How AI assistants actually work

You do not need to understand the maths to use AI assistants well, but a simple mental model explains most of their behaviour.

Predicting the next piece of text

The core of an assistant is a large language model. It was trained on a very large amount of text to predict what comes next, one small piece of text (a token) at a time. Further training taught it to follow instructions and hold helpful conversations. When you send a message, the model generates a reply token by token, each time choosing a likely continuation given everything before it.

This explains several things:

• Fluency is not truth. The model is very good at producing text that sounds right. Sounding right and being right are different things.
• Context matters enormously. The model's answer depends on everything in the conversation. Better context leads to better continuations.
• Answers vary. Ask the same question twice and the wording, and sometimes the content, can change.

What assistants are good at

• Drafting and rewriting text in different tones and lengths
• Summarising documents you provide
• Explaining concepts at a chosen level
• Brainstorming options and arguments
• Transforming information: turning notes into tables, emails into task lists
• Helping with code, formulas and spreadsheets

Where they struggle

• Specific facts they cannot look up, such as recent events, niche details, exact numbers and citations. Without a search tool or your documents, they may invent plausible details. This is often called hallucination.
• Precise counting and arithmetic in long text, unless they use a calculation tool.
• Knowing what they do not know. They often answer confidently even when uncertain.
• Your private context. They do not know your company, customers or yesterday's meeting unless you tell them.

Tools change the picture

Many assistants can now search the web, read uploaded files, run code or connect to other apps. These tools help with facts and calculations, because the model can look things up or compute instead of guessing. But the model still decides what to search for and how to summarise results, so checking remains your job.

The right attitude

Treat an assistant like a fast, knowledgeable and sometimes overconfident colleague. Give clear briefs, provide the materials, and review the work before it goes out under your name.