What is artificial intelligence and how does it work

You open a maps app and it suggests the fastest route. Your email sorts out the spam on its own. A video platform picks the next thing to watch. In many of these cases an artificial intelligence system is at work, even if nobody tells you.

The problem is that the expression “artificial intelligence” has become a label for almost everything, from assistants that hold conversations to simple app features. This article organizes the subject: what the term means, where it came from, how a machine “learns” and where the limits are.

A definition without science fiction

The reference most widely used by governments today is that of the Organisation for Economic Co-operation and Development (OECD). In the version updated in November 2023, an AI system is 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 text adds that these systems vary in their levels of autonomy and adaptiveness after deployment (OECD.AI, 2023).

This definition is worth reading slowly, because it says a lot through what it leaves out:

• It does not mention consciousness. The criterion is functional: receive data and produce a useful output.
• The objectives come from people. Someone decided what the system should predict or recommend.
• The output has an effect. A recommendation changes what you watch; a credit decision changes someone’s life.
• Autonomy is a scale. Some systems only suggest and others act on their own.

The U.S. National Institute of Standards and Technology (NIST) uses a definition adapted from the same OECD basis in its AI Risk Management Framework, published in January 2023 (NIST AI 100-1). In practice, this means technical bodies in different countries treat AI as a category of system, not as a mind.

Where the name came from

The term appears in a research proposal dated August 31, 1955, signed by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, for a summer meeting at Dartmouth College (Dartmouth proposal).

The document started from a bet: that every aspect of learning or any other feature of intelligence could, in principle, be described so precisely that a machine could be made to simulate it. Seven decades later, that bet is still open. What has changed is the path toward it.

Written rules or learned patterns

There are two basic ways to make a computer take an “intelligent” decision.

The first is to write rules. Imagine a spam filter programmed like this: if the message contains “make money fast,” send it to the trash. It works until the sender changes the phrase. Each exception requires a new rule, and the list grows without end.

The second is to let the system discover the rules on its own from examples. This is the field of machine learning, which Google’s documentation describes as the process of training a piece of software, called a model, to make useful predictions or generate content from data (Google for Developers).

In the filter example, instead of writing rules, you show the model thousands of messages already labeled “spam” or “not spam.” It finds on its own the combinations of words, senders and formats that separate one group from the other. Almost everything called AI in everyday life today follows this second path.

How a model learns, step by step

Training seems mysterious, but the logic fits into four steps.

1. Data with answers. In what is called supervised learning, the model receives examples with the correct answer already indicated: labeled photos, classified emails, houses with known sale prices.

2. An initial guess. The model is a mathematical relationship with many adjustable numbers, called parameters or weights. At the start, these values know nothing, and the guesses are poor.

3. Measuring the error. For each example, the guess is compared with the correct answer. The difference becomes a number called loss: the higher it is, the worse the model is doing.

4. Adjust a little and repeat. A technique called gradient descent calculates in which direction to change each weight to reduce the loss and takes a small step in that direction. The process repeats until further steps no longer improve the result meaningfully (Google for Developers).

An image helps: it is like walking down a mountain in fog. You cannot see the valley, but you feel the slope of the ground and take one step down at a time.

Once trained, the model is used on new data. This usage phase is called inference, the same word that appears in the OECD definition. Training tends to be expensive and happens rarely; inference happens with every question you ask.

Other ways of learning

Not every problem comes with ready-made answers. That is why other families of learning exist (Google for Developers):

• Unsupervised learning: the model looks for patterns in unlabeled data, such as grouping customers with similar behavior.
• Reinforcement learning: the model acts in an environment and receives rewards or penalties, adjusting its strategy based on that feedback.
• Generative AI: models that create new content, such as text, images or audio, from a user’s input. The topic has its own article: Generative AI: how models create text, images and code.

Neural networks and deep learning

Many real problems do not have simple patterns. The relationship between the pixels of a photo and the word “cat” is not a straight line. Neural networks are a family of models designed precisely to find nonlinear patterns in data (Google for Developers).

A neural network organizes its weights in layers. Each layer transforms the information it receives and passes the result on. During training, an algorithm called backpropagation distributes responsibility for the error across the layers so that all of them are adjusted. When there are many layers, it is called deep learning.

This approach has gained recognition beyond computing. According to Stanford University’s AI Index 2025, two 2024 Nobel Prizes recognized AI-related work: the Physics prize, for foundations of deep learning, and the Chemistry prize, for protein structure prediction (Stanford HAI, AI Index 2025).

In 2017, researchers introduced a network architecture called the Transformer, based on an attention mechanism that lets the model relate distant parts of a sequence (Vaswani et al., 2017). It became the foundation of large language models, explained in What are LLMs.

The AI you already use

Much of AI does not look like an assistant. It shows up as product recommendations, keyboard autocorrect, translation, auto-generated video captions, card fraud detection or phone face unlock.

A few figures help gauge its reach, always with the date of the information:

• The AI Index 2025 records that the U.S. regulator FDA approved 223 AI-enabled medical devices in 2023, compared with six in 2015 (Stanford HAI, AI Index 2025).
• The AI Index 2026 reports that AI adoption by organizations reached 88% and that industry produced more than 90% of the notable frontier models in 2025 (Stanford HAI, AI Index 2026).

The second figure has a practical consequence: the most advanced systems today are largely company products, not university or government ones. This influences what is published about them and what remains a trade secret.

What AI is not

Understanding how a model learns helps you understand how it fails.

It learns from the data it was given. If the training examples are incomplete or biased, the model reproduces the problem. It has no way of knowing what was left out.

It can fail outside its training conditions. The NIST framework, which is voluntary, recommends documenting a system’s generalization limits, that is, how well it works outside the conditions under which it was developed (NIST AI 100-1). The same document lists the characteristics of trustworthy AI: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair, with harmful bias managed.

Capability is uneven. A model can be excellent at a hard task and get a seemingly simple one wrong. The AI Index 2025 notes that models performing well on math olympiad problems still struggled with complex planning tests. The AI Index 2026 cites that the best model evaluated read analog clocks correctly 50.1% of the time (Stanford HAI, AI Index 2026).

Fluency is no guarantee of truth. Models that generate text can produce convincing and wrong answers. The reasons are in AI hallucinations, and a method for checking outputs is in How to evaluate AI responses.

In the editorial team’s assessment, the most useful stance is to treat AI as a powerful statistical tool: great for speeding up work, but one that needs someone who knows how to check the result when it matters.

Key takeaways

• AI is a category of system. Under the OECD and NIST definitions, it is a system that, from inputs, generates predictions, content, recommendations or decisions with some degree of autonomy.
• Almost all current AI learns from data. A model starts out wrong, measures its own error and gradually adjusts its weights until it errs less.
• Deep neural networks and the Transformer explain much of the recent leap, including the assistants that hold conversations.
• The limits come from the method itself. Poor data, situations outside training and uneven capability mean human verification remains necessary.