What AI Is and Where It Came From

Objectives

Introduction: why start with definitions and history

Artificial intelligence is one of those phrases that everybody uses and few people stop to define. You meet it in news headlines, in product advertising, in conversations at work and in the chat window of tools such as ChatGPT. Sometimes it is described as a helpful assistant, sometimes as a threat to jobs, and sometimes as something close to magic. Because the word is used so loosely, it is easy to come away with a picture that is either far too grand or far too small. This opening chapter exists to give you a solid, plain-language foundation on which the rest of the book can build.

We will do three things. First, we will settle on a working definition of AI and, just as importantly, look at what AI is not, so that you can separate the real technology from the science-fiction version. Second, we will untangle four terms that are often used as if they meant the same thing: artificial intelligence, machine learning, deep learning and generative AI. You will see that they fit inside one another like a set of nesting boxes, and a diagram will make that relationship visible. Third, we will walk through the history of the field, from the first mathematical sketches of artificial neurons in the 1940s, through the famous 1956 workshop at Dartmouth College where the field got its name, through two long periods of disappointment known as the AI winters, and on to the rapid, industry-led boom that has put AI into everyday life.

History is not included here as decoration. Knowing where AI came from explains a great deal about how it behaves today. Many of the ideas inside modern chatbots are decades old; what changed was the amount of data, the cheapness of computing power and a handful of clever engineering breakthroughs. Understanding that story will also help you think clearly about hype. The field has been through cycles of excitement and disappointment before, and recognising the pattern is one of the most useful skills a non-specialist can have.

This is the first chapter of the book, so it assumes no prior knowledge at all. By the end, you should be able to explain to a friend what AI is, how the main terms relate, which milestones matter and why progress has sped up so dramatically. The next chapter then opens the box and shows how machines actually learn from data.

A working definition of artificial intelligence

There is no single definition of AI that every researcher agrees on, but most useful definitions share a common core. IBM describes AI as technology that enables computers and machines to simulate human abilities such as learning, comprehension, problem solving, decision making, creativity and autonomy. The key word in that description is simulate. An AI system does not need to think the way a person thinks; it needs to produce behaviour that, in some useful way, resembles what we would call intelligent if a human did it.

In practice, that covers a wide range of things. Systems with AI can recognise objects in photographs, understand and respond to spoken or written language, learn from new information, make recommendations, and in some cases act on their own, with a self-driving car being the classic example. When your phone groups your photos by the faces in them, when your bank flags a card payment as suspicious, when a streaming service suggests a film, or when a chatbot drafts an email for you, some form of AI is at work.

A helpful way to hold this in your head is to think of AI as a goal rather than a single technique. The goal is to get machines to do tasks that normally require human judgement. Over the decades, researchers have tried many different routes towards that goal. Some wrote down rules by hand; others built programs that search through possibilities; and the approach that dominates today lets the machine learn patterns from large amounts of example data. All of these count as AI, because they all aim at the same target, even though they work in very different ways.

This also explains why the boundary of AI seems to move over time. Tasks that once looked like the height of machine intelligence, such as playing chess well or reading handwritten postcodes, became ordinary software once they were solved. People sometimes joke that AI is whatever computers cannot do yet. The joke contains a truth: our sense of what counts as intelligent behaviour keeps shifting as machines master more tasks. For this book, we will use a practical definition. AI is the field of building computer systems that perform tasks we would normally associate with human intelligence, such as perceiving, understanding language, learning, reasoning and deciding.

What AI is not

Clearing away misconceptions is just as important as stating a definition, because the popular image of AI is shaped more by films and novels than by real systems. Several myths are worth addressing directly.

The first myth is that today's AI is a general-purpose mind. Researchers distinguish between narrow (or weak) AI, which is designed to perform a specific task or set of tasks, and artificial general intelligence (AGI, or strong AI), which would be able to understand, learn and apply knowledge across a wide range of tasks at or beyond human level. Every AI system in use today is narrow in this sense. Even a chatbot that can write poetry, summarise a contract and explain a recipe is still doing one underlying thing very well, namely predicting suitable text, and it fails in ways no thoughtful human would. IBM notes that AGI remains theoretical and that the self-aware machines of science fiction remain in fiction.

