The Single Most Important Thing to Know About AI

Sometimes AI tools make things up.

Not "they're a little off." Not "they round numbers." They will invent — confidently, in well-written sentences, with no warning — facts, statistics, citations, case law, regulations, prices, product features, even people who don't exist.

This is called a hallucination. It is a real, well-documented behavior of every current AI tool, including Claude, ChatGPT, Perplexity, Gemini, and Copilot. Improvements happen with each model release, but no model is hallucination-free.

For a small business owner, this is the #1 risk in using AI. Everything else in this library — privacy, choosing the right tool, prompting well — is secondary to this.

Where Hallucinations Show Up

The same task can produce a great answer one day and a fabricated one the next. Patterns to watch for:

Statistics and percentages

"73% of small businesses say they would switch providers for a 10% discount."

That number sounds authoritative. It may have come from a real survey, a slightly misremembered survey, or thin air. AI tools regularly invent statistics that feel right.

Citations and case law

"According to the 2024 SBA report on..." "Under the Smith v. Henderson ruling..."

If you're including AI-generated citations in a proposal, a legal letter, or anything regulators or attorneys will read, verify every single one. Lawyers have been sanctioned by courts for filing AI-generated briefs that cited cases that did not exist.

Specific details about competitors, vendors, or products

"Vendor X charges $99/month for the basic plan with up to 5 users."

The model may be remembering an old pricing page, mixing up two competitors, or guessing. If you're making a buying decision based on competitive intel, click through to the actual website and verify.

Regulations and compliance specifics

"California requires PTO accrual at 1 hour per 30 hours worked, capped at 48 hours per year."

State-by-state employment, tax, and licensing rules change. AI was trained on a snapshot of the past. Always verify regulatory specifics with a current government source or a professional.

Software and platform features

"In QuickBooks, you can do X by going to Reports → Custom → Y."

AI may be describing the platform from a year ago. Menus change. Features get renamed. Try it; if it doesn't match, ask the AI to find current documentation.

Why Hallucinations Happen

You don't need the technical explanation, but the short version is: AI models predict plausible-sounding text. They are not consulting a database. When they don't know an answer, they often produce a fluent, confident-sounding answer that sounds like what would normally appear in that context — even if it's wrong.

This is why hallucinations are most likely on:

Specific names, numbers, and dates
Recent events (after the model's training cutoff)
Niche topics (less training data → more guessing)
Anything you can't independently verify

And least likely on:

General writing tasks (an offer letter, a follow-up email, a checklist)
Restating, summarizing, or rewriting content you provided
Common business advice and frameworks

Practical Rules for SMB Owners

Rule 1: Trust style, verify substance

If you used AI to write something — an email, a proposal, a job posting — read it for tone and quality, but trust the writing. Hallucinations rarely cause an email to read poorly.

If you used AI to tell you something — a fact, a number, a citation, a regulation, a competitor's pricing — assume it might be wrong until you verify.

Rule 2: Verify anything you'd be embarrassed to be wrong about

Ask yourself: "If this turned out to be made up, would it cost me money, a customer, an employee, a regulator, or my reputation?"

If yes, verify it before sending or acting on it.

Rule 3: Use Perplexity (or web search) for current facts

Perplexity, ChatGPT with web search, and Claude with web search all show you sources. If you're asking about anything time-sensitive or factual, prefer a tool that cites — and click the citations to confirm the source actually says what the AI claims.

Rule 4: Ask for sources, then check them

You can ask any AI tool: "What are your sources for that?" or "Cite the law you're referring to."

Sometimes it will produce real sources. Sometimes it will produce fake ones (yes — fabricated URLs, fabricated case names). Either way, the sources are useful only if you actually click through and read them.

Rule 5: Don't let confidence fool you

AI models write in a confident, fluent tone whether they're right or wrong. "Studies show..." and "It is widely accepted that..." mean nothing in AI output. The tone is generated, not earned.

Rule 6: Cross-check important answers across tools

If a fact, number, or recommendation matters, paste the same question into a second AI tool. If both give the same answer with consistent reasoning, your confidence goes up. If they disagree, you've learned something — investigate.

Rule 7: For high-stakes decisions, get a human expert

AI is fantastic for getting oriented quickly on a topic and asking better questions. It is not a substitute for a real attorney, accountant, HR professional, or regulator when the consequences are serious.

Quick Self-Check Before Sending or Acting

Before you send AI output to a customer, regulator, employee, or your own ledger, ask:

Does this contain any specific facts, numbers, or citations?
If yes, did I verify them?
If I'm wrong about any of them, what's the worst that happens?
Is that worst case acceptable to me?

If the answer to #4 is no, verify before you send.

Hallucinations — When AI Confidently Makes Things Up