Artificial Intelligence vs Machine Learning vs Deep Learning vs LLM: the Differences

I hear "artificial intelligence" used to describe pretty much anything these days, from a spam filter to a chatbot that writes poetry. I get why: these labels all entered everyday language at roughly the same time, over the last few years, so it's tempting to treat them as interchangeable. The problem is that they aren't. Artificial Intelligence, Machine Learning, Deep Learning, and Large Language Model are four different things, nested inside each other like Russian dolls, and knowing where one ends and the next begins makes it a lot easier to read a tech headline or evaluate a new tool without getting lost.

In this article I want to sort things out, starting from the real, dated, source-checked history of these terms rather than half-remembered anecdotes, and ending with concrete examples you can actually point to. This is the first piece in a short series: in the next two I go deeper into how an LLM actually works and into its real limits. But first we need a map, otherwise it's easy to mistake the whole for one of its parts.

The big picture: four concentric circles

The most useful way to think about these four terms is as a series of concentric circles, from the broadest to the most specific. IBM puts it well in one of its technical guides: "AI is the overarching system. Machine learning is a subset of AI. Deep learning is a subfield of machine learning, and neural networks make up the backbone of deep learning algorithms" (IBM, AI vs. Machine Learning vs. Deep Learning vs. Neural Networks).

In other words:

• Artificial Intelligence (AI) is the broadest goal: building systems capable of simulating capabilities we associate with human intelligence, like reasoning, perception, or decision-making.
• Machine Learning (ML) is one way to reach that goal: instead of hand-writing explicit rules, you build algorithms that learn patterns from data.
• Deep Learning (DL) is a subset of Machine Learning: it uses a specific kind of algorithm, deep (that is, many-layered) artificial neural networks, loosely inspired by how neurons connect in the brain.
• Large Language Model (LLM) is a specific application of Deep Learning: deep neural networks, built on an architecture called the Transformer, trained on enormous amounts of text to process and generate language.

These aren't four parallel categories, then, but four levels of increasing specificity. Every LLM is a deep learning system, every deep learning system is a machine learning system, every machine learning system is an AI system. The reverse doesn't hold: not all AI is machine learning, and not all machine learning is deep learning.

Let's go through each one, with a bit of history attached.

Artificial Intelligence: the goal, not the method

The term "artificial intelligence" has a surprisingly precise birth date. It was coined by John McCarthy in the funding proposal for a research workshop held in the summer of 1956 at Dartmouth College, in the United States — an event now widely regarded as the founding moment of the field as an independent discipline (Dartmouth College). The organizers — McCarthy himself, together with Marvin Minsky, Claude Shannon, and Nathaniel Rochester — started from a premise that was fairly bold for the time: that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

Worth noting: the idea of "thinking machines" was already circulating before 1956. Six years earlier, in 1950, the British mathematician Alan Turing had published a paper in the journal Mind titled "Computing Machinery and Intelligence," proposing what we now call the "Turing test" as a way to determine whether a machine could exhibit behavior indistinguishable from a human's (Mind, Oxford Academic). But it's at Dartmouth that the field got a name and became an organized research program.

IBM's current definition is useful precisely because it stays deliberately broad: "AI is technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision-making, creativity and autonomy" (IBM, What Is Artificial Intelligence (AI)?). That breadth isn't a flaw in the definition — it's the whole point. AI is a goal — getting a machine to do something we associate with intelligence — not one specific method for getting there.

That means the "artificial intelligence" umbrella covers a lot of very different things, including techniques that have nothing to do with machine learning at all. A classic example is an expert system: a program that makes decisions by following thousands of rules written by hand by human experts ("if the patient has symptom A and symptom B but not C, then diagnostic hypothesis X"). It learns nothing from data — it simply applies hard-coded rules — and yet it counts, in every meaningful sense, as artificial intelligence, because it simulates a decision-making process that would normally require a human expert.

Other concrete examples that fall under the broader AI umbrella, with or without machine learning: chess engines built on search trees, voice assistants like Siri or Alexa, navigation systems that compute the shortest route, rule-based chatbots ("if the user types X, reply Y").

