How the model decides · AI Tools Field Guide

Illustrative. After "the cat sat on the", the words mat, floor and dragon hold about 70, 26 and 4 percent of the draw, and one of them gets written.

Every word is a weighted draw

The model gives every word it could write next a score, turns those scores into shares of a draw, and takes one. Then it does the same for the word after that. You do not pick words this way, and neither does a dictionary lookup.

Score every candidate word
Turn the scores into shares
Draw one
Repeat

A three-hundred-word reply is about four hundred draws in a row.

The scores come from what training left behind. In text like this, after words like these, these are the words that came next.

How the odds are set

Training set the odds, so the draw is nearly certain wherever the language is nearly certain and spread wherever the language is open. One machine behaves two ways depending on the question.

What you asked	What the shares look like	What varies
"The capital of France is…"	One word holds almost the whole draw	Nothing you would notice
"Open the memo with a line about the deadline"	Dozens of words hold a slice each	Wording, order, emphasis
"What does paragraph 4 of the attachment say?"	Peaked, because the words are in front of it	Phrasing, not substance
"What does our leave instruction say?" (nothing attached)	Flat: nothing in the request decides it	The facts themselves

The last row is the one that costs people. A flat draw still returns a fluent, confident sentence. It had nothing to be confident about.

Temperature: how warm the draw is

Temperature is one number that sets how much the shares are flattened before the draw. Low keeps the leading words dominant. High gives the long tail a chance. It changes how varied the answer is and nothing about whether it is correct.

Setting	What it does	Fits
Zero	No draw at all: the top-scoring word wins every time, so the same request in the same conversation gives the same answer	Extraction, classification, code, anything you will run twice and compare
Cool	The leading words keep most of the draw; small variation in wording	Summaries, structured output, routine drafting
Warm	The tail gets a real share; more variety, more risk of a strange turn	Brainstorming, alternative openings, naming

Top-p does the same job a different way. It throws the tail away instead of flattening it, keeping only the few words that together cover most of the draw. A seed fixes the draws so a run can be repeated.

What is set for you, and what you set

Chat boxes usually expose none of this; the product picks a setting, and it can differ between a chat, a summarizer and a code tool inside the same product. Where you call a model from code, you set it yourself on every call.

"The model is random" means somebody chose how warm the draw is, and it was probably not you. Turning the dial down makes answers more repeatable. It does not make them more true. See Checking output.

What it changes at your desk

Task	What the draw affects	What you do
Draft, rewrite, change tone	Wording and order, not the material you supplied	Ask for three versions and pick one; edit from there
Summarize or extract from an attachment	Phrasing, and which small details get dropped	Run it twice; anything that moves between runs is worth checking
A number, date, count or reference	Everything, if you did not supply it	Supply it, or send it to a tool. Never take it from the draw
The same job over fifty rows	Consistency between rows	Give one worked example in the prompt and keep the wording identical
Anything you will compare run to run	Almost nothing, if the setting is cold	Say so in the request, and keep the source text identical

What to do with this

Next time two answers to the same question differ, look at which row of the table above you were in. If it was the last one, the material was missing and the setting had nothing to do with it. Then read Trust and limits.

The arithmetic is worked by hand, with a video, in Are all AIs just predictive number generators? in the Q&A book.