There are three words almost everyone uses as if they were interchangeable: artificial intelligence, machine learning and deep learning. They show up in the same sentence, get swapped for one another without a second thought, and the result is a fog in which no one is quite sure what is being discussed anymore. The confusion is not foolish: all three describe real, related things. But they are not the same, and the relationship between them has a precise, almost geometric shape.
The shape is this: three circles, one inside the other. The largest is artificial intelligence. Inside it, smaller, sits machine learning. And inside that, smaller still, deep learning. Each circle fully contains the next. They do not half overlap or compete: they nest. Understanding that nesting is understanding what we mean when we talk about AI.
01 · the large circleArtificial intelligence: the ambition, not the method
Artificial intelligence is the outer circle, and it is the oldest of the three. The term was coined in 1956, at a summer workshop at Dartmouth College, where a group of researchers proposed to study how to make machines "use language, form abstractions and solve problems reserved for humans" [1]. Notice the ambition of the statement: it does not say how. It says what is to be achieved.
That is the key to not getting lost. AI is not a technique; it is a goal. It is the name we give to any attempt to make a machine do something that, when a human does it, we call intelligent: playing chess, translating a text, recognizing a face, planning a route. The how can be anything at all. For decades, much of AI learned nothing: it was rules written by hand by experts, long lists of "if this happens, do that". A program that plays checkers by following fixed rules is artificial intelligence, and there is not a single line of learning in it.
Artificial intelligence is a goal, not a recipe. It names what we want, not the way to get there.
When someone says "this uses AI", strictly speaking they have said almost nothing: they have named the largest circle. It is like saying a dish "uses cooking". The interesting question is always the one inside: does it learn from data, or does it follow rules someone wrote? That is where the real distinction begins.
02 · the middle circleMachine learning: when the machine learns from examples
The second circle, inside the first, is machine learning, automated learning. And here is where the most important conceptual turn in the field's recent history takes place. Instead of writing the rules by hand, you show the machine many examples and let it find the rules on its own.
The classic definition comes from Arthur Samuel, who in 1959 described machine learning as the field that gives computers "the ability to learn without being explicitly programmed" [2]. You do not tell the program how to tell junk mail from legitimate mail. You give it ten thousand already classified messages and it adjusts its parameters until it gets them right. The paradigm shifts: from programming the solution to programming a system that learns the solution.
Samuel himself gave the canonical example: a program that played checkers and improved by playing against itself, until it beat its own creator. That program was, at once, artificial intelligence (the large circle) and machine learning (the middle one). All machine learning is artificial intelligence. But not all artificial intelligence is machine learning: the hand written rules of the previous section live in the large circle and outside the middle one.
Within machine learning many families of methods coexist: decision trees, regressions, support vector machines, and also neural networks. Nearly all share the same anatomy: input data, a model with adjustable parameters, and a process that tunes those parameters to reduce the error. What changes from one family to another is the shape of the model and how it is adjusted.
03 · the small circleDeep learning: neural networks with many layers
The innermost circle is deep learning. It is not a rival of machine learning: it is a part of it. Specifically, it is the machine learning done with neural networks of many layers. That is the whole mystery of the word "deep": it refers to the depth of the network, to the number of layers stacked between the input and the output.
Why does depth change anything? Because each layer learns to represent the data at a level of abstraction a little higher than the one before. In a network that recognizes images, the first layers detect edges and patches of color; the middle ones, shapes like eyes or wheels; the last ones, complete objects. No one programs that hierarchy: it emerges from training. As LeCun, Bengio and Hinton summarized in their 2015 article, deep learning lets a model with many layers discover the necessary representations for itself, without a human designing by hand the features to look at [3].
"Deep" is neither praise nor metaphor: it counts layers. It is machine learning with networks of many stacked layers.
The technique is not new (neural networks have existed since the mid twentieth century), but for decades they were a marginal curiosity. What changed was a practical convergence: enormous amounts of data appeared and the hardware to process them arrived. The symbolic moment is usually dated to 2012, when a deep network called AlexNet won an image recognition competition by an enormous margin and made clear that, with enough data and computing power, depth won [4]. Since then, almost everything we associate with "modern AI" (the assistants that talk, the image generators, the language models) lives in this small circle.
| System | AI? | Machine learning? | Deep learning? |
|---|---|---|---|
| Hand written "if… then" rules | Yes | No | No |
| Decision tree over customer data | Yes | Yes | No |
| Spam filter by regression | Yes | Yes | No |
| Image recognizer (AlexNet) | Yes | Yes | Yes |
| Large language model (LLM) | Yes | Yes | Yes |
04 · the reflectionHow to tell which is which in real life
With the three circles in mind, the next time someone mixes the terms up you have a quick method to sort them out. Ask yourself three questions, from the outside in.
Is it trying to do something we would call intelligent? If so, it is artificial intelligence. That is enough for the large circle; it distinguishes almost nothing, but it locates the terrain.
Does it learn from data, or does it follow rules someone wrote? If it learns from examples and adjusts its parameters on its own, you have entered the middle circle: it is machine learning. If instead it runs fixed rules written by a person, it stays in the large circle and outside the middle one.
Does that learning use neural networks of many layers? If so, you have reached the small circle: it is deep learning. If it learns, but with trees, regressions or other methods without deep layers, it is machine learning that is not deep learning.
None of this requires knowing mathematics or programming. It requires remembering the shape: three circles, one inside the other, from the most ambitious and oldest to the most specific and most recent. Artificial intelligence is the question: can a machine do this? Machine learning is one answer: let it learn from examples. And deep learning is the version of that answer which, for now, works best when there is data and computing power to spare: stack layers and let the machine discover on its own what to pay attention to.
Sources
- McCarthy, J., Minsky, M., Rochester, N. & Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. Reprinted in AI Magazine 27(4), 2006, pp. 12-14. Founding document of the term "artificial intelligence".
- Samuel, A. L. (1959). Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development, 3(3), pp. 210-229. doi:10.1147/rd.33.0210.
- LeCun, Y., Bengio, Y. & Hinton, G. (2015). Deep Learning. Nature, 521, pp. 436-444. doi:10.1038/nature14539.
- Krizhevsky, A., Sutskever, I. & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems (NeurIPS), vol. 25. (Network known as "AlexNet".)