What do people mean by AI?

A photo editor, a music service, and a chatbot can all use artificial intelligence. One changes a picture, another suggests a song, and the third can help write an invitation. Calling all of them AI gives them a shared label. It leaves quite a bit to explain.

Artificial intelligence, or AI, covers computer systems that use information to make predictions, recommendations, decisions, or new content. The OECD's explanation of AI systems gives the more precise definition. For now, start with the job a system does.

A spam filter can estimate which messages are unwanted. A music service can recommend a song. A speech-recognition system can turn spoken words into text. An assistant can draft an invitation. One product may combine several of these capabilities.

Some automation follows rules written directly by a person. A reminder that appears at a time you set does not need AI.

Many modern AI systems use machine learning: developers train a model on examples so it can learn patterns useful for a task. A trained model is the part of the system that applies those patterns. During training, numerical settings inside it are adjusted to improve its performance on the training task. That process does not give it a person's experiences or judgment.

Imagine sorting unwanted messages. A fixed rule might block every message containing a particular phrase. A learned model can use patterns across many features of a message. Both can make mistakes. The important question is whether the approach works for the messages it actually encounters.

These are simplified examples, not descriptions of the implementation inside a particular email service.

Generative AI produces content, such as text, images, audio, video, or computer code. An assistant drafting an invitation uses this capability. A system that sorts existing messages into folders does a different job.

Creating content and finding information are also different operations. A convincing paragraph can contain an invented detail. A product may use search to gather information before generating its answer, but generation alone does not establish that the answer is correct.

The NIST guide to generative-AI risks examines unreliable information and harmful bias alongside other risks. Keep it for a deeper question; you can follow this path without reading a technical report first.

Consider these fictional features. You can do this exercise without opening an AI tool.

• A reminder appears at 6 p.m. because you set that time.
• A filter estimates whether a message is spam.
• An assistant writes an invitation from your notes.

Which follows a fixed rule? Which makes a prediction? Which generates content?

The reminder can follow the rule you set. The spam filter makes a prediction. The assistant generates text. These categories describe the examples, not every feature a product could contain. You cannot always tell how software works from its appearance.

For an exact reminder, ordinary automation may be all you need. For an invitation, writing it yourself, adapting a template, and asking an assistant are all reasonable options. AI is useful when its result and the effort to check it fit your task.