The second myth is that AI understands the world the way you do. A system that classifies images has no idea what a cat is beyond the statistical patterns of pixels it has learned to associate with the label. A language model, as later chapters will show, works by estimating which piece of text is most likely to come next. That ability can look remarkably like understanding, and it is genuinely useful, but it is not the same as having experiences, intentions or common sense.

The third myth is that AI is always right, or at least objective. AI systems learn from data, and data reflects the world that produced it, including its gaps and biases. Language models also produce confident statements that are simply false, a behaviour usually called hallucination. On one recent benchmark, hallucination rates across 26 leading models ranged from 22% to 94%, and the models struggled especially when a user personally held a false belief. The lesson is not that AI is useless, but that its outputs need checking in proportion to the stakes.

The fourth myth runs in the other direction: that AI is just hype and nothing more than a fancy search engine or autocomplete. That view underestimates the technology. AI systems now do real work in medicine, science, software development, customer service and many other fields, and later chapters look at both the benefits and the measured limits. A clear-eyed view sits between the extremes. AI is a powerful family of pattern-finding and pattern-generating tools, built by people, trained on data, and shaped by the choices of the organisations that build and deploy it. It is neither a digital person nor a parlour trick.

Two ways to make a machine act intelligently: rules and learning

To see how the different branches of AI relate, it helps to understand the two broad strategies researchers have used to make computers behave intelligently.

The first strategy is to write the intelligence down by hand. A programmer, often working with a human expert, spells out rules: if the patient has this symptom and that test result, consider this infection; if an email contains these phrases, mark it as spam. This approach is usually called symbolic AI or rule-based AI, because it manipulates symbols and logical rules. It dominated the early decades of the field. John McCarthy's programming language Lisp, created in 1958, became the favourite tool of AI researchers precisely because it was good at handling symbolic information. In the 1960s and 1970s, so-called expert systems captured specialist knowledge in rules. DENDRAL helped chemists reason about molecular structures, and MYCIN, developed at Stanford, suggested diagnoses for bacterial infections and recommended antibiotics, although ethical and legal concerns meant it was never used in clinical practice.

Rule-based systems have real strengths. Their reasoning is transparent, because you can read the rules, and they behave predictably. Their weakness is that the world is messy. Writing rules for every situation is slow and expensive, rules contradict one another, and many things people do effortlessly, such as recognising a friend's face or understanding sarcasm, are almost impossible to describe as explicit rules.

The second strategy is to let the machine learn from examples. Instead of telling the computer how to recognise spam, you show it thousands of emails that people have already labelled as spam or not spam, and an algorithm adjusts itself until it can tell them apart. This is machine learning: creating models by training an algorithm on data so that it can make predictions or decisions without being explicitly programmed for each specific task. The term itself dates from 1959, when Arthur Samuel built a checkers program that improved with experience until it played better than he did.

The short Python example that accompanies this chapter shows the difference in miniature. One function decides whether a message is spam using a hand-written rule. The other uses the scikit-learn library to learn a decision from a handful of labelled examples. You do not need to run it yet; Part I of the book will walk you through running and modifying machine learning code step by step. For now, notice the shift in where the knowledge comes from. In the first function, the programmer supplies it. In the second, the data does.

A small case study shows why that shift mattered in the real world. Early email spam filters were largely rule-based: administrators maintained lists of suspicious words and senders. Spammers quickly learned to misspell words, hide text in images and rotate addresses, so the rule lists became a never-ending game of catch-up. Filters that learned statistical patterns from large numbers of messages that users had marked as junk adapted far better, because every new batch of labelled mail taught them something. Modern filters combine learned models with some rules, which is a common pattern: learning handles the messy, shifting bulk of the problem, while rules encode a few firm policies. Most of the AI you meet today is built on the learning strategy, which is why the rest of this chapter, and most of this book, focuses on it.

Nesting boxes: AI, machine learning, deep learning and generative AI

Now we can place the four headline terms in relation to one another. A simple way to picture them, used by IBM among others, is as a series of nested concepts that emerged over more than seventy years: machine learning sits inside AI, deep learning sits inside machine learning, and generative AI is built on deep learning. The diagram for this chapter draws this as boxes within boxes.

Artificial intelligence is the outermost box. It includes everything aimed at making machines behave intelligently, from hand-written expert systems and game-playing search programs to the latest chatbots.

