AI vs machine learning vs deep learning

“Are we using AI or machine learning?” “Is this deep learning?” “Isn’t it all just ChatGPT?” I hear versions of these questions in almost every first meeting, and they’re fair questions — vendors use the terms interchangeably, and the media rarely stops to explain the difference.

In the first post of this series we defined AI. Now let’s sort out its family tree. The good news: once you see how the three terms fit together, you’ll never mix them up again.

Think nested circles, not competing options

The clearest way to picture the relationship is as circles inside circles. The well-known textbook Deep Learning by Goodfellow, Bengio and Courville draws exactly this diagram in its opening chapter: deep learning sits inside machine learning, which sits inside AI^[1].

So all deep learning is machine learning, and all machine learning is AI — but not the other way around. They aren’t rival technologies you choose between; they’re increasingly specific descriptions of how a system works.

AI without learning: rules written by people

The outer circle includes approaches that don’t learn at all. For decades, a lot of AI meant rules-based systems: experts wrote out their knowledge as “if this, then that” rules, and the computer followed them. A loan-approval system might say: if income is above X and existing debt is below Y, approve.

Rules-based AI is still everywhere, and for good reason — it’s predictable, easy to audit and cheap to run. Its weakness is that someone has to anticipate every situation in advance. The moment the world gets messy (handwriting, speech, photos, free-text emails), writing the rules by hand becomes impossible.

Machine learning: let the data write the rules

Machine learning flips the approach. Instead of telling the computer the rules, you show it lots of examples and let it work out the patterns itself. Show it thousands of emails labelled “spam” or “not spam”, and it learns what spam tends to look like.

The idea goes back further than most people expect. In 1959, IBM researcher Arthur Samuel published one of the earliest papers with “machine learning” in its title, describing a checkers-playing program that improved by playing games and learning from the results^[3].

Classic machine learning powers a huge amount of everyday business software:

• Predicting which customers are likely to cancel
• Forecasting next month’s demand
• Flagging suspicious transactions
• Scoring sales leads

These models usually work on neat, structured data — the rows and columns of a spreadsheet or database — and they’re often small, fast and inexpensive. We’ll look at how the learning actually happens in the next post.

Deep learning: many layers, messy data

Deep learning is a particular kind of machine learning that uses neural networks — software loosely inspired by how brain cells connect — stacked in many layers. That’s where the “deep” comes from: depth means layers, not profundity.

In their landmark 2015 review in Nature, Yann LeCun, Yoshua Bengio and Geoffrey Hinton describe it this way: “Deep learning allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction”^[2]. In plain English: early layers pick up simple features (edges in a photo, say), and later layers combine them into more meaningful ones (an eye, a face, a person).

That’s why deep learning shines on unstructured data — images, audio and language — where hand-written rules fail. The turning point many people point to came in 2012, when a deep neural network built by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton won the ImageNet image-recognition competition with a top-5 error rate of 15.3%, compared with 26.2% for the next-best entry^[4]. After that, deep learning took over computer vision, then speech, then language.

Side by side

So where do ChatGPT, Claude and Gemini fit?

The chat assistants everyone is talking about are built on large language models, which are deep learning models trained on enormous amounts of text. They live in the smallest, innermost circle. “Generative AI” is the broader label for deep learning models that create new content — text, images, audio or code — rather than just classifying or predicting. We’ll dedicate a full post to large language models later in this series.

What I want you to take away

1. They’re nested, not competing. Deep learning ⊂ machine learning ⊂ AI.
2. Not all AI learns. Rules-based systems are AI too, and they’re still useful.
3. Machine learning learns from examples. It’s the workhorse behind forecasting, scoring and fraud detection.
4. Deep learning handles messy data. Images, speech and language became practical because of it.
5. Chat assistants sit in the innermost circle. They’re deep learning models trained on text.
6. Pick the simplest tool that works. Fancier isn’t automatically better.