Foundations
AI, machine learning and generative AI: what's the difference?
These four terms get used as if they mean the same thing. They don't, they nest inside one another like Russian dolls. Getting the picture straight makes every other AI conversation easier.
People throw around "AI", "machine learning" and "generative AI" as if they are the same thing. They are not, but they are related, they fit inside each other like Russian dolls.
Artificial intelligence (AI) is the biggest doll: the whole idea of making computers do things that normally need human thinking.
Machine learning is a doll inside it: one particular way of doing AI, where the computer learns patterns from examples instead of being given hand-written rules. Almost all AI today works this way.
Deep learning is a smaller doll inside machine learning: a powerful technique using brain-inspired "neural networks", and it is behind nearly every recent breakthrough.
Generative AI is the doll that makes things, text, images, music, using deep learning. ChatGPT and AI image tools live here.
So when you use ChatGPT, you are using generative AI, which is a kind of deep learning, which is a kind of machine learning, which is a kind of AI. All four at once.
The four terms describe nested sets, from broadest to narrowest.
Artificial intelligence (the whole field)
Artificial intelligence is the broad field of building systems that perform tasks associated with human intelligence, reasoning, perceiving, understanding language, deciding. It includes old rule-based approaches and modern learning-based ones. It is a goal, not a single technique.
Machine learning (learning from data)
Machine learning is the subset of AI where systems learn patterns from data rather than following explicitly programmed rules. Show a model thousands of labelled examples and it works out the pattern itself. Nearly all AI you hear about today is machine learning.
Deep learning (the powerful engine)
Deep learning is a subset of machine learning that uses multi-layered neural networks. Its superpower is learning useful features directly from raw data (pixels, text) without humans hand-designing them. The 2010s breakthroughs, and today's frontier models, are deep learning.
Generative AI (the part that creates)
Generative AI is the subset focused on creating new content, text, images, audio, code, rather than only classifying or predicting. The large language models behind chatbots are generative AI for text.
Putting it together
AI ⊃ machine learning ⊃ deep learning ⊃ (much of) generative AI. A chatbot is all of these simultaneously. The nesting matters because it tells you why modern AI behaves as it does: it learned from data (so it reflects that data), and it generates likely content (so it can be fluent yet wrong).
The terms denote nested scopes, and precision helps because each layer carries different assumptions and failure modes.
The hierarchy
Artificial intelligence is the encompassing discipline (symbolic and statistical). Machine learning is the data-driven subset that fits functions to data rather than encoding rules; deep learning is the further subset using deep neural networks with end-to-end learned representations (LeCun, Bengio & Hinton 2015). Generative AI cuts across as the class of models that estimate and sample from a data distribution to synthesise novel output, today overwhelmingly implemented with deep learning (autoregressive transformers, diffusion).
Why the distinctions are load-bearing
- AI vs ML: not all AI learns; classical search/planning/logic are AI but not ML. Conflating them obscures that today's dominant paradigm is specifically statistical learning, with its data-dependence and distribution-shift brittleness.
- ML vs deep learning: deep learning's representation learning is what enabled perception and language at scale, but also reduced interpretability and increased data/compute demands.
- Discriminative vs generative: much classic ML is discriminative (estimate a decision boundary); generative models estimate p(x) and can sample, which is why generative AI produces fluent output that is optimised for plausibility rather than truth, the root of hallucination.
The pragmatic point
When someone says "AI", they almost always mean a specific deep-learning, usually generative, model such as an LLM. Naming the actual layer, "this is a generative model sampling likely tokens", predicts behaviour and limits far better than the marketing label "AI", and keeps discussion of capability, reliability and governance grounded in what the system actually is.