Article · 27 min read

What is AI?

The word is everywhere, from your phone to parliament. But what is artificial intelligence really, how does it work and why is everyone talking about it now? Here are the basics, explained from the start, with seven figures you can try for yourself.

Short answer

AI, artificial intelligence, is computer systems that can do tasks that would otherwise require human intelligence, such as understanding language, recognising images or making decisions. Most of what is called AI today is built on machine learning: instead of following rules written by people, the programs learn patterns from large amounts of data. Deep learning with neural networks brought the breakthrough in 2012, and since ChatGPT arrived in 2022 generative AI has been used by hundreds of millions of people. In Sweden, 47 percent of everyone aged 16 and over had used generative AI in the past three months in 2026. Today’s AI is uneven: it can reach gold at the Mathematical Olympiad but struggles to read an analog clock.

What does AI mean?

AI stands for artificial intelligence: computer systems that can do tasks that would otherwise require human intelligence, such as understanding language, recognising images or making decisions. The word is older than most people think. It was coined by the computer scientist John McCarthy and first appeared in August 1955, in a proposal from him and three colleagues for a summer study of artificial intelligence at Dartmouth College in the United States the following year. Their starting point was that every aspect of learning and intelligence can in principle be described so precisely that a machine can simulate it.

The question was older still. In 1950 the British mathematician Alan Turing opened a famous paper with the words “Can machines think?”. Instead of trying to define thinking, he proposed a test: if a machine can hold a conversation so well that it cannot be told apart from a person, there is reason to call it intelligent.

Today there are many definitions, and the Swedish government noted as early as 2018 that the term has no single definition. The one that carries most weight in Europe is in the EU AI Act, where an AI system is “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”. At its heart is the ability to infer, which according to the Act goes beyond basic data processing by enabling learning, reasoning or modelling. Systems based only on rules defined by people to execute operations automatically do not count.

The line keeps moving, too. As soon as something works, many people stop calling it AI. The researchers behind Stanford’s AI100 report call this the AI effect: the technology becomes everyday, people get used to it, and then it no longer counts as AI. The computer scientist Larry Tesler put the same thing around 1970: intelligence is whatever machines haven’t done yet.

Figure 1Is it AI?

The figure is interactive and needs JavaScript.

Eight everyday things. The answers follow the EU definition of an AI system and the European Commission’s guidelines on how to read it.

How does AI work?

There are two main ways to make a computer smart. One is to write rules. In the 1970s and 1980s people built so-called expert systems, in which knowledge from experts was written in as long lists of if-then rules. One of the best known, R1 at the computer company Digital Equipment, was already in regular use in the company’s manufacturing in 1980, putting together computer orders.

Rules work well when the world is orderly. But many tasks are almost impossible to describe with rules. Try writing down exactly how to recognise a cat in a picture, or what makes an email spam. The other way is therefore to let the computer learn from examples. This is called machine learning. The pioneer Arthur Samuel at IBM built a draughts program in the 1950s that learned from games it had already played, and he is usually credited with coining the term. He wrote in 1959 that computers that learn from experience should eventually remove the need for much of the detailed programming effort.

Figure 2Rules or examples

The figure is interactive and needs JavaScript.

Made-up emails, but a real little program that learns: the machine adjusts three numbers until it makes as few mistakes as possible. Filled dots are examples it trains on, rings are new emails it has never seen. Click in the chart to add examples of your own.

The program in the figure does on a small scale what today’s large AI models do. It gets examples with the right answers, guesses, sees how wrong it was and adjusts its numbers a little at a time. The difference is size: the program above has three numbers to adjust, the largest language models have hundreds of billions. Notice too that the machine can only become as good as its examples. If the examples are few, skewed or wrong, so is what it learns.

The work happens in two steps. First the model is trained, and for large models that is the part that needs the most computing power. In 2024 Meta trained its largest language model, Llama 3.1, on more than 16,000 graphics processors. Then the finished model is used. Every time you ask a chatbot a question, it works out an answer with the numbers it has already learned, and that step is called inference.

