AI vs Machine Learning vs Deep Learning: What's the Difference?

Start hereA 2016 textbook draws AI, machine learning and deep learning as circles inside each other. Google's glossary notes that some organizations now use AI and machine learning interchangeably. So are they three names for one thing, or three different things?

Deep learning is a kind of machine learning, and machine learning is one way of building artificial intelligence. Picture three circles, one inside the next, with AI as the largest. A 2016 textbook, Deep Learning, draws it that way: deep learning inside representation learning, inside machine learning, inside AI. This page explains each term, where it came from, and how to tell them apart.

011. At a glance

	Artificial intelligence	Machine learning	Deep learning
In one line	Programs or models that can solve sophisticated tasks	Systems that train a model from data	Neural networks with more than one hidden layer
Key date	1955: the term appears in a proposal for a Dartmouth study	1959: the term appears in the title of a paper on checkers	2006: a wave of research begins under this name
Learns from data?	Not always: some systems run on knowledge people typed in	Yes: it finds patterns in data	Yes, and it also learns which features to look at
A textbook example	Knowledge bases	Logistic regression	Multilayer perceptrons (MLPs)

Each column sits inside the one to its left. The reverse does not hold, as the mix-ups below show.

022. A simple picture

Here is our own way to picture it. Say you want a computer to sort spam from real email. There are three broad ways to go about it.

• Write the rules yourself. That is AI without learning. Several AI projects tried to hard-code knowledge about the world in formal languages.
• Let it learn from examples. That is machine learning. A simple method called naive Bayes can separate legitimate email from spam. But the performance of simple methods like this depends heavily on how the data is represented. Often a person designs the right set of features first, then hands them to the method.
• Let it learn the features too. That is deep learning. In the textbook’s picture example, the first hidden layer picks out edges, the second corners and contours, and the third parts of objects.

033. What each term means, and where it came from

Artificial intelligence

In August 1955, four researchers proposed a two-month study of “artificial intelligence” at Dartmouth College, to run in the summer of 1956. The term goes back at least to that proposal. They started from a conjecture: that “every aspect of learning or any other feature of intelligence” could in principle be described precisely enough for a machine to simulate it.

One of the four, John McCarthy, later defined AI as “the science and engineering of making intelligent machines, especially intelligent computer programs”. His page was last revised in 2007.

Even official definitions vary. The NIST glossary lists six definitions of artificial intelligence: five from a US national security glossary that takes them from a US law, and one from a NIST publication. One of them reads: “A set of techniques, including machine learning, that is designed to approximate a cognitive task.”.

Machine learning

By 1959 Arthur Samuel was using the term in the title of a paper, “Some Studies in Machine Learning Using the Game of Checkers”. He reported that a computer can be programmed so that it learns to play a better game of checkers than the person who wrote the program. It could learn this in 8 or 10 hours of machine-playing time, given only the rules, a sense of direction and an incomplete list of parameters.

His introduction sets out the idea in one line: “Programming computers to learn from experience should eventually eliminate the need for much of this detailed programming effort.”

Today, Google’s Machine Learning Glossary describes machine learning as a program or system that trains a model from input data. The trained model can then make useful predictions from new data drawn from the same distribution as the training data.

Deep learning

In the glossary, a neural network is a model with at least one hidden layer. A hidden layer sits between the input (the features) and the output (the prediction). A deep model is a neural network with more than one hidden layer.

A 2015 review in Nature says deep learning lets models made of multiple processing layers learn representations of data at multiple levels of abstraction.

The ideas are older than the name. The 1955 Dartmouth proposal already listed “Neuron Nets” as one of its topics. The textbook describes three waves of this research: cybernetics in the 1940s to 1960s, then connectionism in the 1980s and 1990s. The third, under the name deep learning, began in 2006. The phrase itself is older: a 1986 conference paper by Rina Dechter used “deep learning” in a different sense, for a way of learning while searching for solutions.