Machine learning is the large region inside AI where systems learn from data instead of following only hand-written rules. It contains many different techniques, including linear and logistic regression, decision trees, random forests, support vector machines, k-nearest neighbours and clustering, each suited to different kinds of problems. Machine learning methods are commonly grouped into three families: supervised, unsupervised and reinforcement learning, which differ in what the system is trained to achieve and usually in the kind of data it uses. In supervised learning, the simplest form, humans pair each training example with a correct label, and the model learns to map inputs to outputs so it can predict labels for new data. Chapter 3 explores the three families in detail.

Deep learning is a subset of machine learning built on artificial neural networks with many layers. A neural network is made of interconnected layers of simple units, loosely inspired by neurons in the brain. Classic neural networks typically have one or two hidden layers between input and output; deep networks have many more, often hundreds. Those extra layers let the network discover useful features in raw data by itself. Instead of a person deciding that edges, corners and textures matter in a photo, the network learns its own internal representations from large, often unlabelled datasets. Some form of deep learning powers most of the AI applications in daily life today, including speech recognition, image recognition and language translation.

Generative AI is the innermost box in our picture. It refers to deep learning models that create new content such as text, images, video, audio or code in response to a prompt. Earlier AI systems mostly analysed things: is this transaction fraudulent, what object is in this photo? Generative systems produce things: write a paragraph, draw an image, compose a melody. They work by learning a compressed representation of their training data and then drawing on it to produce new outputs that resemble, but do not simply copy, what they were trained on.

Two cautions make the picture more accurate. First, the boxes are a teaching simplification, not rigid walls. Some generative models are built with techniques that also appear elsewhere, and practitioners mix methods freely. Second, the inner boxes are not automatically better than the outer ones. A simple machine learning model such as a decision tree is often the right choice for a small business dataset, being cheaper, faster and easier to explain than a deep network. Choosing the right tool for the job is a theme we will return to throughout the book.

When you hear a product described as AI-powered, a useful habit is to ask which box it really lives in. Is it following hand-written rules, learning patterns with classic machine learning, using a deep neural network, or generating new content? The answer tells you a lot about what it can do, what data it needed, and how it is likely to fail.

A first look inside generative AI and large language models

Because generative AI is the part of the field most people meet first, through tools like ChatGPT, it is worth a short preview here. Later chapters cover it properly.

The most familiar generative systems are large language models, or LLMs. A language model estimates how likely a piece of text is; in particular, it estimates the probability of the next token, where a token may be a whole word, part of a word or a single character. In English, a token is on average roughly four characters long. When you type a question into a chatbot, the model repeatedly predicts a likely next token, adds it to the text, and predicts again, building the answer one small piece at a time.

The letters in the name GPT summarise the approach: the model is Generative because it produces text, Pre-trained because it first learns from an enormous body of text, and a Transformer because of the kind of neural network it uses. The transformer is now the dominant architecture for LLMs. Its central trick, called attention, lets each token in the input weigh how much every other token affects its meaning. That is how the model can tell that the word model means something different in a sentence about machine learning than in a sentence about fashion.

Raw pre-trained models are good at continuing text but not necessarily good at being helpful assistants. Developers therefore add further training stages. One widely used technique is reinforcement learning from human feedback (RLHF), in which people rank the model's possible answers, those rankings train a separate reward model, and the language model is then fine-tuned to produce answers the reward model scores highly. This is part of why chat assistants feel conversational and follow instructions, and it is also a reminder that human choices shape how these systems behave.

You do not need to master any of these details yet. The important point for this chapter is where generative AI sits in the bigger picture: it is a particular application of deep learning, which is a particular kind of machine learning, which is one approach to the broader goal of artificial intelligence. With that map in place, we can turn to the story of how the field arrived here.

Early ideas: artificial neurons and the Turing Test

People have imagined thinking machines for a very long time, in myths, mechanical automata and fiction. In 1726, Jonathan Swift's Gulliver's Travels satirised a contraption that produced books by mechanically shuffling words, an odd foreshadowing of automated text generation. The scientific story of AI, however, begins in the middle of the twentieth century, when electronic computers made it possible to test such ideas rather than merely dream about them.

One starting point is a 1943 paper by the neurophysiologist Warren McCulloch and the logician Walter Pitts. They proposed that the brain could be understood as a computational system and described simplified mathematical neurons that switch on or off depending on their inputs. This was the seed of the artificial neural network, the idea at the heart of today's deep learning. It is striking that the technology behind modern chatbots traces back to a wartime paper about logic and nerve cells, written before practical general-purpose computers existed.

