In simple terms
Machine learning builds computer systems that adapt and learn from data in order to get more accurate. That is close to the definition NIST uses.
The contrast is with ordinary software, where a person writes the rules. In machine learning the rules are worked out from examples.
How it works
Start with data and a model holding adjustable numbers called parameters. The model makes a prediction, the prediction is compared with the correct answer, and the parameters shift to narrow the gap. Repeat that a very large number of times and you have training.
Supervised learning trains on labeled data, meaning input examples paired with correct outputs. It is what sits behind image classification, spam detection and speech recognition.
Unsupervised learning works on unlabeled data and looks for structure, such as groups of similar customers. Reinforcement learning learns from rewards and penalties as it acts.
Using a trained model on new data is a separate step, called inference, with its own hardware and energy costs.
A model is only as good as its examples. If the data is narrow or skewed, the model repeats that, which is why the European Union, the United States and others are writing rules specifically about training data, testing and disclosure. Those rules differ, so the same model can face different duties in different markets.
Why it matters
Machine learning now sits inside fraud detection, medical imaging, translation, weather forecasting and credit assessment, usually without being named.
Because the rules are learned rather than written, it can be hard to say why a particular output appeared. That difficulty is what drives demands for explainability and audit.
Training large models also consumes real resources, in electricity, specialized chips and data, which shapes who can build them.
Where you’ll see it
- Spam filters and bank fraud alerts.
- Photo apps that group faces and search by subject.
- Voice assistants, dictation and live captions.
- Recommendations in shops, maps and streaming services.
Example
A hospital feeds a model thousands of chest X-rays, each labeled by a radiologist. The model learns which patterns go with which label, then flags likely cases in new images for a human to check.
Often confused with
Artificial intelligence is the wider field. Machine learning is the approach inside it that learns from data, and deep learning is a subset of machine learning built on many layered neural networks.
Key facts
- NIST defines machine learning as the development and use of computer systems that adapt and learn from data with the goal of improving accuracy.1
- Supervised learning trains models on labeled data, meaning input examples paired with correct output answers.2
- Training works by adjusting the internal parameters of a model to reduce the gap between its predictions and the correct answers.2
- Common supervised learning applications include image classification, spam detection and speech recognition.2
- Supervised learning depends on access to large datasets in which every example carries a correct label.2
Related concepts
In the news
Quick checkWhat makes machine learning different from ordinary software?Show answer
In ordinary software a person writes the rules. In machine learning the system derives the rules from data by adjusting its parameters until its predictions match known answers.
Sources
- NIST, Computer Security Resource Center glossary. Machine learning, definition drawn from NIST SP 800-55v1. Undated (accessed 18 September 2026)
- Stanford Institute for Human-Centered Artificial Intelligence. What is supervised learning?. Undated (accessed 18 September 2026)
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed September 21, 2026Researched and drafted with AI assistanceReport an issue