Three ways to teach a machine

Machine learning can work in several ways. In its guidelines on the AI Act, the European Commission lists supervised, unsupervised and self-supervised learning, and reinforcement learning.

In supervised learning the machine gets examples where the answer is already given. The Commission’s own example is the spam filter, trained on emails that people have labelled as spam or not spam. Systems that recognise objects in pictures or look for fraud among payments are trained the same way.

In unsupervised learning there are no answers. The machine gets only data and looks for patterns by itself, such as groups of customers who shop in similar ways, or events that stand out. The Commission mentions drug companies that use such systems to find new medicines.

In reinforcement learning the machine gets no answers either, only a reward when things go well. It tries things out and learns over time which moves pay off. That is how AlphaGo learned to play Go: according to Google DeepMind, the program played thousands of games against different versions of itself and learned from its mistakes. The method is also used to polish chatbots. People judge the answers, and the model is rewarded for the answers people prefer. That technique is called RLHF.

The fourth way explains how language models could grow so large. In self-supervised learning the machine makes its own answer key from the data. A language model hides the next word in a sentence and tries to guess it. The answer is already in the text, so nobody has to label anything, and that is why it can be trained on enormous amounts of text. The Commission mentions exactly language models that learn to predict the next token in a sentence.

Figure 3Three ways to learn

The figure is interactive and needs JavaScript.

Three small programs that really learn, on made-up data. With answers, the machine guesses what a new animal is by comparing it with the five most similar examples. Without answers, it looks for groups by itself, with a method called k-means. With rewards, the robot works its way to the charger, with a method called Q-learning.

Neural networks and deep learning

Almost all the AI that has made headlines in the past fifteen years is built on neural networks. When it awarded the Nobel Prize in Physics in 2024, the Royal Swedish Academy of Sciences wrote: “When we talk about artificial intelligence, we often mean machine learning using artificial neural networks.” The technology was originally inspired by the structure of the brain.

The idea is old. As early as 1943 the neuroscientist Warren McCulloch and the logician Walter Pitts described a mathematical model of how the brain’s nerve cells cooperate: each cell weighs together signals from other cells and passes on a signal of its own. In 1957 Frank Rosenblatt proposed the perceptron, a network for interpreting images that was also built as a machine.

An artificial neuron is simple. It takes in numbers, multiplies each one by a weight, adds them up and passes a new number on. The weights are the network’s parameters, and they are what training adjusts. Every time the network gets something wrong, a method called backpropagation works out which way each weight should change to make the error smaller. The method had its breakthrough in 1986, when David Rumelhart, Geoffrey Hinton and Ronald Williams showed how it could train networks with hidden layers between input and output.

It is the layers that make networks powerful. In a network that recognises images, the first layer finds edges, the next puts the edges together into corners and contours, and layer by layer more complicated concepts are built from simpler ones. No person decided what the layers should look for; the network learned it from the examples. That is why it is called deep learning: the depth is the number of layers.

Figure 4From a line to a network

The figure is interactive and needs JavaScript.

A real neural network trained in your browser, on made-up points. The squares in the network show where each neuron says yes (gold) or no (blue), and the lines are the weights: the thicker, the more they count. The network is trained with backpropagation.

For a long time neural networks were a sideline. At the end of the 1960s Marvin Minsky and Seymour Papert showed that the simplest networks cannot even handle the task “one or the other, but not both”, and the money for research dried up. You can see the same limit in the figure: a single neuron cannot handle the ring. The breakthrough came in 2012. A deep neural network from Geoffrey Hinton’s group in Toronto won a large image recognition competition with an error rate of 15.3 percent, against 26.2 percent for the runner-up. Since then almost all new AI has been built on neural networks, and the largest language models can have more than a trillion parameters.

The likeness to the brain should not be overstated, though. The textbook Deep Learning points out that the networks are only loosely inspired by neuroscience and that the goal is not to model the brain perfectly. An adult male brain has on average about 86 billion nerve cells, according to a study from 2009. The Academy of Sciences put it like this: “Although computers cannot think, machines can now mimic functions such as memory and learning.”

