Supervised vs unsupervised learning: the four types

Open almost any introduction to machine learning and you will meet the same three categories: supervised learning, unsupervised learning, and reinforcement learning. The taxonomy is useful and it has organized the field for decades. It also has a problem, which is that the way nearly every important AI system of the last several years was actually trained does not fit cleanly into any of the three. The types of machine learning are distinguished by what kind of signal a model learns from, an answer that is given, no answer at all, or a reward that must be earned, and the reason the classic three-way split now feels dated is that the paradigm behind almost all modern AI, self-supervised learning, manufactures its own answers out of unlabelled data, capturing the scale of unsupervised learning with the precise training signal of supervised learning. That move is what removed the bottleneck that had constrained the field since its beginning.

This guide explains each type in plain terms, what it is good and bad at, the fourth type that the textbooks are still catching up to, why it mattered so much, and how the systems you actually use combine all of them in sequence rather than choosing between them. By the end, the question "is a chatbot supervised or unsupervised?" should dissolve into a better one.

The question that organizes everything

Before the categories, the principle behind them. Every machine learning system improves by adjusting itself in response to some signal telling it how it is doing. The types of machine learning are really answers to one question: where does that signal come from?

In supervised learning, the signal comes from a correct answer supplied by a human. In unsupervised learning, there is no answer at all, and the signal comes from structure the model finds in the data itself. In reinforcement learning, there is no answer supplied in advance, but the model receives a reward or penalty after acting, and must work out for itself which of its choices earned it. Holding that single question in mind makes the categories feel less like arbitrary buckets and more like the small number of truly different ways a machine can be told it is on the right track.

Supervised learning: learning from answers

Supervised learning is the most familiar and still the most widely deployed type. The model is trained on examples that come with correct answers attached, pairs of inputs and desired outputs, and its job is to learn the mapping between them so it can produce the right output for inputs it has never seen. The name comes from the idea of a supervisor or teacher providing the answers, which are called labels.

Two task shapes dominate. Classification predicts a category: is this email spam, does this scan show a tumour, which of these ten digits is this handwriting. Regression predicts a continuous number: what will this house sell for, how many units will we ship next quarter. Between them they cover an enormous share of practical machine learning, from fraud detection to medical triage to demand forecasting.

The strength of supervised learning is precision. Because the model is told exactly what the right answer is, its training signal is sharp, and its performance is straightforward to measure by checking predictions against held-out labels. Its weakness is the labelled data itself. Labels are produced by people, which makes them slow, expensive, and often requiring of expertise, since annotating medical images or legal documents is not work you can crowdsource cheaply. This is the label bottleneck, and for most of the field's history it set a hard ceiling on how large a supervised model's training set could be. You could only ever learn from as much data as you could afford to have humans annotate.

Unsupervised learning: learning without answers

Unsupervised learning removes the labels entirely. The model is given raw data with no correct answers and asked to find structure in it. There is no target to predict, so the goal shifts from prediction to discovery, from "what is the answer" to "what is in here."

Its main tasks reflect that. Clustering groups similar items together without being told what the groups should be, which is how customer segments emerge from purchase data or how documents sort themselves by topic. Dimensionality reduction compresses complex data into fewer variables while keeping its essential shape, which helps with visualization and with feeding data into other models. Anomaly detection learns what normal looks like and flags what departs from it, which is how unusual transactions or failing equipment get caught.

The appeal of unsupervised learning is that unlabelled data is abundant and nearly free, so it can work at scales supervised learning cannot reach. Its difficulty is the mirror image of its freedom: with no correct answer, there is often no clear way to tell whether the result is any good. If a clustering algorithm proposes five customer segments, nothing in the data confirms that five is right or that the boundaries are meaningful, and judging the output usually falls back on human interpretation. Unsupervised learning is powerful for exploration and much harder to evaluate.

Reinforcement learning: learning from consequences

Reinforcement learning is different in kind from both. There are no labelled examples and no fixed dataset. Instead an agent acts in an environment, receives a reward or penalty for the outcome, and gradually learns a policy, a strategy for choosing actions, that maximizes reward over time. It learns by doing rather than by being shown, which is why it fits problems that are sequences of decisions rather than single predictions: playing games, controlling robots, managing resources, routing traffic.

Its distinctive difficulty is that the feedback is thin and delayed. A move in a game may only prove to have been a mistake twenty moves later, and the agent has to work out which of its many choices deserves the credit or blame, a problem known as credit assignment. Rewards can also be sparse, arriving rarely, and the agent must balance exploiting what already works against exploring what might work better. Reinforcement learning is the most conceptually distinct of the classic three, and it has become newly central to modern language models, which is a point worth returning to.