Machine Learning: learning from data instead of rules

Machine Learning emerges as a subfield of AI once people realize that hand-writing rules for every possible situation is impractical — think of how many rules you'd need to recognize a cat in a photo, across every pose, lighting condition, angle, and breed. The alternative idea is: instead of programming the rules, you program an algorithm that can extract them itself by observing many examples.

Here too, the history has a precise date and name attached. The term "machine learning" is credited (for the origin of the phrase, not the underlying concept, which is older) to Arthur Samuel, an IBM researcher, and his 1959 paper "Some Studies in Machine Learning Using the Game of Checkers," published in the IBM Journal of Research and Development (IBM, What Is Machine Learning?). Samuel had written a program that got better at playing checkers game after game, without every single strategy being explicitly written by a programmer. His goal, in his own words, was for "a computer [to] be programmed so that it will learn to play a better game of checkers than can be played by the person who wrote the program."

IBM's modern definition captures the idea well: machine learning is "the subset of artificial intelligence (AI) focused on algorithms that can 'learn' the patterns of training data and, subsequently, make accurate inferences about new data" (IBM, What Is Machine Learning?). The key word is "inferences": the system isn't applying pre-set rules, it's estimating a plausible answer based on what it has observed.

Concrete examples of machine learning we run into every day, often without noticing: the spam filters in our email inboxes, which learn to tell junk mail from legitimate mail by observing millions of already-labeled examples; recommendation systems on platforms like Amazon, which suggest products based on purchasing patterns (IBM); models that estimate the probability a credit card transaction is fraudulent; demand-forecasting systems used in logistics.

One important detail: a lot of "classic" machine learning doesn't use neural networks at all. Decision trees, statistical regressions, and support vector machines are all machine learning algorithms that existed, and were successfully used, well before deep learning became dominant.

Deep Learning: when neural networks get deep

Deep Learning is the subset of machine learning that uses artificial neural networks, specifically networks with many layers (hence "deep") stacked on top of each other. Each layer processes the information it receives from the previous one and passes a transformed version on to the next. The original inspiration comes, very loosely, from how biological neurons connect to each other — though that's an analogy that's more useful for teaching than as an accurate technical description of how a real brain works.

Neural networks as a mathematical idea have been around for decades, but for a long time they remained a niche field, held back by the scarcity of data and computing power needed to train them well. The turning point that's commonly credited with pushing deep learning to the center of the stage is 2012, when a neural network called AlexNet — developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton — won the annual ImageNet image-recognition competition (ILSVRC) by a margin no one expected from one year to the next: a top-5 error rate of 15.3%, against 26.2% for the runner-up (Krizhevsky, Sutskever, Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NeurIPS 2012). That result convinced much of the research community that deep neural networks, combined with enough data and enough compute (in that case, GPUs), could clearly outperform "classic" machine learning techniques on complex perceptual tasks.

Concrete examples of deep learning: the speech recognition behind virtual assistants, facial recognition, AI-assisted diagnosis from medical images (spotting anomalies in an X-ray, for instance), the computer vision systems in self-driving cars, and Google's search algorithm understanding the meaning behind a query (IBM).

Large Language Model: deep learning applied to language

Which brings us to the last and most recent circle: Large Language Models. An LLM is, in essence, a deep neural network trained on enormous amounts of text to estimate, given a chunk of text, what the most probable continuation is. It's therefore a specific application of deep learning, aimed at language.

The breakthrough that made today's LLMs possible is a neural network architecture called the Transformer, introduced in 2017 by a group of Google researchers in a paper that has become one of the most cited in recent machine learning history: "Attention Is All You Need" (Vaswani et al., 2017, arXiv:1706.03762). As Google's own research blog explains, the Transformer relies on a mechanism called "self-attention," which lets the model "aggregate information from all of the other words" in a sentence to build a representation of each word informed by the entire context, repeating this step multiple times in parallel (Google Research Blog, 2017). Compared to earlier architectures, the Transformer turned out to be far easier to parallelize on modern hardware, which made it possible to train much larger models in a reasonable amount of time — and, as we'll see in the next article, this same architecture underlies practically every modern LLM.

