self-supervised learning (SSL)

advancedwidespread

A machine learning method in which the model learns without human labels: it creates tasks from raw data itself, typically by trying to guess hidden parts of the data.

In plain terms

Imagine learning a foreign language by covering up words in a book, trying to guess them, and then checking if you were right. No teacher stands beside you; the book itself provides the correct answer. Self-supervised learning works exactly like this: the machine creates its own tasks and answer keys from raw data.

This is a big deal because with traditional supervised learning, people must label data one by one: this image is a cat, this email is spam. This is slow and expensive. But from unlabeled text, images, and audio, almost unlimited amounts are available.

Today's large language models, including those behind chatbots, learned fundamentally this way: they practiced on vast amounts of text to predict the next word. In the process, they absorbed not just words, but also grammar, facts, and relationships.

How it works

Developers gather a huge amount of raw data, such as text, images, or audio recordings, without labels.
The system hides part of the data: it masks a word, cuts off the end of a sentence, or removes a section from an image.
The model guesses what the hidden part might have been, and the system automatically compares the guess with the original, which is present in the data.
Based on the error, the model slightly adjusts its internal settings, and this is repeated countless times until its guesses become increasingly accurate.
The general knowledge gained this way can later be adapted to a specific task with few labeled examples, such as translation or medical image analysis.

An everyday example

When your phone keyboard suggests the next word, or a chatbot responds fluently to your question, you are seeing knowledge acquired through self-supervised learning in action. The model practiced on vast amounts of text how a sentence continues, and no one pre-labeled the correct answers for it.

Who is affected

everyoneMuch of the knowledge behind the chatbots, translators, and search engines used daily comes from this training method.

developerYou can build on a pre-trained model, so only a fraction of labeled data is needed for your specific task. → Code

entrepreneur, executiveYour company's unlabeled documents, letters, and recordings can become useful training material without expensive manual labeling.

patient, doctorThere are few expert-labeled examples among medical images, so diagnostic models are often pre-trained on unlabeled scans. → Health

A common misconception

MythSelf-supervised learning means the machine learns and develops on its own without human intervention.

In realityOnly the labeling is left out: humans still determine the data, the type of task to be guessed, the model's architecture, and the training process. The finished model does not continue learning on its own during use.

Limits and risks

The model learns what is in the data, so it also adopts errors, biases, and outdated information.
A good guess is not the same as understanding: the model provides a probable continuation, which sometimes sounds confident but is a false statement.
Training requires enormous computing capacity and energy, so only a few well-funded organizations can produce the largest models.
Massive, uncurated datasets may contain copyrighted material and personal data, leading to legal disputes.

Similar, but not the same

supervised learning – There, the model learns from labels provided by humans; here, the correct answer is contained in the raw data itself.
unsupervised learning – This is the broader category: any learning that does not use human labels, such as clustering data without an answer key. Self-supervised learning is often considered a subtype, where the model generates verifiable puzzle-like tasks for itself.
reinforcement learning – There, the model tries actions and receives rewards or penalties for them; here, there is no reward, only the goal of returning the hidden data portion.
semi-supervised learning – There, labeled and unlabeled data are used together, while here no human labels are needed for pre-training. Its English abbreviation is also SSL, so the two are easily confused.

Related concepts

Broader

machine learning, deep learning, representation learning

Narrower

masked language modeling, next-token prediction, contrastive learning

On the edge

pre-training, fine-tuning, foundation model, large language model, training data

A brief history

The idea's origins go back decades, but the method became dominant in the 2010s: word2vec, introduced in 2013, learned word meanings from context, while BERT and GPT, presented in 2018, showed that guessing masked or next words provides general linguistic knowledge. The term was popularized by Yann LeCun.

Where things stand

By the end of 2026, this is the default method for pre-training large language models, as well as many image, audio, and video models. However, finished systems are rarely built exclusively on this: pre-training is usually followed by fine-tuning based on human feedback, while image generators typically learn from image-caption pairs. One main research direction is developing more efficient variants requiring less data and energy, while the other focuses on models learning from video and the physical world.

Status: 2026 November

Also known as

self-supervised learning · self-supervised machine learning · self-supervised learning · self-supervised learning · self supervised learning · SSL · self-supervised pretraining · self-supervised pretraining

Sources

Devlin et al.: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (arXiv, 2018)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.