Artificial intelligence is software that, given some input, works out for itself how to produce a useful output: a prediction, a recommendation, a piece of text, a decision. That wording follows the definition the OECD agreed in 2023 and the one the European Union wrote into Article 3 of the AI Act, which were aligned on purpose so a company can work from a single description.

What makes something AI

The word that carries the weight is "infers". An AI system does more than run steps a person spelled out; it works out how to reach an objective from what it is given. The OECD's explanatory memorandum names two broad families that do this: approaches that learn the pattern from data, which we call machine learning, and knowledge-based or symbolic approaches that reason from facts and rules a person has encoded. Both are AI.

So AI is not defined by whether it learns from data. A rule-based expert system that reasons over encoded knowledge is AI; a plain spreadsheet that only runs a formula is not, because it infers nothing. The memorandum is clear that this describes what AI systems have in common, not a precise legal test. Whether a particular system is regulated, and how, is a separate legal question you should not settle from this definition alone.

Where the term came from

The name is older than most people assume. It appeared in a 1955 proposal written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, for a summer research project held at Dartmouth College the following year, in 1956. For most of the decades after that, AI meant mainly knowledge-based and rule-driven systems. The ambition has stayed the same, but the dominant method has moved on: many of today's systems learn from very large amounts of data, and that shift is much of why AI feels useful at work now.

Narrow tools, general-purpose tools, and the AI people imagine

AI in use today ranges widely. Some systems are narrow and task-specific: a model that scores card payments for fraud does that and little else. Others are general-purpose: one large language model can summarise a contract, draft an email and translate a message from the same system. Capability is not the neat dividing line it once was.

What none of these are is "general AI", sometimes called artificial general intelligence: a system with broad, roughly human-level ability across almost any task. There is no agreed definition of it and no agreement on whether or when it might arrive, so treat confident claims in either direction with caution. When a vendor hints their product is close to it, read that as marketing. What you can buy today is a set of genuinely useful tools, some narrow and some broad.

Where companies already meet AI

You almost certainly use it already, often without noticing. The spam filter that sorts your inbox, the demand forecast that suggests how much stock to order, the fraud check on a card payment, the transcription of a recorded meeting, and the assistant that drafts or summarises text are all AI, and much of it has been running quietly for years.

The newer part is the generation of tools you talk to directly, the ones that draft text, answer questions and write code. Those belong to a family called generative AI, and they are worth understanding on their own because they behave differently from systems that only sort or score.

What to keep in mind

The limits depend on the kind of system. Any model that learns from data can carry the biases of that data, so results need checking where fairness matters. The tools people most often worry about are generative language models: they can state something wrong in the same calm, fluent tone they use for something right, a failure known as a hallucination, and most of them only know the world up to the date their training stopped, so recent facts can be missing. Those particular limits belong to generative text systems, not to every AI.

None of this makes AI unusable. It means the output is a starting point that a person still checks where it matters, rather than an answer to act on blindly.

Frequently asked questions

What is artificial intelligence in simple terms?
Artificial intelligence is software that, from some input, works out for itself how to produce a useful output such as a prediction, a recommendation, a piece of text or a decision. It can do this either by learning patterns from data or by reasoning from knowledge and rules a person has encoded.

What is the difference between AI and ordinary software?
Ordinary software only carries out fixed instructions a person wrote. An AI system goes further and infers how to produce its output, whether by learning from data or by reasoning over encoded knowledge. The OECD treats inference as the distinguishing feature, while stressing that this is a description rather than a precise legal test.

Is AI the same as machine learning?
No. Machine learning is the most common approach to AI today, where a system learns patterns from data. But AI also includes knowledge-based or symbolic systems that reason from facts and rules a person has encoded. Machine learning is one family within the broader field of AI.

Sources