Self-supervised learning: the one that broke the split

Now the fourth type, and the reason this article exists. Self-supervised learning takes unlabelled data and manufactures a supervised problem out of it, by hiding part of the input and asking the model to predict that part from the rest. Cover the next word in a sentence and ask what it is. Mask a patch of an image and ask what belongs there. No human ever wrote a label, yet there is a definite correct answer to check against, because the answer was in the data all along before it was hidden.

This is why it does not fit the old taxonomy. It uses unlabelled data, like unsupervised learning, but it has a precise right answer and a sharp training signal, like supervised learning. It is often filed as a subtype of unsupervised learning, and it is sometimes described as automated supervised learning, and both descriptions are defensible, which is itself the point: it sits between two categories the textbook treats as opposites. The important thing is not where to file it but what it gets you, which is the best property of each.

Why self-supervision changed everything

The significance of that combination is hard to overstate, and it explains the shape of the last several years of AI. The binding constraint on machine learning was never really the algorithms; it was the labels. Supervised learning gave the strongest training signal but could only ever be as large as a labelling budget allowed. Unsupervised learning could consume unlimited data but extracted a weaker, fuzzier signal from it.

Self-supervision broke that trade-off. Because the supervision is generated from the data's own structure, every piece of raw text, every image, every recording becomes usable training material with a crisp objective attached. The entire internet turns into a labelled dataset that no one had to label. This is precisely what made large language models possible: an LLM is pretrained by predicting the next token in ordinary text, over and over, at a scale of trillions of examples that could never have been annotated by hand. It also explains why the scaling laws mattered so much, because a law saying "more data makes models better" is only useful if you can actually get more data, and self-supervision is what made data effectively unlimited. The same move transformed computer vision, where models now learn strong general representations from unlabelled images before ever seeing a labelled one.

Semi-supervised learning and the practical middle

One more category rounds out the picture, and it is the pragmatic compromise most organizations actually live in. Semi-supervised learning uses a small amount of labelled data alongside a much larger pool of unlabelled data, letting the labels provide direction while the unlabelled mass provides scale. It suits the extremely common real-world situation where you have plenty of raw data and can only afford to annotate a fraction of it. In practice, the workflow that dominates modern applied AI is a version of this logic: take a model that learned general structure from vast unlabelled data, then adapt it with a comparatively tiny labelled set for your specific task, which is what fine-tuning does.

How modern AI uses all of them

This is where the "which type is it?" question dissolves. A current AI assistant is not an example of one paradigm; it is a pipeline that runs through several in sequence, and each stage does something the others cannot.

It begins with self-supervised pretraining, predicting the next token across an enormous body of text, which is where the model acquires its knowledge of language, facts, and reasoning patterns. It continues with supervised fine-tuning, where the model is trained on a much smaller set of human-written instruction-and-response pairs, teaching it to behave as a helpful assistant rather than a raw text predictor. It concludes with reinforcement learning, where the model generates responses that are ranked by humans or checked by automated verifiers, and learns to produce the kind of output that scores well, which is how both preference alignment and the recent leaps in reasoning were achieved.

So the honest answer to whether a chatbot is supervised or unsupervised is that it is self-supervised, then supervised, then reinforcement-trained, in that order, and each stage contributes something distinct. The paradigms are not competing options to choose between. In modern systems they are stages of a single training pipeline.

Choosing between them

For an actual project, the choice usually resolves through a short sequence of questions. Do you have labelled examples, or can you afford to create them? If yes, and you need to predict a specific category or value, supervised learning is the direct answer. Do you have data but no labels, and are you trying to understand what is in it rather than predict something specific? That is unsupervised territory. Is your problem a sequence of decisions where you can define a reward, rather than a one-shot prediction? That points to reinforcement learning. Do you have a large amount of raw unlabelled data and want a model that understands its general structure before you specialize it? That is the self-supervised route, and in practice it usually means starting from an existing pretrained model rather than pretraining one yourself. And if you have a lot of data and a little labelling budget, the semi-supervised pattern of pretrain-then-fine-tune is almost always the efficient path.

The honest state of the taxonomy

The three-way split remains a good teaching map, and it is not wrong so much as incomplete. Its categories were drawn around the question of whether a human supplies the answers, which was the right organizing question when human labelling was the binding constraint. Self-supervision changed what that constraint is, so the boundary that mattered most has moved. The useful modern framing is not which single box a system belongs in but which combination of learning signals it was built from, and in what order. Expect introductions to keep teaching three types for a while, because the map is simple and mostly serviceable, while remembering that the paradigm which actually produced the current era of AI is the one that sits in the gap between two of them.

