
The two-question check
Teach the team a simple two-question habit before anything AI-assisted goes to a customer, a regulator, or anyone external.
First: 'is there anything in here that, if wrong, would matter?'. If the whole document is general framing or summary of your own input, the risk is low. If it contains specific factual claims that would embarrass you if wrong, the risk is high.
Second, for anything in the second category: 'have I personally verified this from a source I trust, not from the AI?'. If the answer is no, either verify it or take it out. 'The AI said so' is not verification.
Five minutes of this on every external piece of work catches almost all hallucination incidents before they happen.
Use grounding wherever possible
Modern AI tools dramatically reduce hallucination risk when grounded in a specific source. In Copilot, that means using /file, /meeting or /chat to anchor the prompt in real content. In ChatGPT, that means using the 'browse' or 'analyse this document' features rather than asking from blank context. The error rate on grounded outputs is a fraction of the error rate on unanchored prompts.
Make grounding a default habit, not a special technique. 'Summarise the attached customer feedback' is much safer than 'summarise what our customers are saying'.
Don't ask the AI to check itself
A surprisingly common mistake: people who suspect a hallucination ask the same AI 'are you sure?'. The AI will then confidently re-confirm the wrong answer, or contradict itself in a way that's no more reliable. The model has no independent verification mechanism - it's making plausibility predictions either way.
Verification means a separate source. A web search, a primary document, a person who actually knows. If the only thing backing a claim is the AI's own confidence, you don't have verification, you have a vibe.
Build it into your review processes
For high-stakes work (legal documents, financial commentary, regulated communications), add a single sentence to your existing review checklist: 'has any AI-generated content been factually verified against a primary source?'. That one line, applied consistently, would have prevented most of the AI-incident news stories of the last 18 months.
When hallucinations are less risky
Be balanced about this. The hallucination risk is real and worth respecting, but it doesn't mean AI is unusable for serious work. For tasks where the AI is shaping, summarising or restructuring content you've provided (a meeting you were in, a document you wrote, a customer email you received), the hallucination risk is low - because the AI is working with your truth, not inventing its own. The risk concentrates in tasks where you're asking the AI to bring information you don't already have.
The honest summary
AI hallucinations are a real risk, a manageable risk, and a risk that gets smaller as the tools improve. The way to manage it is not to ban AI - it's to teach your team to recognise the five places hallucinations hide, ask the two-question check before external work goes out, ground prompts in real sources by default, and verify high-stakes claims against something other than the AI's own confidence. Five minutes of paranoia per important document is a cheap insurance policy for the time AI saves everywhere else.