Artificial intelligence (AI)

The term covers software that takes on tasks that used to need human judgment: understanding language, analyzing images, drafting text, spotting patterns in numbers. The difference from ordinary software lies in how it reaches a result. A conventional program executes rules that a person wrote down step by step. An AI system works out for itself how to turn an input into an output, usually because it learned that behavior from many examples beforehand.

AI is a field of computer science and an umbrella term. It includes machine learning, deep learning and generative AI, and also older knowledge-based methods in which experts encoded their knowledge as rules and facts. Every system in use today is built for a defined set of tasks, which researchers call narrow or weak AI. General AI that could learn and handle any task the way a person can does not exist yet, and whether and when it will arrive is disputed among researchers (as of November 2026).

Under the EU AI Act, AI is also a legal term. Whether a piece of software counts as an AI system helps determine which obligations apply when a company uses it. It also determines whether a vendor can legitimately call its product AI. Anyone buying AI should know which kind of AI sits behind a tool.

The definition in the EU AI Act.

Article 3(1) of Regulation (EU) 2024/1689 defines an AI system as "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments." The amending Regulation (EU) 2026/1744 added to Article 3 but left this definition unchanged.

The European Commission published guidelines on the definition on February 6, 2025. They break it into seven elements: machine-based, operating with varying autonomy, possibly adaptive, pursuing objectives, inferring, producing outputs such as predictions or content, and influencing an environment.

The element that carries the definition is the ability to infer. Recital 12 names two groups of techniques: machine learning, which learns from data, and logic- and knowledge-based approaches, which draw conclusions from encoded knowledge. Systems that only execute rules defined solely by people are excluded. The guidelines are not binding; the authoritative interpretation rests with the Court of Justice of the EU.

What does not count as AI.

The Commission's guidelines list several groups of software that calculate, or even infer in a narrow way, but fall outside the definition. One is basic data processing: a database query such as "all customers who bought product X last month" or a spreadsheet without AI features. Analyses that only describe and visualize data are also excluded, for example a sales dashboard showing totals, averages and trends.

Other groups are classical heuristics, meaning problem solving by fixed rules of thumb. The guidelines use the example of a chess program that evaluates positions with a fixed scoring function without having learned from data. Simple prediction systems that just project an average forward are excluded too, as are systems that speed up established mathematical optimization methods.

In practice this means a product labeled "AI" is not automatically an AI system, and many "smart" features are well-built rule sets. That is not a flaw. It becomes a problem when a vendor promises more than the technology delivers, which is what AI washing describes. Borderline cases can only be decided for the specific system; the guidelines give examples, not an exhaustive list.

A short history of AI.

In his 1950 paper "Computing Machinery and Intelligence," Alan Turing asked whether machines can think. The term itself comes from a funding proposal dated August 31, 1955. John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon asked for support for a summer project on "artificial intelligence," held at Dartmouth College in 1956. Their starting conjecture: every aspect of learning or any other feature of intelligence can in principle be described so precisely that a machine can simulate it.

In 1959, Arthur Samuel of IBM used the term machine learning for a checkers program that improved by playing. In the 1970s and 1980s, expert systems with knowledge encoded as if-then rules dominated the field. They were costly to maintain, and inflated expectations gave way to lean years known as AI winters.

The current wave came from large datasets and cheap computing power. In 2012, a deep neural network by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton won the ImageNet image recognition competition by a wide margin. In 2017, Google researchers introduced the Transformer architecture that today's language models are built on. With ChatGPT, generative AI reached a mass audience at the end of 2022.

The types of AI and how they fit together.

The simplest picture is a set of nested circles. The outer circle is artificial intelligence. Inside it are two broad approaches: knowledge-based systems, which draw conclusions from knowledge that people encoded, and machine learning, where software learns its rules from example data.

Deep learning sits inside machine learning. It uses artificial neural networks with many layers and has greatly improved image, speech and text recognition since about 2012. Generative AI is the part that creates new content such as text, images or code; today it is almost always built on deep learning. Large language models (LLMs) are the text form of generative AI. AI agents in turn build on language models and carry out multi-step tasks using tools.

A second distinction runs across this one: scope. Narrow AI handles defined tasks. General AI (often called AGI) would learn any task at human level and remains a research goal. For a business, the map answers a practical question. If a task has clear, stable rules, conventional software is often enough. If there are many examples but no rules you could write down, machine learning is the route. If the goal is a text, an image or a draft, generative AI is an option.

What decides success in day-to-day operations.

An AI system is only as good as the data it learned from and the data it receives in operation. Incomplete or outdated data produces poor results, and historical data carries old biases forward. A model trained on past hiring decisions, for example, picks up their patterns, including the unwanted ones. Data quality is therefore usually the first job, long before choosing a model.

The second is review. AI works with probabilities, so a certain share of its outputs will be wrong. For a sorting task, that share can be measured on test data. With generative AI, even a wrong answer sounds convincing, which is why outputs need checking before they are passed on. Responsibility stays with the person who uses a result.

The third is cost and structure. In ongoing operation, model calls, data maintenance, training and oversight often cost more than the license. Costs are best tracked per application, and access and permissions settled before tools spread across the company. An AI deployment is measured against the previous process: is it faster, cheaper or better, and by how much?

What the EU AI Act requires of companies.

The AI Act entered into force on August 1, 2024, and applies in stages. It assigns obligations by the risk of a use case, not by the technology. Since February 2, 2025, certain practices such as social scoring have been banned, and Article 4 requires measures for staff AI literacy. Since August 2, 2025, obligations have applied to providers of general-purpose AI models; they usually fall on the model developers.

Since August 2, 2026, the regulation has applied in general, including the transparency obligations in Article 50, for example for chatbots and deepfakes. The amending Regulation (EU) 2026/1744 has been in force since July 27, 2026. It moves the obligations for high-risk systems to December 2, 2027, and for AI in regulated products to August 2, 2028 (as of November 2026).

Most companies using purchased AI are deployers, not providers. For them, three questions matter most. Does a tool fall under the definition at all? Does the use case belong to a high-risk area, such as screening job applicants or assessing creditworthiness? And which transparency obligations apply? The GDPR continues to apply alongside as soon as personal data is processed.

Where the term stops working

Artificial intelligence is an umbrella term and says nothing about how good or useful a system is. Machine learning, deep learning, generative AI and language models are subsets of it, not synonyms. Rule-based automation such as RPA is not AI on its own, though it can include AI components. The legal term "AI system" in the EU AI Act is narrower than everyday use of the word AI.