The short version

The types of machine learning are distinguished by where a model's training signal comes from. Supervised learning uses labelled examples with correct answers supplied by humans, covering classification and regression, and it offers a sharp signal and easy evaluation at the cost of expensive labels. Unsupervised learning uses unlabelled data to find structure through clustering, dimensionality reduction, and anomaly detection, scaling cheaply but with no correct answer to check against, which makes it hard to evaluate. Reinforcement learning has an agent act in an environment and learn a policy from rewards and penalties, suiting sequential decision problems but facing delayed and sparse feedback. The fourth type, self-supervised learning, hides part of the input and asks the model to predict it, manufacturing supervision from unlabelled data, and so combines unsupervised scale with a supervised-quality signal. That combination removed the labelling bottleneck, turned raw text and images into effectively unlimited training data, and is what made large language models and modern vision models possible. Semi-supervised learning mixes a small labelled set with a large unlabelled one. Modern AI assistants use several in sequence: self-supervised pretraining, then supervised fine-tuning, then reinforcement learning.

The idea to hold onto is that the categories of machine learning are really answers to a single question, where the training signal comes from, and the reason modern AI outgrew the classic three is that self-supervised learning found a way to generate a precise training signal from unlabelled data, which lifted the ceiling that human labelling had always imposed. The systems you use are not one type; they are a sequence of them.

Common questions

What is the difference between supervised and unsupervised learning? The difference is whether the training data comes with correct answers. Supervised learning uses labelled data, where each example is paired with the output the model should produce, and the model learns to map inputs to those known outputs for tasks like classification and regression. Unsupervised learning uses unlabelled data and asks the model to find structure on its own, through tasks like clustering, dimensionality reduction, and anomaly detection. Supervised learning is aimed at prediction and is easy to evaluate against known answers but requires expensive human labelling. Unsupervised learning is aimed at discovery and scales cheaply on abundant raw data, but with no correct answer available, judging the quality of the result is much harder.

What are the types of machine learning? The classic taxonomy lists three: supervised learning, which learns from labelled examples; unsupervised learning, which finds structure in unlabelled data; and reinforcement learning, in which an agent learns a strategy by acting and receiving rewards or penalties. Two more are essential in practice. Self-supervised learning generates its own labels by hiding part of the input and predicting it from the rest, and it is the paradigm behind large language models and modern vision systems. Semi-supervised learning combines a small labelled set with a large unlabelled one. Modern AI systems typically combine several of these in sequence rather than using just one.

What is self-supervised learning? Self-supervised learning trains a model on unlabelled data by creating a prediction task from the data's own structure, hiding part of an input and asking the model to reconstruct it from the remaining context. Predicting the next word in a sentence or filling in a masked region of an image are the canonical examples. No human writes labels, yet each example has a definite correct answer, because the hidden portion was present before it was hidden. This gives it the scale of unsupervised learning together with the precise training signal of supervised learning, and it is how large language models are pretrained.

Is ChatGPT supervised or unsupervised? Neither on its own, because a modern assistant is trained in stages that use different paradigms. It begins with self-supervised pretraining, predicting the next token across vast amounts of unlabelled text, which is where it acquires knowledge of language and facts. It then goes through supervised fine-tuning on a much smaller set of human-written instruction-and-response pairs, which teaches it to behave as an assistant. Finally it is refined with reinforcement learning, where responses are ranked by people or checked by automated verifiers so the model learns to produce output that scores well. The paradigms are stages of one pipeline rather than alternatives.

Why is self-supervised learning so important? Because it removed the constraint that had limited machine learning since the beginning: the cost of labelled data. Supervised learning produced the strongest training signal but could only be as large as a human labelling budget allowed, while unsupervised learning scaled freely but yielded a weaker signal. Self-supervision generates supervision from the structure of the data itself, so raw text, images, and audio become usable training material with a precise objective attached and no annotation required. This effectively turned the internet into a labelled dataset, which is what made pretraining at the scale of trillions of examples possible, and therefore what made foundation models and large language models feasible.

What is semi-supervised learning? Semi-supervised learning uses a small amount of labelled data together with a much larger pool of unlabelled data. The labels supply direction about what the model should predict, while the unlabelled data supplies the scale needed to learn general structure. It fits the very common situation where raw data is plentiful but annotation is expensive, so only a fraction can be labelled. The pattern that dominates applied AI today follows the same logic: start from a model that has learned broadly from unlabelled data, then adapt it to a specific task with a comparatively small labelled dataset, which is what fine-tuning a pretrained model amounts to.

Which type of machine learning should I use? It depends on your data and your goal. If you have labelled examples and need to predict a category or a numeric value, use supervised learning. If you have unlabelled data and want to understand its structure rather than predict a specific target, use unsupervised methods such as clustering or anomaly detection. If your problem is a sequence of decisions with a definable reward rather than a one-shot prediction, reinforcement learning fits. If you have abundant raw data and want general understanding before specializing, the self-supervised route applies, which in practice usually means starting from an existing pretrained model. With plenty of data and a small labelling budget, pretrain-then-fine-tune is normally the most efficient path.