AI, machine learning and generative AI

How do the terms fit together? AI is the umbrella, everything that lets machines do tasks that require intelligence. Machine learning is part of AI, where programs learn from data, but not all of it. An expert system with written-in rules is AI too. Deep learning in turn is part of machine learning. The textbook Deep Learning by Ian Goodfellow, Yoshua Bengio and Aaron Courville from 2016 draws it as rings inside each other, and points out that machine learning is not the only route to AI.

The newest and most talked-about branch is generative AI, which creates new content: text, images, audio or code. Its best-known form is the language models behind chatbots such as ChatGPT, Claude and Gemini. They have been trained on vast amounts of text and write their answers by predicting the next word, over and over. So ChatGPT is AI, more precisely generative AI built on deep learning. How language models work is explained in the article How does a language model work?

Figure 5The AI family

The figure is interactive and needs JavaScript.

A simplified picture. The rings are based on the diagram in the textbook Deep Learning (2016), with generative AI and language models added. The definitions follow the International AI Safety Report 2026. Tap a ring.

Examples of AI in everyday life

Most AI goes unnoticed. It is in your phone, in your apps and in services you use every day, often without anyone calling it AI. Here are ten examples, several of them from Sweden.

  • Your phone knows your face

    When you unlock an iPhone with Face ID, the phone lights up your face with more than 30,000 invisible dots. The image is pushed through neural networks that build a mathematical model of your face, and all of it is worked out on the phone.

    Deep learning

  • Spam is sorted out

    According to Google, Gmail blocks more than 99.9 percent of spam, phishing and malware. Since 2024 it also uses a language model trained on exactly that kind of email.

    Supervised learning

  • The map app works out the journey time

    Google Maps uses machine learning to combine live traffic with how traffic usually flows. When Google DeepMind brought in new neural networks in 2020, arrival times became up to 50 percent more accurate in cities such as Berlin and Tokyo.

    Neural networks

  • Music is picked for you

    On Spotify, algorithms choose and arrange what you see in Home, Search and the personal playlists, based on your taste and how listeners in general behave. Spotify describes its DJ feature as an AI-driven tool.

    Recommender systems

  • Translation got better

    When Google began translating with neural networks in 2016, whole sentences were translated at once instead of words and phrases one by one. According to Google, errors fell by 55 to 85 percent on several major language pairs.

    Neural networks

  • The search engine answers

    Since October 2025 Google’s AI Mode has been available in Swedish. You can ask complicated questions there and get a written answer from a version of Google’s Gemini models. Google itself writes that it will not always get things right.

    Generative AI

  • Mammography gets help

    About 106,000 women in Skåne took part in the MASAI study, which compared AI-supported reading of mammograms with standard reading. The AI support found 29 percent more breast cancers and cut the radiologists’ reading workload by 44 percent. Many Swedish regions have already started using the method.

    Image analysis

  • The weather app sees the rain

    Since November 2025 SMHI, the Swedish weather service, has used the AI model CHIMP in its precipitation forecast. It was trained on four years of satellite data and shows up in the weather app’s precipitation map. The European weather centre ECMWF has run an AI model since 2025 that uses about a thousand times less energy per forecast.

    Machine learning

  • Swedish speech becomes text

    The National Library of Sweden has trained KB-Whisper on 50,000 hours of Swedish speech with transcripts, including television with subtitles and recordings from parliament. According to the library, the best version makes on average 47 percent fewer word errors than OpenAI’s Whisper-large-v3.

    Speech recognition

  • The agency answers in the chat

    The Swedish Tax Agency’s chatbot Skatti answers general questions about population registration and tax returns, and the Swedish Pensions Agency has a similar one called Penni. The Tax Agency also uses AI to, among other things, sort incoming email and analyse large amounts of data.

    Chatbot

AI is also used to prepare decisions about people, and there the questions get harder. We come back to that under Is AI dangerous?

