BETA This site is in beta. Information is still being added and reviewed.
Workshop overview
How it works. A companion to Module 1.

How AI actually works, in plain words

Once you see how a chatbot produces text, the failures in the other modules stop being surprising. At its core, a generative AI model works by predicting the next word, over and over, from patterns in a huge amount of text it was trained on. That huge amount of language is why these tools are called large language models. From that one simple engine, today's models do a great deal, and researchers genuinely disagree about whether that adds up to real understanding. You do not have to settle that debate to use the tool well. What matters for this work is that it produces language from statistical patterns, not from a stored set of checked facts, which is exactly why it can sound fluent and still be wrong.

Start activity

Be the model for two minutes

  1. One person writes the first four words of a sentence about a local office on a sheet of paper.
  2. Pass it around. Each person adds exactly one word, the most likely next word, with no discussion.
  3. At twelve words, read the sentence out loud.

Does it sound like a real sentence? Is any of it true? Nobody in the room decided that it would be.

Theofficeisopen until ondowntown
The model builds a sentence one word at a time, each step picking a likely next word from the patterns it learned. That sounds simple, and at a large scale it produces fluent, often impressive results. The catch for our work is that it is optimizing for what sounds likely, and it does not check any of it against facts.
  • It does not look up answers from a stored set of checked facts; it generates what is statistically likely from patterns. That is powerful, and it is also why it can sound fluent and still be wrong, which is what people mean when they say a chatbot hallucinates.
  • Because it learns from text, it is weaker wherever there is less text: lower-resource languages, local rules and local offices, and very recent changes. Its trained memory has a cutoff date, and although many tools now add a live web search to reach past it, an answer is still only as reliable as the sources it pulls in.
A little deeper

What a transformer is actually doing

The kind of AI in these tools is called a large language model, and the design most of them use is called a transformer. A transformer is a kind of artificial neural network, a web of simple connected units loosely inspired by how neurons connect in the brain, though it works quite differently from a real brain. You do not need the math to use it well. A few plain ideas explain both why it is so fluent and why it fails in the ways the modules describe.

Tokens, the pieces it reads and writes

The model does not work in whole words. It breaks text into smaller pieces called tokens, roughly a short word or a part of a word, and it reads and writes one token at a time. English usually breaks into few tokens. Many other languages break into far more tokens for the same sentence, which is one concrete reason the same request can be slower, cost more, and come back less accurate in your clients’ languages.

Attention, how it decides what matters

At each step the model looks back over everything written so far and weighs which earlier words matter most for choosing the next one. This weighing is called attention, and it is the idea the transformer added. It is why the model can keep a pronoun attached to the right name and hold a topic across a paragraph. The model is doing statistical weighing of how words relate, and it is not understanding meaning the way a person does.

Training, where the patterns come from

The model learned by reading an enormous amount of text and adjusting billions of internal numbers, called parameters, until its guess for the next token matched what really came next. Nothing was filed away as a fact you can look up. What it keeps is a web of patterns about which words tend to follow which. This is why it can produce a fluent, confident sentence that is simply wrong.

Sampling, why the same question gives different answers

The model does not always take the single most likely next token. It samples from the likely options, using a setting often pictured as a dial from cautious to creative. Turn the dial up and the writing gets more varied and drifts further from fact. This sampling is why asking the same question twice can give you two different answers.

The context window, why it forgets

The model can hold only a limited amount of text in view at once, called the context window. In a long conversation the earliest parts slide out of view, so it can forget what you told it near the start. Pasting a very long document can push earlier details out the same way.

None of this asks you to trust the machinery. It gives you the reason behind the habits in the other modules. Because the tool produces likely-sounding language, not verified facts, treat anything you will rely on as a fact as unchecked until you confirm it, through a web search or another trusted source. Check what matters, and expect the language gap.

One more layer

From words to numbers, what NLP is doing

The field behind these tools is called natural language processing, or NLP, which means getting computers to work with human language. A computer cannot do math on the word "clinic," so the first thing it does is turn language into numbers. This one move explains a lot of what the tools get right and wrong.

clinichospitaldoctor bicycle
Words used in similar ways sit close together in the model's map, so "clinic," "hospital," and "doctor" cluster while "bicycle" sits far off. The model answers by moving between nearby points.
Words become coordinates

NLP turns each token into a long list of numbers, called an embedding, which you can picture as a point in space. Words used in similar ways end up near each other, so "clinic," "hospital," and "doctor" sit close together, while "clinic" and "bicycle" sit far apart. The model never sees meaning. It works with positions and distances.

Meaning becomes geometry

Because related words sit near each other, the model can answer by moving through this space. It can carry "a place you go when you are sick" toward the neighborhood of "clinic" without ever being given a definition. This is powerful, and it is still association, so it can link things that only look related.

Associations carry bias

The positions come from real human text, so any pattern in that text, including stereotypes, gets built into the map. A model can place certain jobs or traits closer to one group than another simply because the training text did. This is one concrete way bias enters, before anyone asks the model a single question.

Why languages are not equal here

Each language gets its own share of this space, sized by how much text the model saw in it. A language with less text online gets a coarser, less reliable map. This is a deeper reason answers get worse in lower-resource languages, and not only when the model is translating.

You do not need the math. The useful idea is that these tools run on distances between words, learned from human text. That is why they are fluent, why they carry the biases of that text, and why they are weakest in the languages your clients most need.

Good to know
  • It can invent sources, citations, and links that look completely real.
  • It can forget earlier parts of a long conversation, because it holds only a limited amount of text in view at once.
  • Asking the same question twice can give you two different answers.
  • Confidence in its wording is not evidence that it is correct.
  • Anything you paste in has been sent to the tool, so treat it as shared with that company.
Try this

Ask the same question twice

Give the group one factual question about a local rule or office. Ask an AI tool the same question twice, in two separate chats, and read both answers aloud. Note where they differ, and mark the specific details, a phone number, a deadline, an address, that would need checking either way. The point lands on its own. Fluent wording is not the same as a confirmed fact.