The second foundational moment came in 1950, when the British mathematician Alan Turing published Computing Machinery and Intelligence. Turing opened with the question "Can machines think?" and then sidestepped it, because words like think are so hard to pin down. Instead he proposed a practical test he called the imitation game, now known as the Turing Test. In its usual modern form, a human judge holds typed conversations with a hidden person and a hidden machine. If the judge cannot reliably tell which is which, the machine has, in Turing's operational sense, shown intelligent behaviour.

The Turing Test has been criticised from many angles. It rewards imitation rather than understanding, and it focuses on conversation rather than on the many other things intelligence involves. Modern chatbots can hold long, fluent conversations, yet as we saw earlier they still hallucinate and lack common sense, which shows that sounding human and being reliably intelligent are different things. Even so, Turing's move was enormously influential. He turned a philosophical puzzle into a question that could be investigated, and he framed intelligence in terms of observable behaviour. That framing, judging machines by what they do rather than what they are, still underlies how AI systems are evaluated today, from benchmarks to user testing.

Dartmouth and the optimistic early years

The field received its name in 1955, when four researchers wrote a proposal for a summer research project: John McCarthy of Dartmouth College, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Telephone Laboratories. The proposal introduced the term artificial intelligence, and the workshop itself took place at Dartmouth in the summer of 1956. It is generally treated as the official birth of AI as a field of research.

The Dartmouth proposal was famously confident. Its authors believed that every feature of learning and intelligence could, in principle, be described precisely enough for a machine to simulate it, and they hoped a small group could make real progress in a single summer. They did not solve intelligence that summer, of course, but the meeting gave a scattered set of ideas a shared name, a community and an agenda.

The following years produced a burst of creative work. Allen Newell and Herbert Simon built programs such as the Logic Theorist and the General Problem Solver, which imitated human step-by-step problem solving. In 1957, the psychologist Frank Rosenblatt developed the Perceptron, an early neural network that learned to classify patterns by adjusting the weights on its inputs. The Perceptron was genuinely important because it learned from data, but it could only handle problems in which the categories could be separated by a straight line, a limitation that would later matter a great deal. In 1959, Arthur Samuel's self-improving checkers program gave us the phrase machine learning.

The 1960s added more milestones. Joseph Weizenbaum's ELIZA, developed in the mid-1960s, mimicked a conversation by matching patterns in typed input. Weizenbaum meant it to show how shallow such conversation was, and he was surprised by how readily users attributed feelings to it. That reaction, people reading human qualities into fluent machines, is very much alive in the age of chatbots. At SRI, the robot Shakey combined perception, planning and problem solving to reason about its own actions.

Optimism ran high. In 1970, Minsky was quoted predicting that machines with the general intelligence of an average human might arrive within three to eight years. Predictions like this were sincere, but they badly underestimated how hard ordinary human abilities, such as seeing, moving and understanding everyday language, would turn out to be. The gap between promise and delivery was about to have consequences.

The AI winters: when promises outran results

When expectations soar and results lag, money and attention eventually dry up. AI has been through this cycle at least twice, and the resulting downturns are known as AI winters.

The first warning signs concerned neural networks. In 1969, Marvin Minsky and Seymour Papert published Perceptrons, a rigorous analysis of the limits of single-layer neural networks. The book is often blamed for cooling interest in neural network research for years, although its authors later argued that progress had already stalled for lack of theoretical understanding. Either way, the learning approach fell out of fashion, and rule-based methods took centre stage.

The broader crisis came in the 1970s. In 1973, the mathematician James Lighthill delivered a report to the British Science Research Council concluding that AI had failed to deliver on its early promises. Government funding for AI in the UK was sharply reduced, and the report contributed to the first AI winter, a period of reduced interest and investment. Similar scepticism spread among funders elsewhere.

The field recovered in the early 1980s, largely on the back of expert systems, which companies hoped would capture and sell specialist knowledge. Japan launched its ambitious Fifth Generation Computer Systems project in 1982, aiming at machines capable of logical reasoning and natural language processing. Yet the same pattern repeated. At a 1984 meeting of the Association for the Advancement of Artificial Intelligence, Roger Schank and Marvin Minsky warned that inflated expectations would lead to a collapse in investment. Within a few years they were proved right, and the second AI winter set in. Expert systems turned out to be costly to build and maintain, brittle outside their narrow domains and hard to update as knowledge changed, the very weaknesses of rule-based AI we discussed earlier.

