First, three ideas — AI, hallucination and reasoning mode
is not a living creature. It is a people built, and it answers by probability. There are many kinds of AI, such as recognizing faces, translating and recommending videos. This lesson is about the kind that writes text: the language model. A language model builds its answer one small piece, a token, at a time.
“” is only a technical nickname for an answer that sounds right but is wrong or has nothing to back it up — nothing inside is seeing things. When experts on the news say again and again that AI “makes mistakes” or “hallucinates,” the machine sounds alive, but that is not what it means.
“To think” and “to make a mistake” are words we use for living creatures. Yet many AI products have a “thinking (reasoning) mode,” which makes the machine look alive and thinking. A thinking mode only means the program first writes out its working — steps, a plan, a check — before it answers. That writing can be a few sentences or thousands of words, and it is built from probability too. It can help on hard problems, but it does not guarantee a right answer.
Scores and choices — from a bigram model to sampling
Inside the that runs a language model there is a part that scores every possible next piece and picks one. That is why we say an AI's answer is made from probability. You do not have to take this on faith — you can check it yourself. In a little while you will build a tiny of it: a model that only counts which word follows which. Real models are far bigger and score far better, but the last step — score, then choose — is the same.
Say you ask, “What color is the sky?” The AI does not answer because it knows the correct answer. It looks at the words so far and gives every word that could come next a score. Those scores come from patterns it learned from an enormous amount of writing — not from looking up a stored sentence.
Blue 70 points, gray 20 points, orange 10 points… (made-up numbers, just to show the idea). The AI draws one word from this score sheet — the higher the score, the bigger the chance — and there is nothing here worth calling a mind. Then it makes a new score sheet for the word after that, and repeats until the sentence ends.
Ask the exact same question twice and the answers may differ a little. Nothing mysterious is happening: it draws lots among the high-scoring words.
Sounding right is not the same as being right
AI gives convincing answers, doesn’t it? That is not because it understood you. It picked the words most likely to follow naturally from the writing it learned. Convincing and correct are not the same thing: a high score means “often written like this,” not “true.” A wrong claim comes out in exactly the same confident tone as a right one.
Probability describes a choice. Evidence supports a factual claim. Keep these two jobs separate.
No thoughts. No moods. No intent to lie. There is a calculated score sheet, and a draw made from it. A lie needs an intent to deceive, and a machine has none — which is also why it cannot notice when it is wrong. So treat AI as a tool you always check.
When an AI’s answer is wrong, do not say the machine “made a mistake” and stop there. Look for the cause. The writing it learned from may have been wrong or out of date. Its builders may have left out a step that checks facts, or a program may have had a . Or the question may have left out something it needed to know. Even without any of these, drawing by probability can still produce something false. A clear question helps a lot, but the check is still yours.
Three common misunderstandings
-
Misunderstanding AI looks up the correct answer on the internet.
Actually A language model calculates scores for the next token from the it learned during . Even when a search tool is attached, you still need to check the sources it brings back.
-
Misunderstanding A high-probability answer is a true answer.
Actually Probability only says which words often come next in a context like this. Whether a claim is true has to be checked against evidence.
-
Misunderstanding With temperature 0, the answer is always identical.
Actually It is usually very similar, but the model version, attached tools and the way the computation runs can still change the output.
Training, inference, checking — and your first experiment
An AI may run an enormous number of calculations inside, but in the end it uses the patterns of words it learned from huge amounts of human writing to pick the next word by probability. Learning those patterns is called training; using them to answer is called . A document you paste into the chat changes what the AI reads this time — it does not retrain it.
When you get an answer from an AI, ask yourself: “Is this actually true, or is it just a phrase that shows up a lot?” Then check one thing for real. Open the original source, look at its date, or work the number out again. Make that one question a habit, and AI will never push you around.
Your first experiment: after one word, candidate A appears three times and candidate B once. What are the probabilities? Predict them before opening ‘Build AI’ below, then compare greedy selection with sampling.
Check the answer
A appeared 3 times out of 4, so its probability is 3 ÷ 4 = 75%. B appeared once, so 1 ÷ 4 = 25%. Greedy decoding always picks A. If you sample many times, about three draws in four give A and one gives B.
Now let’s build a tiny AI ourselves. Once you have made the score sheet yourself, you can see with your own eyes how an AI picks its answer.
Now see probability with your own eyes and design your own little AI.