Artificial intelligence (AI)

Artificial intelligence (AI) is software that does work which normally takes human thinking, such as understanding a question or recognising what is in a photo. Most AI in everyday use learns this from examples instead of following rules that a programmer wrote out. For a business it turns up in ordinary tools: a webshop chatbot that answers questions about orders, or a bank system that flags unusual payments.

Where does the term come from?

John McCarthy coined the term in 1955, in a proposal he wrote with Marvin Minsky, Nathaniel Rochester and Claude Shannon for a two-month summer study at Dartmouth College in 1956. It rested on the conjecture that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." McCarthy later defined AI as "the science and engineering of making intelligent machines, especially intelligent computer programs" (McCarthy, 2007).

How does it work?

There are two broad ways to build AI. In the first, people write the knowledge into the program, for example how to play chess. In the second, machine learning, the system finds the patterns in examples itself. Stanford professor Christopher Manning describes machine learning as the part of AI that studies how computer systems improve "based on experience or data", and writes that the emphasis has moved to systems that learn (Stanford HAI).

AI is the umbrella term, and IBM describes the terms under it as nested. Machine learning is part of AI. Deep learning, which uses neural networks with many layers, is part of machine learning. Generative AI, which creates text, images and other content, is built on deep learning. A large language model, such as the model behind ChatGPT, is an example of generative AI.

A learning system goes through two phases: training, in which it learns from examples, and inference, in which it applies what it learned to new input.

Narrow or general AI?

Manning calls systems built for one task, such as speech or face recognition, narrow AI. AGI (artificial general intelligence) is the goal of machines that are broadly intelligent and aware of context. Chatbots make that line harder to draw, because one model handles many kinds of tasks. In a 2023 position paper, researchers at Google DeepMind placed ChatGPT and Bard at the lowest general level of their scale, "Emerging AGI": equal to or somewhat better than an unskilled person (Morris et al.). By their account, no public system had reached the next level, "Competent AGI". IBM's explainer, updated in June 2026, calls AGI theoretical.

What counts as AI under EU law?

The EU's AI Act (Regulation (EU) 2024/1689, published in July 2024) has its own definition, and that definition decides which software the law covers. An AI system is a machine-based system that works with some degree of autonomy and infers from its input how to produce outputs such as predictions, content, recommendations or decisions. Recital 12 leaves out software that only executes rules written entirely by people. A spreadsheet formula that adds VAT falls outside the definition. The invoice model in the example below is the kind of system it describes.

Example

Suppose an accountancy firm receives a few hundred invoices a month as PDFs. Classic software needs a template for each supplier, with the total in one box and the VAT number in another, and a new layout breaks it. A machine learning model trained on many invoices has learned what a total and a VAT number look like, so it can also read a layout it has not seen before. It can still misread one, so someone checks the amounts before they go into the books.