The AI winters are worth remembering for a practical reason. IBM's history of AI describes them as evidence of the cyclical nature of AI hype, in which bursts of optimism are followed by disillusionment when the technology fails to meet expectations. Today's boom rests on far stronger foundations than earlier ones, as we will see, but the lesson still applies. When you read a confident prediction about what AI will do in the next few years, it is worth asking what evidence supports it and remembering how often such predictions have missed in the past.

The quiet rebuild: backpropagation, statistics and data

Winters are not periods when nothing happens. Some of the most important groundwork for modern AI was laid while public interest was low.

The most important development was a practical way to train neural networks with several layers. The core mathematical idea, now called backpropagation, was introduced by Arthur Bryson and Yu-Chi Ho in 1969 for control systems. In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams showed how to apply it to neural networks, letting a network adjust its internal weights by passing the error from its output backwards through its layers. This overcame the single-layer limitation that had dogged the Perceptron. Backpropagation remains the standard training algorithm for neural networks today, and modern software libraries perform it automatically so that users never calculate it by hand. You will meet the idea again in the next chapter.

Practical successes followed, though at a modest scale. In 1989, Yann LeCun and colleagues at AT&T Bell Labs trained a multilayer network to recognise handwritten ZIP codes, one of the first practical uses of what we now call convolutional neural networks. Training took about three days on the hardware of the time. In 1997, Sepp Hochreiter and Jürgen Schmidhuber introduced the LSTM network, which handled sequences such as speech and handwriting better than earlier designs.

A second, less glamorous shift was the move from rules to statistics. In 1988, IBM researchers described a statistical approach to language translation that learned from millions of pairs of English and French sentences, rather than trying to encode grammar by hand. The same year, Judea Pearl's work on Bayesian networks gave AI rigorous tools for reasoning under uncertainty. Across the field, researchers increasingly let data, rather than hand-crafted rules, carry the knowledge.

The era also produced public milestones. In 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov in a six-game match under standard tournament conditions. Deep Blue relied mostly on enormous search power and carefully engineered evaluation, not on learning as modern systems do, but it showed the world that machines could beat the best humans at a game long regarded as a test of intellect. In 2011, IBM Watson beat two champions at the quiz show Jeopardy!, demonstrating sophisticated handling of natural-language questions.

By the end of the 2000s, then, many key ingredients existed: neural networks that could be trained in layers, a culture of learning from data, and growing collections of digital information. What was missing was enough data and computing power to make deep networks truly shine.

The deep learning breakthrough: data, GPUs and ImageNet

Two ingredients arrived almost together in the late 2000s, and their combination sparked the modern era.

The first was data on a new scale. In 2007, Fei-Fei Li and her team began the ImageNet project, building a database of millions of labelled images across thousands of categories to support research in visual object recognition. The internet, smartphones and cheap storage were creating vast quantities of text, images and video, but ImageNet showed the value of carefully organised, labelled data for training and comparing models.

The second was suitable hardware. Graphics processing units, or GPUs, were designed to render video-game graphics, which requires performing huge numbers of simple calculations in parallel. Neural network training turns out to need exactly that kind of arithmetic. In 2009, Rajat Raina, Anand Madhavan and Andrew Ng showed that GPUs could far outperform ordinary processors for deep learning, which helped push the field to adopt them. Stanford's AI Index reports that the cost of GPU computation has fallen by more than 99% since 2006, a decline that made later scaling economically possible.

The moment these ingredients came together is a short case study in how a field changes direction. Each year, the ImageNet challenge asked teams to build systems that classified images into categories, with results measured by error rate. In 2012, a team from the University of Toronto led by Geoffrey Hinton entered a deep convolutional neural network trained on GPUs, later known as AlexNet. It achieved an error rate of about 16%, against roughly 25% for the best result the year before. In a competition where progress usually came in small steps, a jump of that size was a shock. Within a few years, researchers in computer vision, speech recognition and language processing had largely switched to deep learning, and technology companies began hiring neural network researchers in large numbers. The lesson of the AlexNet story is that the underlying ideas, layered neural networks and backpropagation, were not new. What changed was that enough data and cheap enough computation finally allowed those ideas to work at scale.

