What generative AI actually is

The mental model that makes everything else make sense. About 30 min, beginner level.

You'll be able to

• Explain how a generative model writes an answer by predicting one token at a time.
• Distinguish what a model learned in training from what tools like web search and file upload add when you ask.
• Predict when an AI answer is likely to vary or be wrong, and decide when to check it.

Where generative AI fits

Artificial intelligence is the broad field of making computers do tasks that seem to need intelligence. Machine learning is the part of AI where a system learns patterns from data instead of following rules a person wrote by hand. Deep learning is machine learning with large neural networks made of many layers. Generative AI is deep learning used to produce new content: text, images, audio, video or code.

Early chatbots matched keywords to scripted replies. A modern language model has instead been trained on a huge collection of text, and it writes a new response each time, shaped by the patterns it learned.

It also helps to separate the model from the product. ChatGPT, Claude and Gemini are products: apps that wrap one or more models with features like web search, file upload and memory.

The key idea: generative AI makes new content that follows learned patterns. It is not a search engine looking up a stored answer.

Trained once, used many times

A model is built in two very different phases.

• Training happens once, before release. The model reads an enormous dataset and gradually adjusts billions of internal numbers, called parameters or weights, so it gets better at predicting text. This takes a great deal of computing power and time.
• Use (often called inference) happens every time someone sends a prompt. The weights are fixed; the model applies what it learned to your input.

Two consequences follow. First, the model does not keep a copy of its training documents to look things up in; what it has is a compressed set of patterns. Second, its knowledge stops at a training cut-off, the point where its training data ends.

The model also doesn't learn from your conversation in the moment. It can use your corrections within the current chat, but its weights don't change. Some products add memory features, or may use chats to train future models depending on your settings; that's the product, not the model changing mid-chat.

The key idea: what the model knows was fixed when it was trained.

One token at a time

Language models read and write in tokens: whole words, pieces of words, or punctuation.

Given all the text so far, the model calculates a probability for every token it could write next. One token is picked, added to the text, and the process repeats until the answer is complete. Every answer, however long, is built this way.

After “The capital of France is”, one continuation is overwhelmingly likely. After “Suggest a title for my essay”, many continuations are plausible. Chat tools usually pick among likely tokens with some deliberate randomness (a setting often called temperature), so the same question in fresh chats can lead down different paths.

That's why asking the same question several times is revealing. Parts that stay the same reflect strong patterns in the training data. Parts that change are where the model has many plausible options. But consistency isn't proof: a common misconception can be repeated just as reliably as a fact.

The key idea: an answer is a chain of likely next tokens, not a retrieved record.

Fluent is not the same as accurate

Training rewards the model for producing text that looks like the text it learned from. That makes it very good at sounding right: correct grammar, the right tone, the familiar structure of an explanation or a citation.

What the model doesn't have, by default, is a separate step that checks each claim against reality. When its training data holds strong, consistent patterns, the likely answer is usually the correct one. When the patterns are thin, such as niche topics, recent events, exact numbers, specific quotes or references, it still produces fluent text in the right shape. The result can be a confident, well-written answer that is partly or entirely made up. This is usually called a hallucination.

So tone tells you nothing about accuracy. A wrong date and a right date read the same. Specific, checkable details are the ones most worth checking.

The key idea: judge an answer by checking its claims, not by how confident it sounds.

Tools that ground the model

Most assistants now put tools around the model:

• Web search finds pages and passes their text to the model.
• File upload gives the model your document to read.
• Code execution lets the model write and run a program, for example to do exact arithmetic on a dataset.

In each case, the tool fetches or computes something, and the result is placed in the model's context, the text it can see while writing. The answer can then draw on that material instead of only on training patterns. This is called grounding, and it's how an assistant can tell you about last week's news despite its training cut-off.

Grounding reduces errors but doesn't remove them. The model can misread a source, blend it with its own patterns, or cite a page that doesn't quite say what the answer claims. Open the source when it matters.

The key idea: know whether an answer came from the model's memory or from a tool, and check the source either way.

Catching the prediction in the act

A student runs the module's activity in three steps.

1. Same question, three fresh chats. They ask “What are the three biggest challenges for cities adapting to extreme heat?” The three answers share a core (tree cover, vulnerable residents, energy demand), but the wording, order and third point change each time. The shared core is where the training patterns are strong; the changing parts are where the model had several plausible options.

2. A recent fact, search off. They ask what a city council decided at a meeting last week. With search off, a good answer says the model can't know events after its training data. A worse answer invents a plausible decision. Either way, nothing in the model's training covers that meeting.

3. The same question, search on. Now the answer cites a news article. The student opens it and confirms the decision is described correctly.

Their two sentences: “Without search, the model can only predict what sounds likely, so it can't know recent events. With search, it can answer from a real source, but I still had to open the source to know it was right.”

Try it

Three answers to the same open-ended question with notes on what stayed the same and what changed, the search-off and search-on answers to a recent-fact question, and two sentences on what the difference tells you.