What happened and why, in plain words. If the experiment names a term, it says what the term means here, with the plain English beside it.

So: what to do about it, and when it matters most. The two paragraphs make one point, and they never claim more than the results show.

You and Bit couldn't say where one coin would stop. But you could both say where most coins would stop. AI answers work the same way. This is called probabilistic determinism. You can't predict one answer, but you can predict where most will land. So two people who ask AI the same homework question get different answers that say about the same thing.

A small change to your question, called a nudge, works like dropping from lower down. It moves where most answers land. But it never makes one answer certain. Even from waist height, Bit never hit the bullseye. The size of your circles decided what counted as a hit, just as your rules decide what a good answer is. So decide what is good enough before you ask, not after.

The same three words got seven different orders. Which order came back depended on which AI Bit asked. This is what probabilistic means. Every answer is one draw from a range of answers, where some are common and others are rare. You'll see this when you ask AI to rank anything, like the best books for the summer or the order of your chores.

An exact request made the order we wanted the most common answer. Even then, some answers gave a different order. So be exact when one answer matters, and check every answer against your rule. A loose request is still useful when any fair order is fine, like the songs for a party. Then the range of answers can show you orders you hadn't thought of.

Bit's first roll was a 10, a total that comes up about once in 12 rolls. Alone, it said nothing about which total is most common. After 30 rolls, the shape was clear. Most totals were near 7, and the totals at the ends were rare. One AI answer is like one roll. From one answer, you can't tell if it's the most common one.

So don't judge AI on one answer, good or bad. When it matters, ask more than once, and see which answers come up most. Even 30 rolls gave only a rough shape. A 12 came up twice, and a 6 only once. The more answers you see, the closer you get to the true shape.

Bit's draws matched his bag. He drew mostly red, some yellow and no green, though 2 of his 20 candies were green. AI's answers follow its bag the same way. AI's bag is called its training data. It is everything AI learned from, before you asked. What's common there comes up often, and what's rare may not come up at all. So AI knows lots about pizza and nothing about your grandma's soup.

You can't change what AI learned from, but you can add to the bag. When your case is rare, put your own facts in the request. Adding green changed every color's odds, even yellow's. Whatever you add to a request changes the odds for the whole answer. For help with grandma's soup, type in her recipe and ask about that, not about soup in general.

When asked plainly, Pip and Max left the bakery and the cat blank every time. Nobody could know those. But the bread and tomato blanks seem to have answers. Pip and Max filled them every time with the most common answer, like May for tomatoes. This is called next-token prediction. AI writes a likely next word, then the next, so a guess comes out as the most common answer. AI never checks whether the answer is true where you live.

For you, the price of bread may be a fact you know. For Pip and Max it was a guess, written like the facts beside it. So when an answer depends on where you are, ask AI to mark what it knows and what it's guessing. Then check every guess against your own facts. You'll see this with anything that depends on where you live: bus times, store hours, prices, the rules at your school. Those are the answers to check first.

Bit's first 30 flips had two long runs of tails, and he decided the coin liked tails. The next 30 flips came out 15 heads and 15 tails. The runs were just chance. A run of four or more is common in 30 flips of a fair coin. About 9 people in 10 will see one. A run is what chance looks like, not a sign that something isn't fair. Runs fool people in other places too. Three rainy Saturdays in a row can make it feel like it always rains on weekends.

AI's answers come in runs too. Three strange answers in a row can be chance, and so can three good ones. Before you decide AI always does something, write what would prove you wrong. Then count enough new answers to see. Imagine AI got three math problems right in a row. That doesn't mean it always will, so check the fourth too. And when someone says AI is always wrong about something, ask how many answers they counted.

Pip and Max both chose Rise first. Fifteen of their 20 names began with Rise. And 9 had Crumb in them. This is called the usual answer. It's the one most people would give, so it's the one AI gives first. Your own first three names probably shared its words too. AI learned from what millions of people wrote. So the words most people choose are the ones AI chooses first.

The usual answer isn't wrong. It's just the answer everyone gives. When a name, a line or an idea needs to be different, first cross out the usual answers, yours and AI's. Then start from what's left. You'll see this at school too. If you ask AI for a science fair idea, your classmates may get the same one. Something only you know, like your street, your pet or a family recipe, is a good start.

Pip and Max said 7 and Starry Night all 10 times. They said mango 9 times. Max was even more predictable, and gave the same answers in 4 runs of 5. A question with a usual answer gets it nearly every time, so you can often write it yourself. AI usually picks the answer people give most often. Lots of people pick 7 when asked for a number. Everyday requests have usual answers too, like a gift for a teacher or a quick dinner. Sometimes the usual answer is just what you want.

So guess first. If your guess matches, you already had the usual answer, and you can ask for something different. If Pip or Max surprised you, like mango or Biscuit, AI's usual answer can differ from yours. Only a guess you write down first shows you this. A guess kept in your head often changes once you see the answer. To get a different answer, add one detail only you know, like your teacher's favorite color or how big the dog is. If the usual answer comes back anyway, say what to leave out.