AI 101 — AI vs machine learning: what's the difference?
Artificial intelligence is the goal — getting a machine to do things we would call intelligent — and machine learning is the technique the field now overwhelmingly uses to get there: instead of a programmer writing the rules, the system is handed examples and adjusts itself until its answers get better.
The two terms get swapped for each other in almost every story you read, including ours. That is mostly harmless, and occasionally it hides the interesting part. When Justin Fanelli, the US Navy's chief technology officer, put "machine learning and increasingly agentic software" at the top of his shopping list this month, the wording was doing real work: it tells you the system learns from data rather than being told what to do.
Where the split came from
The phrase "artificial intelligence" was coined in the 1955 funding proposal for a summer workshop at Dartmouth College, written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. The workshop ran in the summer of 1956 and is generally treated as the founding event of the field. Its bet was that every feature of intelligence could be described precisely enough for a machine to simulate it — and the early work took that literally, hand-coding rules. Chess engines, medical diagnosis programs, language parsers.
In 1959, an IBM researcher named Arthur Samuel described a different approach. He had written a checkers program that played against itself and adjusted the weights in its own scoring formula as it went, and reported that it learned to play better than the person who wrote it, in eight to ten hours of machine-playing time. He called the technique machine learning. It won for the reason that always wins — it kept working as the problems got harder.
The one-paragraph mental model
Think in nested circles. The outer circle is artificial intelligence — the ambition, which includes systems that never learn anything, such as a rules engine that approves insurance claims by a fixed checklist. Inside it sits machine learning, the family of methods that improve by fitting themselves to data: a spam filter, a demand forecast, a credit score. Inside that sits deep learning, machine learning done with large neural networks stacked in many layers, which is what took over around 2012. And inside that sits the transformer, the network design every major language model is built on — which is why one 2017 paper gets cited in everything you read about GPT, Claude, Gemini, Llama or Qwen.
The OECD's working definition captures the boundary precisely: a machine-based system that, for explicit or implicit objectives, infers from its input how to generate outputs such as predictions, content, recommendations or decisions. Note what it does not say. It requires inference, not learning — a system can qualify as AI and never learn a thing. The European Union's AI Act uses a near-identical definition for the same reason: regulators have to write rules for the whole circle, not just the part that trains.
The apprentice analogy
Old-fashioned AI is a recipe book. Someone who knows how to cook writes down every step in advance, and the kitchen follows it exactly — reliable, inspectable, and useless the moment a dish arrives that the author never imagined.
Machine learning is an apprentice who never sees the recipe. They watch a thousand plates go out, taste the ones that come back, and adjust. Nobody tells them what "properly seasoned" means; they build their own internal sense of it from the corrections. That is the trade: you give up the ability to read the rules, and in exchange you get a system that handles situations nobody wrote down.
Two consequences follow, and they explain most AI news. First, the corrections are the whole game — if the apprentice only ever hears feedback about speed, they will optimise for speed. Second, an apprentice is only as good as what they were shown, which is why the field argues so much about data. The machinery that performs those corrections is backpropagation, the algorithm that works backwards from the error to figure out which connection deserves the blame.
Common misconceptions
"They're the same thing." They overlap heavily in practice and are not synonyms. A rules engine is AI without machine learning; a model that predicts shipping times is machine learning most people would never call AI. When someone claims their product "uses AI," the useful follow-up is whether it learns from data — and from whose.
"Machine learning means the system understands." It means the system found patterns that reduce error on the examples it saw. Whether those patterns correspond to anything real is a separate question, which is why a model can be excellent at benchmarks and still fail in a way that looks careless — the failure mode we went through in what an AI hallucination is.
"AI means human-level." The field has a long habit of moving the goalposts. Optical character recognition and speech recognition were once core AI problems; once they worked reliably, they became ordinary engineering. The phenomenon has a name — the AI effect — and a one-line summary, Tesler's theorem: "AI is whatever hasn't been done yet."
"If it's AI, it keeps learning after launch." Most deployed models do not. They are trained once, frozen and shipped — the weights stop changing the day training ends.
Where to learn more
The OECD's explanatory note on its definition of an AI system is the clearest account anywhere of where the line between AI and not-AI is drawn. Arthur Samuel's 1959 checkers paper is short, readable and where the term comes from. For the moment the field switched, the 2012 AlexNet paper is the hinge: a network with 60 million parameters that won the ImageNet contest with a top-5 error rate of 15.3%, more than 10 percentage points better than the runner-up, and made everyone rebuild around the transformer's ancestors instead of hand-written rules.
Related reading: AI 101 — What is backpropagation? · AI 101 — What is an AI benchmark? · AI 101 — What is a transformer?
If a company says its product "uses AI," should it have to say whether the thing learns — and from whose data? Tell us in the comments.
Sources: OECD — What is an AI system? · EU AI Act, Article 3: Definitions · Arthur Samuel — Some Studies in Machine Learning Using the Game of Checkers (IBM Journal, 1959) · Dartmouth Summer Research Project on Artificial Intelligence, 1956 · Krizhevsky, Sutskever & Hinton — ImageNet Classification with Deep Convolutional Neural Networks (NeurIPS 2012)
