What Is Artificial Intelligence (AI)?

A working definition

Artificial Intelligence is software that learns patterns from data and uses them to make decisions, predictions, or new content — instead of following rules a programmer wrote by hand. That one distinction is what separates AI from ordinary software. A traditional program does exactly what it was told; an AI system infers what to do from examples.

There is no single agreed definition, so it helps to hold two at once:

• The field: the branch of computer science concerned with building systems that perform tasks normally requiring human intelligence — perceiving, reasoning, learning, deciding.
• A system: a specific model or pipeline that ingests data, finds patterns in it, and acts on those patterns toward a defined goal.

Everything else on this page is detail underneath those two ideas.

Why it matters

AI is now embedded in systems most people touch daily — search ranking, recommendation feeds, spam filters, fraud detection, voice assistants, and the chat tools that draft text or code. Its practical value comes down to a few things it does well: it processes far more data than a person can, it handles repetitive work without fatigue, it personalizes output at scale, and it surfaces patterns that would otherwise stay hidden. None of that makes it a substitute for judgment — but it changes what a small team can realistically take on.

The building blocks

“AI” is an umbrella. The parts that matter most in practice:

• Machine Learning (ML) — the core idea: algorithms that improve at a task by learning from data rather than being explicitly programmed. Nearly everything below is a form of ML.
• Deep Learning (DL) — ML built on multi-layered neural networks. It excels at messy, unstructured data — images, audio, language — and underpins most of today’s frontier systems. See Machine Learning vs. Deep Learning.
• Natural Language Processing (NLP) — the subfield that lets machines read, interpret, and produce human language. See Natural Language Processing.
• Computer Vision — interpreting images and video: detecting objects, reading scenes, classifying content.
• Generative AI — models that produce new text, images, audio, code, or video rather than just classifying existing data. See Generative AI.

How a system runs

Most AI systems, whatever their subfield, move through the same loop:

1. Input — data goes in, structured (tables, sensor readings) or unstructured (documents, images, audio).
2. Processing — a trained model matches that input against the patterns it learned, producing an internal representation.
3. Output — the system returns a result: a prediction, a classification, a ranking, or generated content.
4. Feedback — corrections and new data feed back in, refining the model over time.

Training (learning the patterns) and inference (applying them to new input) are separate phases. Training is expensive and periodic; inference is the cheap, constant work of actually serving answers.

Narrow, general, and beyond

AI is commonly placed on a capability spectrum. Every system in production today is Artificial Narrow Intelligence (ANI) — very good at a bounded task, with no general awareness. Artificial General Intelligence (AGI), a system that matches human flexibility across domains, remains a research goal, not a product. Artificial Superintelligence (ASI) is purely speculative. Modern chat assistants can feel general because they cover many tasks within software, but that breadth is still narrow intelligence wrapped in a flexible interface. The full breakdown lives in Types of AI.

How AI systems are built

Real deployments are layered, from raw compute up to the app a user touches: infrastructure, data and pipelines, a pre-trained foundation model, an orchestration layer that connects the model to tools and knowledge, and the application on top. Understanding those layers tells a team where a given tool fits and what still has to be built around it. The full architecture is covered in The AI Stack.

The human role

AI amplifies people; it does not replace judgment. The durable human contributions — strategy, ethics, taste, accountability, knowing which problem is worth solving — are exactly the things models don’t supply. As systems grow more autonomous, the operator’s job shifts from doing each step to directing and reviewing the systems that do them: setting the goal, defining guardrails, and checking the output before it ships. More capability raises the value of good oversight, it doesn’t remove the need for it.

What AI still can’t do

• It needs data. Quality and coverage of the training data cap what the system can do; thin or skewed data produces a thin or skewed model.
• It can be confidently wrong. Generative models “hallucinate” — producing fluent output that is simply false. Fluency is not accuracy.
• It has no common sense. There is no genuine understanding, empathy, or reasoning about the physical and social world underneath the statistics.
• It inherits bias. Patterns in the training data, including harmful ones, get reproduced and sometimes amplified.
• It raises unsettled questions around ownership, privacy, and misinformation that are legal and ethical, not technical.

These aren’t reasons to avoid AI — they’re the reasons human review stays in the loop. The rest of this section builds out each idea, starting with how AI got here and the types of AI.