The Confidence Trick

Here is a strange fact about the machines we just invited into our homes: they are most confident precisely when they are most wrong. Ask a chatbot something easy and it hedges. Ask it something it has no idea about, and it answers like a game-show host — fast, smooth, certain. Your kid talks to these machines every day.

So here is the question that should keep us up at night: who is teaching them to notice?

Let me tell you where the word comes from, because the history is doing half the work in this talk. In 1849, a man named William Thompson walked the streets of New York in a fine coat, struck up conversations with strangers, and asked a simple question: do you have the confidence to trust me with your watch until tomorrow? Remarkably often, they did. The newspapers called him the confidence man. That phrase got shortened over the years, and now we just call it a con.

Notice what the con man never did. He never picked a pocket. He never forced anything. He simply sounded sure, and sounding sure was enough. The victim handed over the watch voluntarily. The whole trick, then and now, is this: humans use confidence as a shortcut for truth. We can't check everything, so we check tone instead. Someone who answers quickly, smoothly, without hedging — some ancient part of our brain files that as this person knows.

For most of history, that shortcut worked well enough. Confidence was expensive to fake. You had to hold eye contact. You had to keep your story straight. You could get caught.

Then we built a machine that produces confidence for free.

I want to be precise about what a large language model actually does, because the precision is what makes this teachable to a ten-year-old. The model has read more text than any human ever will, and from all that text it learned one skill supremely well: given some words, predict the words that would plausibly come next. That's the whole engine. When it answers your question, it is not consulting an inner library of facts and checking each one. It is producing the shape of a good answer — the rhythm of it, the vocabulary of it, the confidence of it — because good answers are what filled its training data.

And here's the part that matters. The shape of a good answer looks identical whether the contents are true or false. The model produces fluent certainty the way a river produces water. It is not lying, exactly — lying requires knowing the truth and choosing against it. It's doing something stranger: it is sincere and sure and sometimes completely wrong, all at once, with no inner alarm bell that rings on the wrong ones.

Researchers call these failures hallucinations, which I've always found too gentle. A hallucination sounds like a glitch, a rare misfire. But the confident wrong answer isn't a malfunction of the system. It is the system, running exactly as designed, on a question it happened not to have the goods for. The fluency is constant. The truth is variable. That's the trick.

Now — why am I telling this to families, and not just to engineers?

Because your kid is on the receiving end of this trick more hours a day than any adult you know. Homework help, curiosity questions, settle-an-argument questions, what-should-I-watch questions. And the machine answers all of them in the same voice: warm, instant, unhesitating. A kid who grows up inside that voice, without anyone naming what's happening, learns a quiet lesson we never intended to teach: that certainty is what knowledge sounds like.

Adults have the opposite scar tissue, mostly. We've been burned. We've met the salesman, the smooth uncle, the internet stranger who was so sure. Kids haven't yet. They're meeting the most fluent speaker in human history before they've met the con man he learned his manners from.

Here's the good news, and it is real news: kids are naturally magnificent at exactly the skill this moment requires. There is no sound a ten-year-old loves more than an adult being wrong. Wait — actually, that's not quite it. It's not the being wrong they love. It's the catching. The moment where you point at the confident claim and say, prove it. Kids do this for sport. Our whole approach at Hi, Bot rests on one bet: that you can aim that instinct at the machines.

So let me give you the three moves. They're small enough to teach at dinner.

Move one: ask where it came from. When the machine gives you a fact, ask it for the source. Sometimes it names one. Then — and this is the move — go look. Half the time the source is real and says what the machine claimed. Sometimes the source is real and says something different. And occasionally the source doesn't exist at all: a plausible-sounding book by a plausible-sounding author that was never written. Every kid who finds their first invented citation levels up permanently. You can watch it happen. Something in their face changes.

Move two: ask twice, differently. Ask the same question two ways, in two chats, and compare. On solid ground the answers agree. On thin ice the model gives two smooth, confident, incompatible answers — and nothing teaches the lesson faster than seeing certainty contradict itself in under a minute.

Move three: find the edge of the map. Ask about something your family actually knows — the street you live on, grandpa's trade, the rules of the weird card game your cousins play. Watch how the machine handles the parts it can't know. Watch it fill the gaps with plausible texture, still in that same assured voice. Once a kid has seen the machine be smoothly wrong about their own street, they carry the right question into every future answer: is this one of the things it knows, or one of the things it's filling in?

None of this requires distrusting the machine, by the way. That's the framing I'd ask you to resist. The opposite of gullibility isn't cynicism — a kid who decides the machine is all lies is as badly calibrated as one who believes everything. The skill is checking. It's the difference between a reader and a believer. We spent a century teaching kids to read words. The next literacy is reading confidence — hearing a fluent answer and noticing, calmly, that fluency is a costume truth and falsehood both wear.

William Thompson eventually went to prison, because enough New Yorkers compared notes. That's worth sitting with. No single victim caught him. The catching was collective — people talking to each other about the man who sounded so sure. That's what a family dinner table can be. That's what a clubhouse of kids fact-checking chatbots together is.

The confidence trick is one hundred and seventy-seven years old. The machine running it is five. Your kid can learn to catch it in a weekend — and honestly, they'll enjoy it more than most things we call education.

The watch stays in your pocket. All anyone has to do is ask the second question.