Supervised learning
In short
Supervised learning is a way of training an AI model on examples that already have the right answer attached, called a label. The model is adjusted until it can predict that label for a new example it has not seen.
- Also called
- Supervised machine learning
- Topic
- How models are trained
- Level
- Beginner
- First described
- A paper that coined the name was not identified. Sutton and Barto trace the method to systems studied by Widrow and Hoff in 1960 and by Rosenblatt in 1962. Those authors wrote in the language of rewards and punishments, but the systems they studied learned labelled examples for pattern recognition2.
Supervised learning is the kind of machine learning in which a model learns to predict a label, a correct output that has already been attached to each example. The US standards body NIST describes those labels as explicit and often produced by people1. In NIST's terms, two common forms are classification, where the predicted label is a category, and regression, where it is a number1.
Richard Sutton and Andrew Barto, in their textbook on reinforcement learning, define it as learning from a training set of labelled examples given by an external supervisor. Each example pairs a situation with the correct action or category. The aim is to generalise, so the system responds correctly in situations that were not in that set2.
It is not unsupervised learning, which looks for structure in data with no labels, and it is not reinforcement learning, where the system is never shown the correct action and instead tries things in order to increase a reward2. NIST notes that a supervised model is usually tested before it is put into use1.
The method sits inside products that do not advertise it. Google's BERT model is pre-trained on raw text, then fine-tuned on labelled data for tasks such as question answering3. OpenAI's November 2022 ChatGPT release began the same way: contractors wrote example conversations, playing both the user and the assistant, and that set was used for supervised fine-tuning before a later stage that used reinforcement learning4. OpenAI's page now says ChatGPT has changed substantially since that post4.
Example
A mail app can be shown thousands of messages that people have already marked as spam or not spam. Supervised learning is the step that uses those marks to sort the next message that arrives.
Why it matters
It is how models learn to put things in a category or predict a number1, and it was the first training stage of the ChatGPT model OpenAI introduced in November 20224.
Questions people ask
Is supervised learning the same as reinforcement learning?
No. In supervised learning an external supervisor supplies the correct action or category for each example. In reinforcement learning the system is not shown the correct action. It tries actions and is scored with a reward2.
Does supervised learning need people to label the data?
Often. NIST defines it as learning to predict explicit labels or output values, and says those labels are often human-generated1. Without an explicit label to predict, the setup is not supervised learning in that definition1.
Was ChatGPT trained with supervised learning?
The November 2022 release was, as a first stage. Contractors wrote example conversations, and OpenAI used them for supervised fine-tuning, then trained further with reinforcement learning from human feedback. OpenAI says ChatGPT has changed substantially since that announcement4.
Related terms
Tools that use it
Sources
- NIST AI 100-2e2025 (March 2025), Adversarial Machine Learning, glossary entries for supervised learning, classification and regression, and section 2.1.1
- Sutton and Barto (2018), Reinforcement Learning: An Introduction, 2nd edition, MIT Press, sections 1.1 and 1.7
- Devlin et al. (2018), BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, arXiv:1810.04805
- OpenAI (30 November 2022), Introducing ChatGPT