Safekipedia

Artificial intelligence

Discoverer · ages 10+Adapted from Wikipedia
Kismet, a robot head designed to show facial expressions and interact with people.

"AI" also has other uses. For those, see AI (disambiguation) and Artificial intelligence (disambiguation).

Artificial intelligence (AI) is the ability of computational systems to do tasks that are usually linked to human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making. It is a field of research in engineering, mathematics, and computer science. Researchers in this field build and study methods and software that help machines sense their surroundings and use learning and intelligence to take actions that give them the best chance of reaching set goals.

Many well-known uses of AI exist today. These include advanced web search engines, chatbots, virtual assistants, autonomous vehicles, and play and analysis in strategy games such as chess and Go. AI is also used for content generation, which can create text, images, audio, and videos.

The traditional goals of AI research include learning, reasoning, knowledge representation, planning, natural language processing, and perception. Research also supports robotics. To reach these goals, researchers use methods such as state space search, mathematical optimisation, formal logic, and artificial neural networks. They also use ideas from statistics, operations research, and economics. AI draws on other fields too, including psychology, linguistics, philosophy, and neuroscience. Some companies, such as OpenAI, Google DeepMind, and Meta, aim to create artificial general intelligence (AGI). This is AI that can complete nearly any cognitive task at least as well as a human.

Artificial intelligence was founded as an academic discipline in 1956. Over its history, the field went through several cycles of optimism. Each one was followed by periods of disappointment and loss of funding, known as AI winters. Funding and interest grew a great deal after 2012, when graphics processing units (GPUs) began being used to speed up neural networks, and deep learning performed better than earlier AI techniques. This growth sped up further after 2017 with the transformer architecture. In the 2020s, an AI boom happened alongside advances in generative AI, which became widespread and made it possible to create and change media.

Along with the use of AI, people have raised concerns about AI safety and unintended consequences and harms. Ethical concerns, AI's long-term effects, environmental effects, and potential existential risks have led to discussions about AI regulation.

Goals

Researchers have broken the big problem of creating intelligence into smaller parts, called subproblems. Early researchers developed algorithms that copied the step-by-step reasoning people use to solve puzzles or make logical deductions. Many of these methods became too slow as problems grew larger. AI programs also use knowledge to answer questions, and large language models learn from the text of millions of books and billions of websites.

An ontology represents knowledge as a set of concepts within a domain and the relationships between those concepts.

A rational agent has goals or preferences and takes actions to make them happen. Machine learning lets programs improve at a task on their own, and there are several kinds, including supervised learning and reinforcement learning. Natural language processing allows programs to read, write, and communicate in human languages. Machine perception uses sensors, such as cameras and microphones, to understand the world. Affective computing is a field that recognises, interprets, and simulates human feelings. A machine with artificial general intelligence would be able to solve a wide variety of problems, much like human intelligence.

Techniques

AI research uses a wide variety of techniques to reach its goals. One kind of search is state space search, which looks through a tree of possible states to find a goal, and planning programs use a process called means-ends analysis to find a path to a target. Simple exhaustive searches are rarely enough, because the search space grows to astronomical numbers and becomes too slow. "Heuristics," or "rules of thumb," help choose the options most likely to reach a goal. Adversarial search is used in game-playing programs such as chess or Go, looking for a winning position.

Illustration of gradient descent for three different starting points; two parameters (represented by the plan coordinates) are adjusted in order to minimise the loss function (the height).

Local search uses mathematical optimisation to improve a first guess step by step. Gradient descent adjusts numbers to minimise a loss function, and variants of it are commonly used to train neural networks through the backpropagation algorithm. Evolutionary computation improves a set of solutions by "mutating" and "recombining" them, selecting only the fittest each generation. Swarm intelligence algorithms include particle swarm optimisation, inspired by bird flocking, and ant colony optimisation, inspired by ant trails. Formal logic is used for reasoning, with propositional logic working on statements that are true or false, and predicate logic working on objects and relations. Deductive reasoning proves a new conclusion from premises that are assumed to be true. Fuzzy logic gives a "degree of truth" between 0 and 1, which helps with statements that are partly true, and non-monotonic logics handle default reasoning.

Many AI problems require working with incomplete or uncertain information. Researchers use tools from probability theory and economics, such as decision theory, game theory, and Markov decision processes, to help an agent make choices and plan. Bayesian networks can be used for reasoning, learning, planning, and perception, and probabilistic methods such as hidden Markov models and Kalman filters help analyse data that changes over time.

A simple Bayesian network, with the associated conditional probability tables

The simplest AI applications are classifiers, which use pattern matching to find the closest match, and they can be improved with supervised learning using labelled examples. The decision tree is the simplest and most widely used symbolic machine learning algorithm, while the support vector machine displaced k-nearest neighbour in the 1990s. The naive Bayes classifier is reportedly the "most widely used learner" at Google.