From there, scale grew fast. One well-documented example: in 2020 OpenAI published the paper on GPT-3, a model with 175 billion parameters, ten times larger than any previous non-sparse language model at the time (Brown et al., Language Models are Few-Shot Learners, 2020, arXiv:2005.14165). From that point on, the "Large" in LLM had well and truly earned its place.

It's worth noting that an LLM doesn't "know" things the way we do: it doesn't consult a database of verified facts, and it has no memory of lived events. What it does, at a very simplified level, is estimate which sequence of words is statistically the most plausible continuation of a piece of text, based on all the patterns it observed during training. That's a central point I'll dedicate a whole article to later on, because it's also the root of the "hallucination" phenomenon.

Examples of LLMs that are widely known today: OpenAI's GPT family, Anthropic's Claude, Google DeepMind's Gemini, Meta AI's Llama. They all share the same underlying architecture (the Transformer) and the same basic training philosophy (predicting the next token over enormous text corpora), even as they differ in size, training data, and the alignment techniques applied afterward.

A necessary detour: where does "generative AI" fit?

There's a fifth term you've probably heard just as often as the other four, and it's worth placing on the same map: generative AI. It isn't an extra layer in the hierarchy — it's a category that cuts horizontally across deep learning and LLMs: it's the artificial intelligence "that can create original content such as text, images, video, audio or software code in response to a user's prompt or request," as IBM defines it (IBM, What is Generative AI?). An LLM that writes text is generative. But there are also generative deep learning models that aren't LLMs — image generators, for instance — which share the deep learning foundation with LLMs, and often pieces of the Transformer architecture too, but don't operate on language and so don't qualify as LLMs.

The reason these terms all exploded into everyday use at once has a fairly precise date attached, too: November 30, 2022, when OpenAI made ChatGPT public. From that moment on, as IBM itself notes, "generative AI, and specifically the arrival of ChatGPT... has thrust AI into worldwide headlines and launched an unprecedented surge of AI innovation and adoption" (IBM, What is Generative AI?). It's understandable that in such a short span of time these four or five terms would blur together in everyday use: the technology spread faster than the technical vocabulary needed to describe it precisely.

A practical example to keep them straight

Let's line up all four levels using a single concrete scenario: an online customer service system.

• A rule-based system that replies "For returns, go to the returns page" whenever a customer types the word "return" is artificial intelligence (it simulates a function a person would normally perform), but it isn't machine learning: it hasn't learned anything, someone wrote that rule by hand.
• A system that analyzes thousands of past support tickets to predict which department should handle a new request, learning from patterns in historical data, is machine learning.
• A system that analyzes a customer's tone of voice or facial expressions during a video call to estimate their frustration level, using neural networks trained on thousands of hours of labeled recordings, is deep learning.
• A chatbot that reads a customer's request written in natural language, understands the context of the conversation, and generates a relevant written reply, all powered by a Transformer model trained on huge amounts of text, is an LLM.

All four can perfectly well coexist in the same product. And it's precisely this overlap — more than sloppy communication — that makes it natural to blur the terms together in everyday speech.

Why the difference is worth knowing

This isn't just an exercise in vocabulary precision. Knowing which "circle" a technology sits in helps you know what to expect from it: a rule-based system is predictable but rigid; classic machine learning is flexible but needs task-specific labeled data; deep learning can handle complex inputs (images, audio) but needs huge amounts of data and compute; an LLM can converse in natural language about almost any topic, but — as we'll see — precisely because of how it works, it can also confidently generate information that's wrong.

In the next article I go deeper into the technical details of this last circle: how an LLM actually works, from tokenizing text all the way to how the model generates, quite literally, one word at a time, the response you end up reading. And in the one after that, I tackle the part that causes the most misunderstandings: why LLMs hallucinate and how to use them without getting fooled.