How an AI Model Picks the Next Word

Every AI writing tool you've used works the same way underneath. It reads what's been written so far, guesses which word should come next, adds it, and repeats. That's it. There's no outline in its head and no point it's trying to make. Once you see this, the flat, polished tone of AI text stops being a mystery.

One word at a time

A language model is trained on a huge pile of text: books, websites, forums, documentation. From all that reading it learns one skill. Given some words, it can tell you how likely each possible next word is.

Strictly speaking it works in tokens, which are word pieces. A rough rule is that 100 tokens make about 75 English words. For this lesson, thinking in whole words is close enough.

Here's what that looks like for a simple sentence. The model has the words The meeting ran and has to pick the next one.

The model will almost always pick from the top of that list. "Long" is the most likely word because it's the most common thing people have written after "the meeting ran". "Sideways" is a better sentence. It tells you something happened. It also sits at the bottom, so the model rarely picks it.

Why the result sounds safe

Now stretch that one choice across a whole paragraph. At every step, the model leans toward the most expected word. Each choice is reasonable. Together they produce text with no surprises in it.

That's the core of the AI voice. It isn't wrong and it isn't badly written. It's the statistical average of everything the model has read, which is why it sounds like everyone and no one.

Two more training steps make it stronger:

• Instruction tuning. After the first round of training, models are taught to be helpful assistants. They learn to be polite, balanced and thorough. That's where the hedging and the tidy summary paragraphs come from.
• Human ratings. People rate which answers they prefer, and the model learns to produce more of what got high ratings. Confident, well-organized, pleasant text tends to win, so the model drifts toward it.

Researchers have measured the side effects. One study of medical paper abstracts found the word delves appeared about 28 times more often in 2024 than you'd expect from earlier years. Nobody started delving more. Writers started pasting in AI text. Lesson 1.4 covers the full list of tells like this.

What "temperature" changes

Most AI tools have a setting called temperature. At low temperature the model nearly always picks the top word. At high temperature it's more willing to pick from further down the list.

You might think turning it up fixes the problem. It helps a little, but it mostly adds randomness, not meaning. A high-temperature draft can pick "sideways" by luck, and it can just as easily pick a word that makes no sense. The model still doesn't know that your meeting actually went sideways because the client brought a new budget.

The model can only predict what's likely. It can't know what's true for you. Every fix in this course comes back to that gap.

What this means for your writing

Once you know how the words get chosen, the fixes are easy to see:

1. Give the model your specifics before it writes. If the prompt contains the client, the budget and the date, the "likely" next word is now about your situation. Module 2 starts with a brain-dump before you prompt.
2. Expect to edit the safe words out. The draft will lean on the most common phrasing. That's the part you replace with what you'd actually say. The 10-minute edit shows you how.
3. Add what the model couldn't predict. Your opinion, your numbers, the thing that went wrong. These are, by definition, the least predictable words in the piece, and they're the ones readers remember.

Next, lesson 1.2 puts numbers on this idea with perplexity and burstiness, the two measures most detectors rely on.