Expectation–maximisation clustering of Old Faithful eruption data starts from a random guess but then successfully converges on an accurate clustering of the two physically distinct modes of eruption.

An artificial neural network is loosely based on the neurons in a biological brain. It is trained to recognise patterns and can then find them in new data. A network is typically called a deep neural network if it has at least 2 hidden layers. Deep learning uses several layers of neurons to pull out higher-level features, such as edges in an image at lower layers and faces at higher layers. It improves computer vision, speech recognition, and natural language processing, and the reason it works so well was not known as of 2021. Generative pre-trained transformers (GPTs) are large language models that generate text by predicting the next token after learning from a large corpus of text. They power chatbots such as ChatGPT, Claude, and Gemini, and they sometimes produce false statements called "hallucinations."

Main articles: Programming languages for artificial intelligence and Hardware for artificial intelligence

In the late 2010s, graphics processing units (GPUs) replaced central processing units (CPUs) as the main tool for training large machine learning models, often with TensorFlow software. Early AI research used specialised languages such as Prolog, but general-purpose languages like Python are now the most common. The number of transistors in integrated circuits has been observed to roughly double every 18 months, a trend known as Moore's law, named after Gordon Moore of Intel. Improvements in GPUs have been even faster, a trend sometimes called Huang's law, named after Jensen Huang of Nvidia.

Applications

Main article: Applications of artificial intelligence

AI Overviews, an example of AI use on search engines

AI and machine learning are used in many everyday tools. These include search engines such as Google Search, recommendation systems on Netflix, YouTube and Amazon, and virtual assistants such as Siri and Alexa. They also help power autonomous vehicles, including self-driving cars, and automatic language translation like Google Translate. Chatbots and facial recognition are other examples. Games have been used for a long time to test AI. In 1997, Deep Blue became the first computer chess-playing system to beat a reigning world chess champion, Garry Kasparov. In 2016, AlphaGo won 4 out of 5 games of Go against champion Lee Sedol.

AI is also used in health and science. In 2021, AlphaFold 2 showed it could approximate the structure of a protein in hours rather than months. Researchers have also used machine learning to speed up the search for Parkinson's disease drug treatments. In the military, AI is being studied for command and control and for target acquisition. AI has also been used to help with legal research, but judges have issued sanctions and bar associations have issued warnings after lawyers submitted court documents with fake case law hallucinated by AI tools.

Ethics

Main article: Ethics of artificial intelligence

Fuelled by a growth in AI, data centres' demand for power increased in the 2020s.

AI has both possible benefits and possible risks. It may help advance science and solve serious problems. Demis Hassabis of DeepMind hopes to "solve intelligence, and then use that to solve everything else." As AI has spread, though, several unintended consequences and risks have been found. Some AI systems do not always take ethics and bias into account during training, especially deep learning systems whose decisions are hard to explain. AI also raises concerns about privacy and copyright. For example, Amazon recorded millions of private conversations to build speech recognition systems and let temporary workers listen to some of them. Large companies such as Alphabet Inc., Amazon, Apple Inc., Meta Platforms, and Microsoft dominate the commercial AI market. AI also uses a lot of electricity, which has increased greenhouse gas emissions and water scarcity in some places. Other concerns include misinformation, conspiracy theories, and biased systems that can treat people unfairly. The U.S. courts and Google Photos examples show how these problems can appear in real life.

Some experts worry that AI could be misused by authoritarian governments, terrorists, or criminals, and that it could change jobs, with some white-collar workers and paralegals at risk. Others warn about lethal autonomous weapons and even the idea that AI could become so powerful that humanity might lose control of it. Many people disagree about how serious these risks are. Some researchers focus on AI safety and try to make AI aligned with human values, while others point out that AI can also be used for good. Governments are also making rules. The EU Artificial Intelligence Act entered into force on 1 August 2024, and in November 2023 the first global AI Safety Summit was held at Bletchley Park in the UK.

Main article: History of artificial intelligence

For a chronological guide, see Timeline of artificial intelligence.

In 2024, AI patents in China and the US numbered more than three-fourths of AI patents worldwide. Though China had more AI patents, the US had 35% more patents per AI patent-applicant company than China.

Long ago, philosophers and mathematicians began studying mechanical reasoning. Alan Turing's theory of computation showed that a machine could work with simple symbols like "0" and "1" to copy any kind of mathematical reasoning. In 1950, Turing's paper introduced the Turing test. The field of AI was founded at a workshop at Dartmouth College in 1956, where the first AI program, Logic Theorist, was presented. In the 1960s, researchers created programs that could play checkers, solve algebra word problems, and speak English. Many of them believed machines with general intelligence would soon be built, but they had underestimated how hard this would be. Funding was cut in the 1970s, and a period when it was hard to get money for AI projects, called the "AI winter", followed.