A short history of AI

The history of AI is one of great hopes and setbacks. After Dartmouth many believed thinking machines were close. When they did not arrive, the mood turned. In Britain the mathematician James Lighthill concluded in a report in 1973 that no discoveries in the field had produced the major impact that had once been promised, and funding for AI research was cut. The period is called the first AI winter, although historians debate how much the report itself caused.

In the early 1980s a new spring came with expert systems, and Japan invested heavily in an AI project of its own. But the systems only managed relatively simple problems and could not be scaled up, and in the late 1980s the next winter came. The AI researcher Melanie Mitchell recalls that when she finished her PhD in 1990, she was advised not to use the term artificial intelligence in her job applications.

Then the breakthroughs came faster. In May 1997 IBM’s chess computer Deep Blue beat the world champion Garry Kasparov by 3.5 to 2.5. After deep learning’s breakthrough in 2012, Google DeepMind’s AlphaGo beat one of the world’s best Go players, Lee Sedol, by four games to one in 2016, watched by more than 200 million people according to DeepMind. In 2017 researchers at Google introduced the transformer, the architecture behind today’s language models. On 30 November 2022 OpenAI released ChatGPT, which according to an estimate by the bank UBS had 100 million monthly users after two months. In September 2026 OpenAI said that more than a billion people use ChatGPT every week.

In 2024 AI came to Stockholm. The Royal Swedish Academy of Sciences awarded the Nobel Prize in Physics to John Hopfield and Geoffrey Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks”. Half of the chemistry prize went to Demis Hassabis and John Jumper of Google DeepMind for AlphaFold, which has predicted the shapes of more than 200 million proteins. In the summer of 2025 an AI model from Google DeepMind reached gold-medal level at the International Mathematical Olympiad, graded by the olympiad’s own coordinators. More milestones are on our timeline.

Why is so much happening now?

AI research is seventy years old. Why has it only taken off now? Researchers usually point to three ingredients: computing power, data and better algorithms. A fourth can be added: money.

The biggest change is in computing power. Researchers at Epoch AI have worked out how many calculations it took to train hundreds of well-known AI models. Before 2010 the amount doubled roughly every twenty months, in step with computers getting faster. Since deep learning took off it has doubled roughly every six months, and for the most advanced language models it has grown about five times a year since 2020. The calculations are mostly done on graphics processors, chips that can carry out a large number of calculations at the same time rather than one after another.

Figure 6The computing power behind AI

The figure is interactive and needs JavaScript.

Each dot is a well-known AI model, and its height shows how many operations it took to train it. The scale is logarithmic: each line up means 100,000 times more. The numbers are estimates from Epoch AI (CC BY 4.0), read on 15 November 2026. For many of the newest models the companies have not said enough for the compute to be worked out.

The second ingredient is data. When the image database ImageNet was presented in 2009 it held 3.2 million images, and by 2012 it had grown to more than 15 million. The network from Toronto was trained on 1.2 million of them. Language models are trained on far more. Meta trained its Llama 3 on more than 15 trillion tokens from publicly available sources. Researchers at Epoch AI have calculated that if the trend continues, models will at some point between 2026 and 2032 be trained on about as much text as people have published in public.

The third ingredient is better methods. The transformer of 2017 could be computed in parallel and took far less time to train than earlier methods. Researchers have estimated that the computing power needed for a language model to reach a given level has halved roughly every eight months thanks to better algorithms. Even so, the growth in computing power has mattered more. The AI researcher Rich Sutton called it the bitter lesson in 2019: the biggest lesson from 70 years of AI research, he wrote, is that general methods that leverage computation are ultimately the most effective, and by a large margin.

The fourth ingredient is money. According to Stanford University’s AI Index, private investors put 344.7 billion dollars into AI worldwide in 2025, more than twice as much as the year before. Nearly half went to generative AI companies.

What AI can and cannot do today

According to Stanford University’s annual AI Index from April 2026, progress is not levelling off but accelerating. The best models meet or exceed human baselines on PhD-level science questions, on tasks that require reasoning about both text and images, and in competition mathematics.