044. What each one is used for

• AI in the broad sense covers anything built to solve sophisticated tasks. Google’s glossary gives two examples: translating text, and identifying diseases from radiologic images.
• Classic machine learning works from features a person has chosen. In a textbook example, a simple method called logistic regression can determine whether to recommend a cesarean delivery. It learns from features such as the presence or absence of a uterine scar, but it cannot change how those features are defined.
• Deep learning has done well with images, sound and text. Deep convolutional nets brought breakthroughs in processing images, video, speech and audio, while recurrent nets shed light on sequential data such as text and speech. The 2015 review says it improved the state of the art in speech recognition, visual object recognition, object detection, drug discovery and genomics.

Large language models are deep learning too. Google’s glossary describes a large language model, informally, as any language model based on the Transformer. Google researchers introduced the Transformer in 2017 as a new neural network architecture. Google’s Machine Learning Crash Course adds that a complete Transformer model stacks multiple self-attention layers on top of one another. Our page on how large language models work goes further.

055. The differences that matter in practice

• Who supplies the knowledge. In rule-based AI, people do. The textbook calls Cyc one of the most famous such projects. Its statements are entered by a staff of human supervisors. The textbook says none of these hard-coded projects has led to a major success. In machine learning, the system acquires its own knowledge by extracting patterns from raw data.
• Who chooses the features. In classic machine learning, a person usually decides what the model looks at. Learned representations often perform much better than hand-designed ones, and deep learning learns representations of data with multiple levels of abstraction.
• How much data and computing it needs. The textbook says neural networks have grown in size, driven by faster computers with larger memory and by larger datasets. Our reading: if a task has little data, it is worth asking whether a simpler method would do.

066. Common mix-ups

AI and machine learning mean the same thing. They overlap, but they are not the same. Google’s glossary says that formally, machine learning is a sub-field of artificial intelligence. It adds that some organizations have begun using the two terms interchangeably. Our reading: when a product says “AI”, it may simply mean machine learning.

If it’s AI, it learned from data. Not always. The textbook’s diagram says machine learning is used for many but not all approaches to AI.

Deep means it understands deeply. The word is about layers, not insight. The textbook notes there is no consensus about how much depth a model needs to count as “deep”.

Neural networks are copies of the brain. The textbook describes them as inspired by the brain, but generally not designed to be realistic models of biological function.

Generative AI is a fourth, separate thing. Our reading: it is a label for what some of these systems do, not a separate layer. Google’s glossary calls generative AI “an emerging transformative field with no formal definition”, and lists large language models among its examples.

077. How to tell them apart

Our rule of thumb: two questions sort most cases.

1. Did people write the rules, or did it learn them from data? Rules written by people point to AI without machine learning.
2. If it learned, did it learn its own features through many layers? If yes, it is deep learning. If a person chose the features, it is classic machine learning.

When a product page only says “AI”, look for its documentation or model card. Our AI glossary explains 22 related terms in plain words.

Where to go next

• Google’s Machine Learning Glossary: short definitions of machine learning terms, from Google’s developer site.
• Machine Learning Crash Course: Google describes it as a practical introduction with videos, visualizations and exercises.
• Deep Learning, the textbook: its authors say the online version is complete and will remain free. Chapter 1 includes a section on the field’s history.
• NIST’s AI Risk Management Framework: a January 2023 framework from the US standards body. It defines an AI system as one that generates outputs “such as predictions, recommendations, or decisions”.
• The 1955 Dartmouth proposal: the short document where the term appears, hosted on John McCarthy’s Stanford pages.

08Quick answers

09Use this

10Quick quiz

How we checked this · Last checked October 2026

• We read the 1955 Dartmouth proposal and John McCarthy's own question-and-answer page on his Stanford pages.
• For Samuel's 1959 paper we read the abstract and the introduction. We do not say who first used the term machine learning, only that Samuel was using it by 1959.
• For the deep learning textbook we read chapter 1 on its official website. For the 2015 Nature review we read the abstract.
• We read Google's Machine Learning Glossary and the NIST glossary entries directly. Glossary wording can change, so we will check them again.

How we check every article →

11Sources