What is AI, actually?
Not a brain. A prediction machine.
You've spent four lessons on what AI means for your job. None of it works while you still believe the machine is thinking: you can't decide when to trust AI, when to double-check it, or when it's about to embarrass you.
So let's kill the myth right now. A large language model doesn't understand your question. It doesn't know facts. It doesn't reason, not the way you reason. What it does is much stranger, and much more useful once you see it clearly: it predicts the next word.
One word. Then the next. Then the next. Thousands of times per answer, each time asking the exact same question: given everything written so far, what's the most statistically likely thing to come next?
That's the whole trick. The fluent paragraphs, the working code, the jokes that actually land, all of it is that one mechanism, run enough times, at a scale large enough to feel like magic.
You be the model
Before you read another sentence about how this works, do the job yourself. For each prompt, pick the word you think a language model would output next, not the "correct" answer, the statistically likely one.
The capital of France is▍
I couldn't find my keys, so I▍
2 + 2 =▍
My favorite color is▍
There's no lookup table in there
Where does "statistically likely" come from? Training. The model read an almost incomprehensible pile of text, books, code, forums, articles, and for every stretch of words, adjusted billions of internal numbers (weights) so the actual next word in that text scored a little higher next time it saw something similar.
Do that trillions of times and the weights settle into a shape that's very good at one thing: given some text, guess what a human would plausibly write next. Not what's true. Not what's smart. What's likely, based on patterns it's already seen.
That's why round 3 above catches people. "2 + 2 =" isn't a calculation for the model, there's no arithmetic circuit running underneath. It's a memory of a pattern repeated an enormous number of times in training. Ask something with a similarly shaped but far less common answer, and this exact mechanism starts producing confident, fluent, wrong answers. That's not a bug you patch later. It's what the mechanism is.
Why the words we use lie to us
"The AI understood my question." "It knows the answer." "It's trying to help." Every one of those sentences smuggles in a mind that isn't there.
This isn't pedantry. The words you use to describe a system shape how much you trust it. Call it "understanding" and you'll assume it grasped your intent the way a colleague would. Call it "the highest-probability continuation given the input" and you immediately ask the right follow-up question: probable according to what data, and what happens when my situation isn't in that data?
Same system. Completely different amount of scrutiny.
Accurate or anthropomorphized?
Same event, described two ways. Sort each line by what it actually claims is happening inside the machine.
"The AI understood exactly what I meant."
"The model assigned the highest probability to 'Paris' given the input tokens."
"The AI knows the capital of France."
"Training adjusted the weights so this continuation is now scored higher."
"The AI wants to be helpful, so it tries its best."
Why this is the lesson that unlocks the rest
This reframe is a tool, not a footnote. Before you trust an AI answer, run it through this: is this the kind of thing that shows up constantly and clearly in the patterns it was trained on? Common historical fact, common code pattern, common phrasing, trust goes up. Obscure, recent, numerically precise, or logically multi-step, trust goes down, no matter how confident the sentence sounds.
"Confident" and "correct" are not the same property in this system. A model is exactly as confident predicting a wrong pattern as a right one, fluency is not evidence.
That single sentence is worth more than most of what you've read about AI in the news this year.
Next module goes under the hood of that training process itself, how a machine actually learns those patterns in the first place, starting with the three different ways machines learn.
What you learned
- ✓AI doesn't think or understand, it predicts the most statistically likely next word, over and over, based on patterns in its training data.
- ✓There's no fact database or reasoning engine inside, just billions of weights shaped by training to make plausible text score higher.
- ✓Words like "understand", "know", and "want" smuggle in a mind that isn't there, swap them for "predicted" and "trained on" and your trust calibrates correctly.
- ✓Confidence and correctness are unrelated in this system, a model sounds equally sure whether the pattern it's completing is common (usually right) or rare (often wrong).