Where the term "artificial intelligence" comes from
The term "artificial intelligence" was coined in 1956 at a conference in Dartmouth – it meant machines performing tasks that would require human intelligence. That definition has stayed deliberately broad: a simple 1990s chess program falls under it just as much as a modern language model. That breadth is exactly why the term is used so loosely today – and why it's worth breaking it into four clearly distinct layers.
The four layers that actually matter in practice
Layer 1: AI – the umbrella term
AI is the big umbrella over everything: any system that performs tasks that would require human intelligence – even if it never learns anything.
Layer 2: Machine learning – learning instead of programming
Machine learning (ML) is a subfield of AI with one decisive property: the system learns patterns from data instead of a human specifying every rule individually.
Layer 3: Deep learning – learning with many layers
Deep learning is a subfield of machine learning: artificial neural networks with many stacked layers that recognize increasingly abstract patterns – at the cost of significantly more data and compute.
Layer 4: Generative AI and LLMs – creating content
Generative AI is a subfield of deep learning that doesn't just recognize patterns but creates new content. Large language models (LLMs) like GPT or Claude are specialized in understanding and generating language. An LLM is therefore all four layers at once – but not every AI is an LLM.
Why this distinction matters for you as a decision-maker
The practical payoff shows up the moment a vendor says: "Our system uses AI." That statement is almost always true – and almost always meaningless, because it can apply to any of the four layers. The decisive follow-up question: "Which layer – fixed rules, a learning model, or a generative language model?" A rule-based system always behaves exactly the same way and can be fully traced. A generative language model can handle vague inputs creatively – but can also produce confidently wrong answers (see the "Limits & Risks" module).