More landmarks followed quickly. In 2016, DeepMind's AlphaGo defeated Lee Sedol, one of the world's top Go players, by four games to one; Go had long been considered far harder for computers than chess because of its vast number of possible positions. In 2020, DeepMind's AlphaFold 2 made a breakthrough in predicting the three-dimensional shapes of proteins from their sequences, a problem that had challenged biologists for decades. The importance of this line of work was recognised in 2024, when the Nobel Prizes in Physics and Chemistry honoured foundational work on neural networks and on AI-driven protein structure prediction.

From transformers to ChatGPT: AI goes mainstream

The final step towards today's generative AI came from a new neural network design. In 2017, researchers at Google introduced the transformer architecture, built around the attention mechanism described earlier in this chapter. Transformers process all the words in a passage together rather than strictly one after another, which made them well suited to training on very large amounts of text using many GPUs at once. The architecture quickly became the dominant design for large language models.

OpenAI's GPT-1 appeared in 2018, and in 2020 GPT-3, with 175 billion parameters, showed that a single very large model could perform a wide range of language tasks, such as answering questions, translating, writing code and producing creative text, with little or no task-specific training. Parameters are the adjustable numbers inside a model that training tunes, and GPT-3's size was unprecedented at the time. Image generation advanced in parallel: DALL·E, introduced in 2021, generated images from text descriptions, helping establish text-to-image generation as a major area of generative AI.

The turning point for public awareness came in November 2022, when OpenAI released ChatGPT, built on GPT-3.5, as a free research preview. Its conversational interface let anyone ask questions, request revisions and follow up in ordinary language, and it spread with extraordinary speed. GPT-4 followed in March 2023, handling more complex tasks and accepting images as well as text. Competitors such as Google's Gemini appeared, open-weight models such as Meta's Llama family made powerful models available for others to download and adapt, and governments responded too: the European Union adopted its AI Act in 2024, setting out a risk-based framework for regulating AI. By 2025, much development effort had shifted to AI agents, systems that can plan tasks, use software tools and carry out multistep work on a user's behalf.

The pace of adoption is part of what makes this period distinctive. Stanford's 2026 AI Index estimates that generative AI reached about 53% population adoption within three years, faster than either the personal computer or the internet, though with large differences between countries. The same report finds that 88% of surveyed organisations used AI in 2025 and 70% used generative AI in at least one business function. Few technologies have moved from research laboratories into daily life so quickly.

Why AI has accelerated so quickly

Looking back over this history, it is natural to ask why progress, after decades of fits and starts, has become so fast. There is no single cause. Instead, several forces have reinforced one another.

The first force is data. Learning systems improve with more and better examples, and the digital world now produces examples on a staggering scale: web pages, books, photos, videos, code, sensor readings and records of countless transactions. Foundation models are trained on huge volumes of raw, mostly unlabelled data drawn from sources such as the internet. Quantity is not everything, though. The AI Index notes that careful data curation can let much smaller models match far larger ones on some benchmarks, a reminder that the quality of data matters as much as its volume.

The second force is cheap, plentiful computation. As we saw, the cost of GPU computation has fallen by more than 99% since 2006. Training a frontier model still requires thousands of clustered GPUs running for weeks and typically costs millions of dollars, but such projects would have been unthinkable at earlier prices. Cloud computing has also made powerful hardware available to people who could never buy it outright; even a free browser-based notebook service such as Google Colab gives learners access to GPUs for experiments.

The third force is better algorithms and architectures. Backpropagation made deep networks trainable, convolutional networks suited them to images, and transformers let them scale on text. Training techniques such as RLHF turned raw text predictors into usable assistants. Each advance multiplied the value of the data and computation available.

The fourth force is money and industry. Once deep learning showed commercial value, technology companies invested on a vast scale, and the centre of gravity moved from universities to industry. The AI Index reports that industry produced more than 90% of notable AI models in 2025, and that the most capable models disclose less about their training data, size and compute than before. Commercial competition speeds up development, but it also concentrates power in a few organisations and makes independent scrutiny harder, a concern echoed by falling scores on transparency measures.

