AI, Machine Learning, Deep Learning And Generative AI: How They Fit Together

Vendor brochures, board papers and news stories use “AI”, “machine learning” and “generative AI” interchangeably. That causes real confusion when a team tries to decide which tools need review or which risks apply. The terms describe different things, and most of them sit inside one another like a set of nested boxes.

This article explains each term with an everyday example, shows how they relate, and clears up the one term that does not fit neatly into the nesting: generative AI. The diagram below is the short version; the sections after it fill in the detail.

Artificial Intelligence

AI is the widest category. Policy bodies define an AI system as one that works out, from the input it receives, how to produce outputs such as predictions, content, recommendations or decisions.¹ The first article in this group, What Is AI?, covers that definition in detail.

AI is also older and broader than machine learning. Expert systems, popular long before today’s models, reasoned over knowledge that specialists had written down by hand rather than learning from data. The EU recitals still count this kind of logic-based or knowledge-based reasoning as inference, so such a system can be AI in the legal sense. What falls outside is software that only carries out rules people wrote, with no reasoning of its own.²

Machine Learning

Machine learning is the part of AI where the system learns its behaviour from examples. Instead of a programmer writing “flag emails containing these 200 phrases”, a spam filter is shown thousands of emails already labelled as spam or not spam and adjusts itself until it separates them well. The learned behaviour lives in internal numbers called parameters. How that adjustment works is covered in How A Machine Learning Model Is Built.

Many familiar business uses of AI are machine learning: credit scoring, demand forecasting, fraud detection, product recommendations and document classification. In each case the system was shaped by historical examples, so the quality of those examples matters as much as the code.

Deep Learning

Deep learning is machine learning that uses artificial neural networks with many stacked layers. Each layer transforms its input a little and passes it on, so early layers pick up simple patterns and later layers combine them into more abstract ones. In a 2015 review in Nature, Yann LeCun, Yoshua Bengio and Geoffrey Hinton described image models whose early layers respond to edges, later layers to combinations of edges and parts of objects, and the deepest layers to whole objects.³

The same review credits deep learning with large gains in speech recognition and image recognition, and it is also the basis of today’s language models.³ In general terms, it tends to need more data and computing power than simpler methods, and because what it learns is spread across a very large number of parameters, its reasoning is harder to explain. NIST notes that organisations may face a trade-off between predictive accuracy and interpretability.⁴

Foundation Models And General-Purpose AI

In a 2021 paper, a large team of researchers introduced the term foundation model for a model “trained on broad data at scale” that can then be adapted to a wide range of tasks.⁵ The paper’s own examples include the language models BERT and GPT-3. In principle, one model, trained once at great expense, can be adapted to drafting emails, summarising contracts, writing code or answering support questions.

The EU AI Act uses a legal term, the general-purpose AI model, that overlaps with this research idea without being identical to it. In plain terms, the Act means a model that can handle many different kinds of task well and can be plugged into many other products, and it expressly includes models trained on very large datasets with self-supervision.¹ The legal definition has its own tests and exclusions, so whether a given foundation model qualifies has to be checked against the Act itself. Providers of these models have their own obligations under Chapter V of the Act, separate from the rules for AI systems built on top of them. Those obligations have applied since 2 August 2025, although models already on the market before that date have until 2 August 2027 to comply.⁶

Where Generative AI Fits

Generative AI is the term that breaks the neat nesting. It describes systems whose output is new content, such as text, images, audio, video or code, rather than a label or a score. That makes it a description of what a system does, not of how it is built.

Chat assistants built on large language models are a clear example of generative AI that runs on deep learning and is built on a foundation model.⁵ Not every foundation model is used generatively, though, and simple generative techniques existed long before deep learning. When the OECD revised its AI definition in 2023, it added “content” to the list of outputs specifically so that generative systems are clearly inside the definition.⁷

Why The Distinction Matters

Each layer brings its own questions. Machine learning raises questions about the training data and whether it reflects the people the system will affect. Deep learning adds questions about explainability. Generative AI adds the risk of fluent but false output, which the Large Language Models group explains, and new attack routes covered in AI Security. Naming the right layer is the first step to asking the right questions.