These four terms are not more or less learned synonyms: they nest inside one another. Artificial intelligence contains machine learning, which contains deep learning, which contains generative AI. Each level is a subset of the previous one. Once you have that picture, the whole sector's vocabulary becomes readable.
This is probably the most widespread confusion among those new to the subject, and it is kept alive by sloppy media usage. Let's clarify.
Level 1: artificial intelligence
This is the largest doll, and the oldest term: it dates back to the 1950s. It refers to any technique that lets a machine carry out a task that would seem to require intelligence.
The point many miss: it involves no learning whatsoever. A chess program from the 1990s, exploring millions of moves according to rules hand-written by humans, is artificial intelligence. So is a recommendation system based on rules like if the customer bought A, suggest B.
This category is sometimes called symbolic AI, and it dominated for decades. It still works very well for problems where the rules are known and can be stated.
Level 2: machine learning
Machine learning is a subset: these are the techniques where the machine infers the rules from examples rather than having them dictated to it.
The difference is fundamental. To recognise a cat, the classic approach would require describing what a cat is: pointed ears, whiskers, fur. That is impossible to do properly. Machine learning flips the problem: you show a hundred thousand labelled photos, and the system finds the regularities on its own.
Many systems you encounter daily fall under this level without being deep learning: bank fraud detection, churn prediction, spam filtering.
Level 3: deep learning
Deep learning is a subset of machine learning that uses neural networks with many layers.
The word deep simply refers to the number of stacked layers. Each layer transforms the information it receives and passes it on to the next. On an image, the first layers pick out edges, the next ones shapes, the last ones whole objects. No one programmed this hierarchy: it emerges from training.
This is what unlocked, from the 2010s onwards, image recognition, translation and speech recognition. The parameters we constantly hear about are the settings of these layers.
Yet another subset: deep learning systems that produce new content rather than classify or predict. This is where language models and image generators sit. In other words, what we talk about every day in these pages occupies the smallest of the four dolls. It is a fraction of AI, and not the most widely used by volume in the economy.
Why this distinction matters
Three practical reasons.
For reading announcements. When a company says it uses AI, that could mean a cutting-edge language model or a rule-based system written fifteen years ago. Both claims are accurate and they are not worth the same.
For choosing the right method. Many problems are better solved with a simple statistical model than with a large language model: faster, cheaper, more reliable, and explainable. Using a generative model to sort documents into three categories is often a poor technical choice.
For understanding the debates. Questions of opacity mainly concern deep learning. A decision tree, which falls under level 2, is perfectly readable. Opacity is not a property of AI in general; it is a consequence of one specific technique.
What to remember
Remember the nesting: all generative AI is deep learning, all deep learning is machine learning, all machine learning is AI. The reverse statements are false.
And keep in mind that most AI deployed in the world is not generative. It sorts job applications, assesses risks, optimises delivery routes, detects anomalies. These systems have been making decisions that affect you for far longer than the models we talk about every day, and they make far less noise.