But the ability is uneven, and researchers speak of a jagged frontier. The same report notes that AI can win gold at the Mathematical Olympiad but cannot reliably tell the time: the best model read an analog clock correctly 50.1 percent of the time. AI agents still fail roughly one in three tasks on structured tests, and robots manage only 12 percent of ordinary household tasks. The International AI Safety Report from February 2026 describes the same thing: systems can be derailed by simple errors in projects with many steps, and they still make things up.

That makes today’s AI useful and unreliable at one and the same time. Why it makes things up is explained in Why does AI make things up?, how AI that carries out tasks on its own works in What is an AI agent? and what it all costs in electricity in How much energy does AI use?

Narrow AI, general AI and superintelligence

A common way to divide up AI is by how broad it is. For a long time AI was built for one task at a time. A chess program cannot translate and a translation program cannot play chess. AI like this is called narrow AI. Today’s chatbots are something else: they can write, calculate, program and translate in the same conversation. The International AI Safety Report calls them general-purpose AI.

Yet they are not what researchers mean by AGI, artificial general intelligence: a system that can do roughly any intellectual work a person can. Stanford’s AI100 report wrote in November 2026 that a large and growing share of humanity uses powerful general-purpose AI tools, but that they are not yet AGI, and that there is still no agreement on exactly what would count as AGI. More on that in the article What is AGI, really?

One step further away is superintelligence, an intellect that vastly outperforms the best human minds in practically every field. That is the question all of SI-bolaget is about, and which we describe on the front page.

Is AI dangerous?

It depends on what you mean, and researchers do not agree on everything. The International AI Safety Report 2026, written by over 100 experts and backed by more than 30 countries and international organisations, sorts the risks into three groups: someone using AI to cause harm, AI going wrong, and effects on society as a whole. At the same time the report notes that AI is already being put to good use in healthcare, research and education.

Errors and fraud

Today’s AI sometimes makes up facts, writes flawed code and gives misleading advice, including about health. Fraudsters use AI to clone voices and make fake videos, so-called deepfakes, to trick people out of money. In March 2026 Sweden’s financial supervisor, Finansinspektionen, warned that it can be almost impossible to tell that such content is fake, and according to the agency’s survey six in ten Swedes worry about how AI can be used for fraud.

Unfair decisions

AI that learns from data can also learn prejudice. The International AI Safety Report from 2025 found that general-purpose AI systems frequently display biases with respect to things like gender, culture, age and disability. In Sweden, the newspaper Svenska Dagbladet and Lighthouse Reports in 2024 investigated a machine learning model that the Swedish Social Insurance Agency, Försäkringskassan, used to pick out claims for temporary parental benefit for checks. According to the investigation, the model was more likely to wrongly flag women, migrants, people on low incomes and people without a university education. The Swedish data protection authority, IMY, opened a supervision, and in 2025 Försäkringskassan stopped using the model. In November 2026 IMY opened a review of an AI system that the Swedish Public Employment Service uses to judge what support jobseekers can get.

Jobs

About 60 percent of jobs in advanced economies and 40 percent in emerging economies are exposed to AI, according to the international report. So far there is no effect on overall employment, but some signs of falling demand for early-career workers in occupations that are highly exposed, such as writing. Economists disagree about the future. Some expect new jobs to replace those that disappear, others think automation could push down both employment and wages.

Bigger risks ahead

The report also covers risks that may grow as AI becomes more powerful. AI can find security holes and write malicious code, and criminal groups and state-linked attackers already use it. In 2025 several AI companies released new models with extra safeguards because they could not rule out that the models might help novices develop biological or chemical weapons. And there is the question of whether people could lose control of AI systems. Today’s systems are not capable of that, but expert opinion on how likely it is varies greatly. We go through the question in Can we control something smarter than us?

What does the law say?

The EU AI Act regulates AI according to how big the risk is. Most AI, such as spam filters and AI in video games, faces no new rules at all. Since February 2025 some uses are banned, for example social scoring, scraping facial images from the internet to build facial recognition databases, and reading people’s emotions in workplaces and schools. Since August 2026 people must be told when they are talking to a chatbot, and anyone who spreads deepfakes must disclose that they were made with AI. From December 2026 AI that creates nude images of real people is banned too. The strictest rules, for AI in areas such as education, employment and border control, apply from December 2027, after the EU postponed them in the summer of 2026.

