How Artificial Intelligence Actually Works, and Why It Matters Now

The Organisation for Economic Co-operation and Development (OECD), whose member governments coordinate policy on the technology, defines an AI system as "a machine-based system 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 OECD Council revised this wording in 2023 to keep it technically accurate as generative AI tools spread, according to the organisation's own explanation of the definition.

From written rules to learning from data

Older software followed instructions that programmers wrote line by line: if X happens, do Y. Machine learning, the approach behind most current AI, instead gives an algorithm large amounts of examples and lets it work out the patterns itself, according to IBM's technical explainer on the subject. Once trained, the system can classify new information, predict a value, flag something unusual, or recommend an action without being given the exact rule for each case.

The OECD breaks an AI system's operation into three elements. Inputs are the data, code or rules fed into the system during development or use. Inference is the step in which the system works out an output from those inputs, which can happen while a model is being built or after it has been deployed. Outputs are the predictions, content, recommendations or decisions the system produces, which can then affect a physical or virtual environment — for example, a car's steering, a list of recommended videos, or a generated paragraph of text.

What a neural network does inside the system

The most widely used machine-learning method today is the neural network, a structure loosely modelled on the brain, made of interconnected layers of nodes that process data together, according to IBM. Unlike a traditional program, a neural network is not told the rule connecting input to output in advance; instead it is shown many pairs of inputs and outputs and adjusts itself, through repeated exposure, to discover what connects them.

Deep learning refers to neural networks with many layers — typically an input layer, at least three (and often hundreds of) hidden layers, and an output layer, IBM's explainer states. In an image-recognition system, early layers tend to respond to simple features such as edges or colours, while deeper layers combine these into shapes and, eventually, recognisable objects. This layered structure lets systems automatically extract relevant features from unlabelled data, which is part of what makes current generative AI tools possible.

Building a large generative AI model generally involves three stages, per IBM: an initial training phase on massive, often unlabelled datasets that can require thousands of specialised processors and cost millions of dollars; a tuning phase in which the model is adapted to specific tasks, sometimes using human feedback; and ongoing evaluation after deployment to check and refine its outputs.

Why it matters: adoption and investment figures

AI's reach has grown quickly by several measures tracked in Stanford University's Human-Centered AI Institute 2025 AI Index Report. The share of organizations reporting any use of AI rose to 78% in 2024, up from 55% the year before. Generative AI specifically drew $33.9 billion in global private investment in 2024, an 18.7% rise from 2023. In the United States, private AI investment reached $109.1 billion in 2024, far exceeding China's $9.3 billion and the United Kingdom's $4.5 billion, the report said.

The report also recorded a jump in AI benchmark performance: on three tests introduced in 2023 to measure advanced systems — MMMU, GPQA and SWE-bench — scores rose by 18.8, 48.9 and 67.3 percentage points respectively within a year. In regulated fields, the number of AI-enabled medical devices cleared by the U.S. Food and Drug Administration grew from six in 2015 to 223 in 2023, according to the same report.

Where the competition stands

The Stanford report also tracked which countries produced the most notable AI models in 2024: U.S.-based institutions produced 40, compared with 15 from China and three from Europe. At the same time, the performance gap between top U.S. and Chinese models on major benchmarks narrowed from double digits in 2023 to near parity in 2024, the report said.