What is artificial intelligence and how does it work?
What is artificial intelligence?
Artificial intelligence (AI) is intelligence in machines, the counterpart of the natural intelligence of humans and other living organisms. The field took its name and became an official area of research in 1956. Early work produced new tools for solving maths problems.
Researchers found that building an AI is extremely hard, and progress slowed in the 1970s. More recently, growth in computing power and the availability of enormous datasets have laid the ground for fresh advances.
How did the idea of AI begin?
In 1955, four researchers proposed a study of artificial intelligence lasting 2 months and involving 10 people, to be held at Dartmouth College in the summer of 1956. The proposal rested on the assumption that any aspect of learning, or any other feature of intelligence, could in principle be described so exactly that a machine could imitate it. The aim was to find out how to get machines to use language, build abstractions and concepts, solve problems reserved for people, and improve themselves.
According to the proposal, the main obstacle was not a lack of machine capacity but our inability to write programs that make full use of what we already have. It also listed neural networks, asking how a set of hypothetical neurons could be arranged so as to form concepts.
What is machine learning?
Scientists have made huge strides in machine learning, one application of AI. Machine learning refers to systems that teach themselves from the data they examine and the results they get, and so become better at handling that data. For instance, it powers apps that estimate how long a trip will take and then suggest a route. It suits specific, well-defined jobs that can be split into logical steps.
Before the boom in generative AI, talk of AI usually meant machine-learning models that could learn to make predictions from data. Such models are trained on millions of examples to predict, for instance, whether an X-ray shows signs of a tumour or whether a borrower may fail to repay a loan.

What is deep learning?
The next level of AI is known as deep learning. It deals with complex jobs that have many variables that are hard to pin down. It uses neural networks, which try to make computers work like our brains. It relies on many neural networks piled on top of each other to build one large, complex network.
Deep learning needs a great deal of computing power, both to train a model on huge amounts of data and to use trained models to make decisions. Digital assistants are one example: Amazon Alexa, Apple Siri and Google Assistant.
In 2011, IBM’s Watson won the quiz show Jeopardy! in the US. In 2016, Google DeepMind’s AlphaGo AI beat a human grandmaster at Go, the complex Chinese game. People say that AI is not very good at games that call for teamwork.
How does generative AI work?
Put simply, generative AI is a machine-learning model that is trained to make new data instead of predicting something about a given dataset. A generative AI system learns to make more items that resemble the data it was trained on. These powerful machine-learning models build on research and advances in computing that reach back more than 50 years.
An early form of generative AI was a far simpler model, the Markov chain. The method is named after the Russian mathematician Andrey Markov, who in 1906 brought in this statistical technique to model how random processes behave. Within machine learning, Markov models have for a long time been used to predict the next word, as in an email program’s autocomplete. When predicting text, a Markov model produces the next word of a sentence by looking at the word before it, or a few words before it. Because such simple models can look back only so far, they are poor at writing believable text.
Under the hood, ChatGPT and comparable systems operate in largely the same fashion as a Markov model. The big difference is size: ChatGPT is much larger and more complex, with billions of parameters. It was trained on a vast amount of data, in this case a large share of the text freely available on the internet.
In this huge heap of text, words and sentences come in sequences with certain dependencies. This repetition helps the model learn how to divide text into statistical pieces that are somewhat predictable. The model picks up the patterns in these chunks of text and uses that knowledge to suggest what might come next.
Which architectures are used?
In 2014, researchers at the University of Montreal proposed a machine-learning architecture called the generative adversarial network (GAN). GANs use two models that work together: one learns to produce a target output, such as an image, and the other learns to tell real data apart from the producer’s output.
A year after that, Stanford University and University of California, Berkeley researchers presented diffusion models. By refining their output step by step, these models learn to produce new samples that look like the ones in their training dataset.
Google researchers unveiled the transformer architecture in 2017, which has been used to build large language models such as the ones behind ChatGPT. A transformer encodes every word of a text as a token and then produces an attention map showing how each token relates to all the others. Using this map, the transformer keeps track of context while it writes new text. What these approaches share is that they turn inputs into sets of tokens, which are numerical representations.
On tasks that involve predicting from structured data, such as tables, generative AI models often fall behind traditional machine-learning methods.
Risks and environmental cost
Using generative AI chatbots in call centres highlights one possible red flag for these models: the replacement of workers. Generative AI can inherit and spread biases found in its training data, and can amplify hate speech and false statements. The models can produce content that appears to come from a particular human creator, which can raise copyright problems.
According to a report by United Nations University published in 2026, the world’s data centres that power AI are expected to use 945 terawatt-hours of electricity by 2030. The report finds that the environmental cost of AI is being measured wrongly in a systematic way.
Once a model is put into use, inference, meaning the constant running of models to answer users’ everyday prompts, becomes the dominant cost at 80 to 90 percent of all AI energy use. A typical chat query takes about 200 times more energy than basic text classification. Producing a single AI image can need roughly 1,450 times that baseline.
Test yourself
Three questions from the article. Pick an answer to find out.
Question 1 of 3
In which year did artificial intelligence get its name and become a formal field of research?
The field took its name and became an official area of research in 1956. Source: Energy.gov
Show answer
Right answer: 1956
The field took its name and became an official area of research in 1956. Source: Energy.gov
Question 2 of 3
What do machine-learning systems do?
Machine learning refers to systems that teach themselves from the data they examine and the results they get, and so become better at handling that data. Source: Energy.gov
Show answer
Right answer: They learn automatically from the data they analyse and from their results, and so improve at working with that data
Machine learning refers to systems that teach themselves from the data they examine and the results they get, and so become better at handling that data. Source: Energy.gov
Question 3 of 3
How does the transformer architecture encode each word of a text?
A transformer turns every word of a text into a token and then produces an attention map showing how each token relates to all the others. Source: MIT
Show answer
Right answer: As a token
A transformer turns every word of a text into a token and then produces an attention map showing how each token relates to all the others. Source: MIT