AI in Sweden

Swedes have taken to AI quickly. According to Statistics Sweden (SCB), 47 percent of everyone aged 16 and over had used generative AI tools in the past three months in 2026, against 25 percent two years earlier. Among 16 to 24-year-olds the share was 71 percent, and the biggest rise was among older people: among 65 to 74-year-olds it went from 4 to 20 percent.

Figure 7Sweden and AI

The figure is interactive and needs JavaScript.

People: the share of the population aged 16 and over who used generative AI tools in the past three months, according to Statistics Sweden (SCB). Companies: the share of companies with at least ten employees that use any AI technology, according to Eurostat. Choose a year to see the change.

The most used service is ChatGPT, which according to the Internet Foundation’s survey Svenskarna och internet 2026 is used by 52 percent of everyone aged 8 and over who uses the internet. Three in four AI users say they usually double-check the answers, but among retired people who use AI nearly four in ten do not. At the same time many are sceptical. In the SOM Institute’s survey in the autumn of 2025, 38 percent were mainly pessimistic about AI’s effect on society, against 25 percent who were optimistic.

Companies are well ahead too. According to Eurostat, 35 percent of Swedish companies with at least ten employees used AI in 2025, nearly twice the EU average. Only Denmark and Finland were higher. Among large companies the share was 72 percent according to SCB.

Policy has followed. The AI Commission presented a roadmap in November 2024 with 75 proposals, estimated to cost 16.7 billion kronor over ten years, and warned that Sweden was falling behind. In February 2026 the government presented Sweden’s first comprehensive AI strategy, with the goal that Sweden should be among the ten leading countries in the world in AI. The national centre for applied AI, AI Sweden, built GPT-SW3 as early as 2023, the first large generative language model for the Nordic languages.

How to get started

  • Try a chatbot on something you know a lot about yourself. You will quickly see both what it can do and where it goes wrong.
  • Check facts, sources and figures before you use them. AI can sound sure even when it is wrong.
  • Be careful about what you type in. Personal details, health information and business secrets do not belong in a chatbot when you do not know how it handles data.
  • Be sceptical of calls and videos in which a relative or a famous person asks you for money or recommends an investment. Voices and faces can be made with AI.
  • Learn the basics. The free course Elements of AI, from the University of Helsinki and MinnaLearn, is available in Swedish through Linköping University. Our glossary explains the terms.

Frequently asked questions about AI

What does AI stand for?

AI stands for artificial intelligence. The term was coined in 1955 by the computer scientist John McCarthy.

Who invented AI?

No single person. Alan Turing asked as early as 1950 whether machines can think, and John McCarthy coined the term artificial intelligence in 1955, in a proposal for a summer study at Dartmouth College the following year. Thousands of researchers have built on it since.

What is the difference between AI and machine learning?

AI is the overall term for machines that can do tasks that require intelligence. Machine learning is one way to build AI, in which the program learns from data instead of following rules someone has written. Almost all AI today is built on machine learning, but not all AI is.

Is ChatGPT AI?

Yes. ChatGPT is a chatbot built on a large language model, which makes it generative AI. It has been trained on vast amounts of text and writes its answers by predicting one word at a time.

Can AI think?

It depends on what you mean by thinking, and researchers disagree. Today’s AI can solve hard problems but also makes mistakes no person would make. Whether an AI can become conscious is discussed in Can an AI become conscious?

Is AI dangerous?

AI is already used for fraud and deepfakes, can give wrong answers and can make unfair decisions if it has learned from skewed data. Experts disagree about how big the future risks are. Read more under Is AI dangerous? above.

Will AI take our jobs?

About 60 percent of jobs in advanced economies are exposed to AI, according to the International AI Safety Report 2026. So far there is no effect on overall employment, and economists disagree about what will happen next.

How much energy does AI use?

