The Terminology Decoded

The vocabulary is genuinely confusing, partly because the terms overlap and partly because marketing departments use them interchangeably. Here is the map.

The nesting

Read it as nested boxes: deep learning sits inside neural networks, inside machine learning, inside AI. When a vendor says "our AI", they almost certainly mean machine learning, and quite likely deep learning.

Neural network and deep learning are so often used together that many people assume they are the same thing. A neural network is the structure; "deep" describes having a lot of it.

What a neural network actually is

Forget the brain analogy — it has caused more confusion than clarity. Here is a more honest picture.

Imagine estimating a house price. You might reason in stages. From size and number of rooms you form a rough sense of how much space you are getting. From postcode and school ratings you form a sense of how desirable the location is. From age and recent renovation you form a sense of condition. Then you combine those three impressions into a price.

You just described a neural network. The input facts feed into an intermediate layer of derived impressions, which feed into a final answer.

The difference is that nobody tells the network what the intermediate impressions should be. You do not instruct it to compute "desirability". You give it inputs and correct answers, and during training it invents whatever intermediate quantities help it predict well. In a network with many layers, the early layers build simple things and later layers combine them into complicated ones. In an image network the first layers respond to edges, later ones to shapes, later still to faces.

That is deep learning. Many layers, each building on the last, all of it discovered from examples rather than designed by a person.

It also explains why these systems are hard to interpret. The intermediate quantities were never meant to be human-readable. There is no "desirability" column to inspect — just numbers the network found useful.

Data science, analytics and machine learning

These three get muddled constantly, and the cleanest way to separate them is by what comes out at the end.

A machine learning project ends with software running in production. A data science project ends with a slide deck that changes a decision. Both are valuable; they are different jobs with different staffing and different definitions of success.

This distinction matters more than it sounds. Many "AI projects" that stall were actually data science questions dressed as engineering ones. If what you need is to understand why churn rose last quarter, you want an analysis, not a model — and being clear about that saves months.

Generative AI and large language models

The newest layer of vocabulary, and the one most likely to be used loosely.

Generative AI produces new content — text, images, audio, video, code — rather than a prediction or a classification. The difference from everything above is the shape of the output. A classic model outputs a number or a category. A generative model outputs an artefact.

Large language models (LLMs) are the text version, and they are what sits behind the chat assistants everyone now uses. An LLM works by repeatedly predicting the next chunk of text given everything before it. Trained on enormous quantities of writing, that turns out to produce fluent, useful, often correct output.

Two things follow that are worth carrying into every conversation about them:

• They generate, they do not look up. An LLM writing a citation is producing text that looks like a citation. It may be real. It may not. The system has no mechanism for knowing the difference, which is why confidently invented facts — usually called hallucinations — are a property of the technology rather than a bug awaiting a patch.
• They are not the same category of tool as a prediction model. A fraud model gives you a score you can act on automatically. An LLM gives you a draft that a person should read. Deploying the second as though it were the first is the most common AI mistake being made in organisations right now.