AI was revived in the 1980s by expert systems, which copied the knowledge of human experts. Later, Geoffrey Hinton and others brought back neural network research, and in 1990, Yann LeCun showed that convolutional neural networks could recognise handwritten digits. Deep learning began to lead industry tests in 2012, helped by faster computers and large amounts of data. In 2015, AlphaGo, developed by DeepMind, beat the world champion Go player. In 2022, ChatGPT, made by OpenAI, became the fastest-growing consumer software application in history, gaining over 100 million users in two months. This started an AI boom, in which large companies invested billions of dollars in AI research.

Philosophy

Main article: Philosophy of artificial intelligence

The Turing test can provide some evidence of intelligence, but it penalises non-human intelligent behaviour.

Philosophers have long asked what intelligence is and how to make intelligent machines. Alan Turing asked whether machines can show intelligent behaviour and think. In 1950, he proposed the Turing test, which measures whether a machine can copy human conversation. Other ideas from philosophy, such as epistemology and free will, are also linked to AI, and fast progress has led to more public talk about the ethics of AI.

Another big question is whether machines can be conscious or have subjective experiences like people do. Many AI researchers think this does not matter for building machines that solve problems. Some philosophers, such as John Searle, argue that a computer could copy human behaviour without having a mind. Others think that advanced AI that can feel might deserve certain rights or protection, similar to animals.

Future

A superintelligence is a hypothetical agent that would have intelligence far beyond that of the brightest and most gifted human mind. If research into artificial general intelligence produced smart enough software, it might be able to reprogram and improve itself. Each improved version would be even better at improving itself, which I. J. Good called an "intelligence explosion" and Vernor Vinge called a "singularity". In Artificial Intelligence: A Modern Approach, Russel and Norvig doubted that a singularity was possible. They argued that "so far, every technology has followed an S-shaped curve", and that "sometimes it is not possible to keep the growth going."

Main article: Transhumanism

Robot designer Hans Moravec, cyberneticist Kevin Warwick and inventor Ray Kurzweil have predicted that humans and machines may merge in the future into cyborgs that are more capable and powerful than either one alone. This idea, called transhumanism, has roots in the writings of Aldous Huxley and Robert Ettinger. Edward Fredkin argues that "artificial intelligence is the next step in evolution." This idea was first proposed in Samuel Butler's "Darwin among the Machines" as far back as 1863, and was later expanded upon by George Dyson in his 1998 book Darwin Among the Machines: The Evolution of Global Intelligence.

In fiction

The word "robot" itself was coined by Karel Čapek in his 1921 play R.U.R., the title standing for "Rossum's Universal Robots".

Thinking artificial beings have appeared in stories since ancient times, and they have been a lasting theme in science fiction. A common trope began with Mary Shelley's Frankenstein, where a human-made creation becomes a threat to its makers. Examples include Arthur C. Clarke's and Stanley Kubrick's 2001: A Space Odyssey (both 1968), with HAL 9000, a dangerous computer on the Discovery One spaceship, as well as Blade Runner (1982), The Terminator (1984) and The Matrix (1999). Loyal robots, such as Gort from The Day the Earth Stood Still (1951) and Bishop from Aliens (1986), are rarer and less well known in popular culture.

Isaac Asimov introduced the Three Laws of Robotics in many stories, most notably with the "Multivac" super-intelligent computer. People often bring up Asimov's laws when discussing machine ethics. Almost all AI researchers know the laws from popular culture, but they generally consider them useless for many reasons, one of which is that they are unclear. Several works also use AI to make us ask what makes us human, showing artificial beings that have the ability to feel, and so can suffer. This appears in Karel Čapek's R.U.R., the films A.I. Artificial Intelligence and Ex Machina, and the novel Do Androids Dream of Electric Sheep? by Philip K. Dick.

Quiz time!

5 questions about Artificial intelligence

Question 1 of 5

What is artificial intelligence (AI)?

Pictures

A diagram comparing supervised learning, where a computer learns from labeled examples, with unsupervised learning, where it finds patterns in data without labels.
The Raspberry Pi AI Kit, a set of add-on hardware that gives a Raspberry Pi 5 computer extra power for artificial intelligence tasks.
Vice President Kamala Harris poses with world leaders and officials at the 2023 AI Safety Summit in the United Kingdom, where countries discussed shared rules for artificial intelligence.
The Shadow Dexterous Robot Hand gently holds a light bulb, showing how robots can be designed to grip delicate objects.

Keep exploring

This article is a child-friendly adaptation of the Wikipedia article on Artificial intelligence, available under CC BY-SA 4.0.

Images from Wikimedia Commons. Tap any image to view credits and license.