A single question to a chatbot uses little electricity, but AI companies are building data centres that use as much electricity as cities. The numbers are in How much energy does AI use?

What is AGI?

AGI, artificial general intelligence, is AI that can do roughly any intellectual work a person can. It does not exist yet, and researchers disagree about when, or whether, it will come. More in What is AGI, really?

How can I learn more about AI?

Try it yourself, ideally on something you know well. The free course Elements of AI is available in many languages, and our glossary explains the terms. More in-depth articles are listed below.

Read more

This article is a beginning. Our in-depth articles pick up where it ends.

Our view

This section is SI-bolaget’s opinion. The rest of the article reports the research. Here we say what we ourselves think.

AI is not one thing but a whole family of techniques, and it is growing faster than any other technology of our time. That is why we think:

  • A right of public access to intelligence. Everyone should be able to understand what AI is and use it, not just those who build it. Knowing about AI is part of a general education.
  • People at the centre. AI should make us wiser and freer, not replace our judgement. Whoever uses AI keeps the responsibility for what they do with the answers.
  • Humility. Today’s AI is jagged: brilliant one moment and wrong the next. We do not know how far it will go, and that is a reason to follow its development closely.

We believe AI may become the most significant thing humanity has created. That is exactly why it matters that as many people as possible understand what it is.