These forces form a feedback loop. Better models attract users and investment, investment buys more computation and data, and these produce better models still. The loop also has costs. Models have grown faster than efficiency has improved, so the power needed to train frontier systems keeps rising, with the largest training runs drawing more than 100 megawatts. Estimated training emissions have risen from about 0.01 tonnes of CO2-equivalent for AlexNet in 2012 to about 72,800 tonnes for a leading 2025 model, although these figures are uncertain estimates. Documented AI incidents have also climbed, with 362 logged in 2025 compared with 233 in 2024.

For a thoughtful non-specialist, the key insight is that today's acceleration is real and rests on solid foundations, unlike some earlier booms, yet it is not a law of nature. It depends on continuing flows of data, energy, money and public trust. Keeping that in view will help you read both enthusiastic and alarmed headlines with a clear head, a skill later parts of this book develop further.

Summary and what comes next

This chapter has given you a map and a timeline that the rest of the book will build on.

We began with a working definition: artificial intelligence is the field of building computer systems that simulate abilities we associate with human intelligence, such as learning, understanding language, solving problems and making decisions. AI is best understood as a goal that many different techniques have pursued, which is why the boundary of what counts as AI keeps moving. We then cleared away common myths. Every AI system in use today is narrow rather than general; systems produce intelligent-looking behaviour without human-like understanding; they can be wrong, biased or confidently mistaken; and yet they are far more than hype, doing real work across many fields.

Next we contrasted two strategies for building intelligent behaviour. Rule-based, symbolic AI writes knowledge down by hand, which makes it transparent but brittle. Machine learning lets systems learn patterns from data, which makes them adaptable and is the basis of nearly all modern AI. From there we built the nesting-box picture. AI is the outer box; machine learning sits inside it; deep learning, using neural networks with many layers, sits inside machine learning; and generative AI, which creates new text, images and other content, is built on deep learning. Large language models such as the one behind ChatGPT are generative, pre-trained transformers that produce text by repeatedly predicting a likely next token.

The history showed how we arrived here. McCulloch and Pitts sketched artificial neurons in 1943, Turing proposed his imitation game in 1950, and the 1955 Dartmouth proposal named the field before the 1956 workshop launched it. Early optimism gave us the Perceptron, machine learning and ELIZA, but overpromising led to two AI winters. Quiet progress in backpropagation, statistical methods and convolutional networks laid the groundwork, and the combination of large datasets such as ImageNet with cheap GPU computing produced the 2012 AlexNet breakthrough. Transformers then enabled large language models, and ChatGPT brought generative AI to a mass audience in 2022.

Finally, we saw that recent acceleration comes from four forces that reinforce one another: abundant data, dramatically cheaper computation, better algorithms and heavy industry investment. That loop has brought rapid adoption, and with it rising energy demands, reduced transparency and a growing number of incidents.

One question has been left deliberately open: what does it actually mean for a machine to learn? We have used words such as training, model and prediction without opening them up. The next chapter, How Machines Learn: Data, Models, Training and Prediction, does exactly that, following the path from raw data to a trained model making predictions, and showing you how to run that process yourself in Python.

# Two ways to decide whether a message is spam.
# Requires scikit-learn: pip install scikit-learn
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB

# 1. Rule-based AI: a programmer writes the knowledge down by hand.
def rule_based_is_spam(message):
    suspicious_words = ["winner", "free", "prize", "urgent"]
    text = message.lower()
    return any(word in text for word in suspicious_words)

# 2. Machine learning: the knowledge comes from labelled examples.
messages = [
    "You are a winner, claim your free prize now",
    "Urgent: your account wins a cash reward",
    "Cheap watches, limited offer, click here",
    "Are we still meeting for lunch tomorrow?",
    "Here are the notes from today's class",
    "Can you send me the report by Friday?",
]
labels = [1, 1, 1, 0, 0, 0]  # 1 = spam, 0 = not spam

vectorizer = CountVectorizer()          # turns text into word counts
features = vectorizer.fit_transform(messages)
model = MultinomialNB()                 # a simple learning algorithm
model.fit(features, labels)             # training: learn from examples

new_message = "Limited offer: click here for a cash reward"
print("Rule says spam:", rule_based_is_spam(new_message))
prediction = model.predict(vectorizer.transform([new_message]))[0]
print("Model says spam:", bool(prediction))

A hand-written rule compared with a tiny scikit-learn model. The rule misses the new message because it contains none of the listed words; the trained model can flag it because it learned that words such as offer, click and reward appeared in spam examples.

Key takeaways

Common mistakes

See the full ebook: AI for Everyone