You know the person who explains a decade to you that you actually lived through? That's the clip, and it's the best hallucination analogy I've heard. My teenage staff love telling me '90s facts. I was there. Total confidence. Half of it wrong. Not one ounce of hesitation. If you didn't already know better, you'd believe them, because certainty reads as truth. That's your AI. It delivers a made-up fact with the identical tone it uses for a real one. There's no wobble, no "I think," no accuracy meter. So the takeaway isn't that AI is dumb. It's that confidence is a worthless signal, and you verify because you can't feel the difference. Who's the confidently-wrong person this instantly reminded you of?

Best mental model I've got for hallucinations, straight from the clip. Show your brain part of an image and it fills in the rest. That's what the LLM does. Your brain doesn't ask permission. It completes the picture and hands you a whole image, even though half of it was a guess. Language models do the same thing with facts. They fill in what they think should be there, and that filled-in piece can be something that never existed. Once you see it as pattern-completion instead of lookup, the made-up citations and confident wrong dates stop being surprising. They're the same mechanism working exactly as built. What's a hallucination you've caught that looked totally real at first?

We keep calling it lying. Wrong word, and the wrong word makes you check for the wrong thing. It's not lying to you, because lying involves intent. Lying means I know the truth and I choose to tell you a falsehood. The model has no idea it's wrong. It thinks it's handing you the truth every time, with the same certainty whether it's right or making something up. Paul's version in the clip: it's a really excited five-year-old. Completely sure, occasionally completely off. So confidence tells you nothing here. You need a check that doesn't rely on how the answer feels. What's your current check when an answer really matters?

The scariest AI story I know isn't sci-fi. It's a lawyer who filed real court papers built on six cases that never existed. An AI made them up, quotes and all, and he didn't check. He got sanctioned. And it's happened hundreds of times since, mostly to people who trusted a confident answer and skipped the thirty seconds of verifying. Here's the part that matters for us as operators: the same tool that saves you hours will hand you a fabricated fact in the exact same tone as a real one. Nothing flashes red. So before you put AI anywhere near a client, a contract, or a number, three free habits carry most of the weight: • Ask it where the answer came from • Click any source it cites • Verify anything that carries a real consequence Which of those three do you actually skip when you're busy?

If it's customer-facing, you own it. Ryan says that at the top of this clip, and it's worth unpacking what it actually means. Not the tool. Not the automation vendor. Not the model. You. And businesses have already been burned on this, both in customer loyalty and in actual dollars. The practical framework he lands on: automate the boring, predictable stuff first with formula-driven tools. Then add AI to the one step in the process that actually needs judgment. Could you run the whole business through AI? Maybe. Should you? Probably not. And definitely not as your first move. The 'could you versus should you' question is the most important one we ask when working with operators. What would your answer be for the customer-facing parts of your own business right now?