Sources

  1. J. McCarthy, M. L. Minsky, N. Rochester and C. E. Shannon, A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, 31 August 1955, and Stanford School of Engineering, on John McCarthy, October 2011.
  2. A. M. Turing, Computing Machinery and Intelligence, Mind 59, 1950.
  3. Government Offices of Sweden, Nationell inriktning för artificiell intelligens, 2018, Regulation (EU) 2024/1689 (the AI Act), Article 3(1) and recital 12, Official Journal of the European Union, 12 July 2024, and European Commission, Commission Guidelines on the definition of an artificial intelligence system established by Regulation (EU) 2024/1689 (AI Act), including paragraphs 32 to 38, 29 July 2025.
  4. Stanford University, One Hundred Year Study on Artificial Intelligence (AI100), 2016 Report: Defining AI, September 2016, and Larry Tesler, Adages and Coinages.
  5. John McDermott, R1: An Expert in the Computer Systems Domain, AAAI 1980.
  6. A. L. Samuel, Some Studies in Machine Learning Using the Game of Checkers, IBM Journal, July 1959, and IBM, on Arthur Samuel.
  7. Meta, Introducing Llama 3.1, July 2024, and Introducing Meta Llama 3, April 2024.
  8. Google, What is clustering?, from the Machine Learning course, 2025.
  9. Google DeepMind, AlphaGo.
  10. International AI Safety Report, International AI Safety Report 2026, February 2026, and International AI Safety Report 2025, January 2025.
  11. The Royal Swedish Academy of Sciences, press release on the Nobel Prize in Physics 2024, 8 October 2024, popular science background and Scientific Background to the Nobel Prize in Physics 2024.
  12. D. E. Rumelhart, G. E. Hinton and R. J. Williams, Learning representations by back-propagating errors, Nature 323, 1986, and Y. LeCun, Y. Bengio and G. Hinton, Deep learning, Nature 521, 2015.
  13. Ian Goodfellow, Yoshua Bengio and Aaron Courville, Deep Learning, MIT Press, 2016, chapter 1 and chapter 6.
  14. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NIPS 2012.
  15. F. A. C. Azevedo et al., Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain, Journal of Comparative Neurology, 2009.
  16. Apple, press release on iPhone X and Face ID, September 2017.
  17. Google, on Gmail’s protection against spam, December 2024.
  18. Google DeepMind, Traffic prediction with advanced Graph Neural Networks, September 2020.
  19. Spotify, Förstå rekommendationer and DJ (in Swedish).
  20. Google Research, A Neural Network for Machine Translation, at Production Scale, September 2016.
  21. Google, on AI Mode in Google Search in Sweden (in Swedish), 8 October 2025.
  22. Lund University, on AI support in breast cancer screening, 30 January 2026, with the Swedish version, and Hernström et al., Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI), The Lancet Digital Health, 2025.
  23. SMHI, on the AI model in the precipitation forecast, November 2025, and ECMWF, on its AI forecasts, 25 February 2025.
  24. KBLab, National Library of Sweden, on KB-Whisper, March 2025.
  25. Government Offices of Sweden, Sverige kliver fram på AI-området, September 2025.
  26. James Lighthill, Artificial Intelligence: A General Survey, Science Research Council, 1973, with an overview of its consequences, IEEE Spectrum, Freddy the Robot Was the Fall Guy for British AI, April 2025, and Jon Agar, What is science for? The Lighthill report and the purpose of artificial intelligence research, UCL, 2020.
  27. Melanie Mitchell, Why AI is Harder Than We Think, 2021.
  28. IBM, Deep Blue.
  29. Vaswani et al., Attention Is All You Need, 2017.
  30. History.com, on the launch of ChatGPT, Reuters via Cyprus Mail, ChatGPT sets record for fastest-growing user base, February 2023, and The Decoder, on OpenAI’s user figures, September 2026.
  31. The Royal Swedish Academy of Sciences, The Nobel Prize in Physics 2024 and The Nobel Prize in Chemistry 2024, October 2024, and Varadi et al., AlphaFold Protein Structure Database in 2024, Nucleic Acids Research, 2024.
  32. Google DeepMind, Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad, July 2025.
  33. J. Sevilla et al., Compute Trends Across Three Eras of Machine Learning, 2022, and Epoch AI, Trends in AI, February 2026.
  34. Epoch AI, Data on AI models, CC BY 4.0, read on 15 November 2026, and the United Nations, World Population Prospects 2024.
  35. Center for Security and Emerging Technology, AI Chips: What They Are and Why They Matter, April 2020.
  36. J. Deng et al., ImageNet: A large-scale hierarchical image database, CVPR 2009.
  37. P. Villalobos et al., Will we run out of data? Limits of LLM scaling based on human-generated data, 2024.
  38. A. Ho et al., Algorithmic progress in language models, 2024.
  39. Rich Sutton, The Bitter Lesson, March 2019.
  40. Stanford HAI, The 2026 AI Index Report, April 2026.
  41. Stanford University, AI100 2026 Report: AI Goes Mainstream, November 2026.
  42. Finansinspektionen, on AI fraud (in Swedish), 6 March 2026.
  43. Lighthouse Reports, Sweden’s Suspicion Machine, November 2024, and the Swedish Authority for Privacy Protection (IMY), Avslutad tillsyn efter att Försäkringskassan tagit AI-system ur bruk, 18 November 2025, and IMY granskar Arbetsförmedlingens AI-system för arbetssökande, 5 November 2026.
  44. European Commission, AI Act, updated 19 October 2026, and Regulation (EU) 2024/1689, Articles 5 and 50.
  45. Statistics Sweden (SCB), use of generative AI tools, 2024 to 2026, Statistical Database, October 2026, and Artificiell intelligens i Sverige 2025.
  46. Eurostat, 20% of EU enterprises use AI technologies, 11 December 2025, and dataset isoc_eb_ai, accessed 15 November 2026.
  47. The Internet Foundation, Svenskarna och internet 2026, chapter on AI, September 2026.
  48. University of Gothenburg, on the SOM Institute’s survey, April 2026.
  49. Government Offices of Sweden, AI-kommissionen överlämnar sina förslag till regeringen, 26 November 2024, the AI Commission, Färdplan för Sverige, SOU 2025:12, and Sveriges första heltäckande AI-strategi ska ge en topp tio-placering, 20 February 2026.
  50. AI Sweden, GPT-SW3.
  51. Elements of AI, A free online introduction to artificial intelligence for non-experts.

Terms in this article

How we work: the articles are written with the help of AI, and every fact is checked against the sources listed above. Found a mistake? Write to hej@sibolaget.se.

More questions?

The glossary explains the terms, and the front page has straight answers to the most common questions. Missing something? Write to us.