In short
AI takes many forms: some systems predict outcomes, some create content, and some act in the physical world. Machine learning is a way to build systems that learn patterns from examples. Prediction, generation, conversation, and physical action describe different jobs those systems can perform.
Why it matters
The same label appears on a chatbot, a spam filter, and a robot, but that doesn’t make them interchangeable. You might want a category for an incoming message, an explanation of a topic, or a machine that moves a package. Understanding the job helps you ask useful questions about the input, the output, and the consequences of an error. It also makes a headline about “AI” easier to unpack.
How it works
Methods: how a system learns
Supervised learning uses examples with labels: the answers or categories the system is supposed to learn to predict. Unsupervised learning looks for structure without the same kind of supplied target. Reinforcement learning uses rewards to learn which actions help achieve an objective. Tom Mitchell’s machine-learning report describes these different learning problems.
Self-supervised learning creates learning targets from the data itself, such as predicting a missing part of a sequence. A neural network is a model built from connected numerical operations with adjustable parameters. Deep learning uses networks with multiple layers. These terms describe how learning happens, and they can appear together in the same system.
Capabilities: what it does
Classification assigns a category. Regression estimates a numerical value. Generation produces content such as text or an image. A system can classify a message and then generate a response, so a single product can contain both predictive and generative components. The introductory chapters of Dive into Deep Learning explain these learning tasks and their basic ingredients.
Language, vision, and speech describe kinds of information. Multimodal systems work across more than one kind. These labels help you understand the input, but they don’t establish that every feature is available in every application.
Systems: how it is put to work
Retrieval supplies reference material. An agent combines a model with tools and a process for taking steps. A robot adds sensing and physical control. Ask what goes in, what comes out, and who checks it. That question remains useful even when the marketing vocabulary changes.
Example
Illustrative example: this scenario is invented.
Sam is a learner trying to make sense of a course advertisement. It calls an assistant “generative,” a recommendation system “predictive,” and a classroom robot “intelligent.” They describe each one by its job.
The assistant can draft an explanation. The recommendation system can rank suggested lessons from available information. The robot can act in the room, which introduces physical constraints. Sam realizes that the labels don’t tell them whether any of those systems gives good advice. They still need to examine the task and its evidence.
What it doesn’t mean
- These categories don’t form a ladder toward human intelligence. Different systems solve different problems.
- A system that generates text can also use prediction and retrieval. The categories overlap.
- A handwritten software rule isn’t automatically machine learning. Some automation follows explicit instructions without learning a model from data.
Try it
For about ten minutes, ask an assistant to classify these sample tasks: draft a welcome message, estimate next week’s demand, group similar articles, and move a box around an obstacle.
Ask it to name the likely input, output, learning approach, and surrounding system for each. Tell it to identify alternatives when the task doesn’t determine the method. Review whether it confuses the job with how the system was trained. The useful result is a clearer map of the vocabulary.
Keep going
- Models, assistants, and products
- Predictive AI and structured data
- Generative AI is good at possibilities. You provide the standard.
- Explain it three ways
Sources
- The Discipline of Machine Learning, Carnegie Mellon University, 2006. Checked October 10, 2026.
- Dive into Deep Learning: Introduction, Cambridge University Press, 2023. Checked October 10, 2026.