When AI Gets It Wrong — Accuracy, Bias, and Who's Responsible

Lesson 8 · taught by Pip · free preview

We've guarded what the AI does. Now we guard what it says, because the most common AI mishap isn't a dramatic data breach — it's the bot stating something false in a confident, friendly voice, and a customer acting on it.

The plain idea: AI can be wrong while sounding completely sure of itself. This isn't a rare glitch; it's a known feature of how these tools work. They're built to produce a fluent, plausible-sounding answer — and "plausible" is not the same as "true."

The two ways your AI can be wrong

It makes something up. When an AI fills a gap with a confident invention, people in the field call it a hallucination. Your bot might cheerfully announce a sale you're not running, a service you don't offer, or a policy that doesn't exist — not out of malice, but because inventing a smooth answer is exactly what an ungrounded AI does. To the customer, it looks authoritative. That's the danger.

It carries a bias. Because AI learns from huge piles of human-made text, it can absorb human unfairness too. In a small business this shows up most in places like hiring help or screening customers — an AI quietly treating people differently based on a name or background. If you ever point AI at decisions about people, this moves from "good to know" to "get real guidance," because anti-discrimination rules are serious and old.

The cure for the made-up answer: grounding

The fix for hallucination isn't hoping the AI behaves. It's grounding — forcing the assistant to answer from your approved facts (your real hours, prices, services, policies) instead of its general guesswork. A grounded assistant pulls from a curated set of true things about your business. An ungrounded one improvises, and improvisation is where false promises are born. When you evaluate any business AI, "where does it get its answers?" is the question that separates the two.

Who's responsible? You are — and that's fine

Let's be direct about liability. If your AI tells a customer something false and they're harmed by relying on it, "the AI did it" is not a defense that protects you. You chose the tool, you pointed it at customers — it's your representative. (This is the same principle as lesson 3.)

This sounds heavy, but it's actually empowering: because the responsibility is yours, so is the control. Ground the AI in your real facts. Keep a human in the loop for anything that matters. Tell customers it's AI so they calibrate their trust. Do those three and the risk shrinks to a manageable size — the same size as the risk of a human employee occasionally misspeaking, which you already live with every day.

Your turn

Think of one fact about your business an ungrounded AI might get wrong — a price, a policy, a service you don't offer that sounds like you might. That's exactly the kind of thing grounding protects against. Note it; it's a clue about what your assistant most needs to be anchored to.

🔦 You can now spot a wrong answer before it costs you. In the final lesson, we tie it all together: how to judge an AI tool or vendor, and how to write your own one-page policy.