his is a story about stolen intelligence. It is a long but necessary history — about the deceptive illusions of AI, about the Silicon Valley Goliaths and the ordinary Davids who stand against them. It is about secret treasure chests and vast libraries of pilfered work, and about the internet's investigators trying to solve one of the greatest heists in history. It is about what it means to be human, to be creative, to be free. And it is about what the end of humanity — post-humanity, trans-humanity, even the apocalypse — might actually look like.

This is an investigation into what it means to steal, to take, to colonise, to conquer. Along the way, we will learn what AI really is, how it works, and what it can teach us about our own intelligence — about ourselves as human beings. Drawing on some of history's great philosophical and scientific minds, we will try to understand why the speed and tempo of modern life are accelerating at an ever-increasing rate, and why, without organisation and system, the result would be chaos.

We carry with us a deeply held idea that intelligence is something abstract, transcendent, disembodied — something unique and special. But we will see how intelligence is far more about the deep, deep past and the far, far future: something that reaches powerfully through bodies, through people, through infrastructure, through the concrete empirical world.

Sundar Pichai, the CEO of Google, reportedly claimed that AI is one of the most important things humanity has ever worked on — more profound, he said, than electricity or fire. It sounds like the kind of hype you would expect from a big tech CEO. But we will see how it might well be true. Artificial intelligence may change everything — dazzlingly quickly, and just like electricity and fire before it. We need to find new ways of ensuring that this vast, consequential, and truly unprecedented change is used for good, for everyone — not for evil.

We will get to the future. But it is important that we begin with the past.

A History of AI: God Is a Logical Being

Intelligence. Knowledge. Brain. Mind. Cognition. Calculation. Thinking. Logic. We often use these words interchangeably, or at least with a great deal of overlap — but when we drill down into what something like intelligence actually means, we find surprisingly little agreement. Our control over a bewildering environment has been facilitated by new techniques for handling vast quantities of data at incredible speeds. The tool that made this possible is the high-speed digital computer, operating with electronic precision on great quantities of information.

Can machines be intelligent in the same way humans can? Will they surpass human intelligence? What does it really mean to be intelligent? When the first computers appeared, the media referred to them as electronic brains — machines that could process enormous amounts of data at very high speed, saving time and effort.

In Britain in the 1950s, a national debate was already underway about whether machines could think. After all, the computer, even then, was in many ways more capable than any human in certain respects — it could make calculations faster, with fewer mistakes. The father of both the computer and AI, Alan Turing, contributed to this debate in a BBC radio broadcast in 1951, claiming that it was "not altogether unreasonable to describe digital computers as brains." The coincidence between computers, AI, and intelligence strained the idea that intelligence was a single, unified thing. A thorough history would require encompassing transistors, electricity, computers, the internet, logic, mathematics, philosophy, neurology, and society itself. Is there any real understanding of AI without these things? This impossible totality will echo throughout this history. But there are two key moments at which we can begin: the Turing Test and the Dartmouth College Conference.

Turing wrote his now-famous paper, Computing Machinery and Intelligence, in 1950. It began with: "I propose to consider the question: can machines think?" He suggested a test — that if a person, not knowing whether they were conversing with a machine or a human, found the machine indistinguishable from a human, it could be considered intelligent. Ever since, the conditions of the Turing Test have been debated: How long should it last? What kinds of questions should be asked? Should it be text-based only? What about images? Audio? One competition, the Loebner Prize, offered $100,000 to anyone who could pass the test before a panel of judges.

A few years later, in 1955, one of the founding fathers of AI, John McCarthy, and his colleagues proposed a summer research project to debate the question of thinking machines. In choosing a name, McCarthy settled on the term artificial intelligence. In the proposal for the summer conference, they wrote: "An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves." The conference was attended by at least twenty-nine well-known figures, including the mathematician John Nash — famous for his contributions to game theory and portrayed by Russell Crowe in A Beautiful Mind.

Along with Turing's paper, the Dartmouth College Conference was a foundational moment marking the beginning of AI history. But difficulties were already emerging that anticipated problems the field would face to this day. Many objected to the word "artificial" in the name McCarthy had chosen. Does calling it artificial intelligence not constrain what we mean by intelligence? What makes something artificial? What if the foundations are not artificial at all, but the same as those of human intelligence? What if machines surpass human intelligence? There were already suggestions that the answers to these questions might not be technological but philosophical — because despite machines being, in some ways, more capable than humans, something seemed to be missing.

The Symbolic Approach

The first major approach to AI — one that dominated the first few decades of research — came to be called the symbolic approach. The idea was that intelligence could be modelled symbolically, by imitating or encoding a digital replica of the human mind. If the mind contains a movement area, an emotional area, a calculating area, a visual area, and so on, then one could simply code each of those. Symbolic approaches essentially aimed to create digital maps of the real world. If the world could be represented symbolically, it could be approached logically.

John McCarthy, a proponent of this approach, wrote: "The idea is that an agent can represent knowledge of its world, its goals, and the current situation by sentences in logic, and decide what to do by deducing that a certain action or course of action is appropriate to achieve its goals."

The appeal to computer programmers was obvious. This kind of thinking aligned neatly with binary — the root of computing, in which a transistor is either on or off, one or zero, true or false. If the traffic light is red, then stop. If hungry, then eat. If tired, then sleep. Building a symbolic, virtual, logical picture of the world quickly became the most influential approach.

In his book A Brief History of AI, computer scientist Michael Wooldridge writes that this was because "it makes everything so pure. The whole problem of building an intelligent system is reduced to one of constructing a logical description of what the robot should do."

But a problem quickly emerged. Knowledge turned out to be far too complex to be represented neatly by these simple, binary, if-then rules. One reason is the gradient of uncertainty: if hungry, then eat is not exactly true or false — there are degrees of hunger. But another problem was that calculating what to do from these seemingly simple rules required far more knowledge, and far more computation, than first assumed. The computing power of the period simply could not keep up.

Consider the Towers of Hanoi — a game in which you must move a stack of discs from one pole to another in as few moves as possible, never placing a larger disc on top of a smaller one. With three discs, the game is solvable in seven moves. With five, thirty-one. With ten, 1,023. With twenty discs, over a million moves are required. With sixty-four, the number of moves is so vast that, at one move per second, it would take nearly 600 billion years to complete.

This problem was named combinatorial explosion — the phenomenon by which, as the number of possible options for each action increases, the total number of combinations quickly becomes incomprehensibly large and technologically impossible. The Towers of Hanoi is a simple game. Combinatorial explosion became even more severe with games like chess or Go, and with human problems like driving — infinitely more complex still.

In robotics, a similar approach was being tested and proving even more complicated. At MIT in 1971, Terry Winograd developed a program called SHRDLU that symbolically modelled something he called Blocks World — a virtual environment in which a user could ask the program to manipulate blocks in various ways. In a simple, constrained setting, it worked impressively well. But when researchers at Stanford built a real-life version — a robot called Shakey, equipped with bump sensors and laser rangefinders — the problems multiplied. The room had to be painted in a specific way for the sensors to even function. The technology of the time could not keep pace. By the 1970s and 1980s, combinatorial explosion and the sheer complexity of any real environment had become so severe that the field entered what is now known as the AI winter.

A History of AI: The Impossible Totality of Knowledge

By the 1970s, some researchers were beginning to argue that something crucial was being left out: knowledge. The real world is not made of Towers of Hanoi puzzles or robots and blocks. Knowledge about the world is central.

The result was the expert systems approach, which led to one of the first significant AI breakthroughs. Researchers at Stanford used this method to work with doctors to create a system called MYCIN, designed to diagnose blood diseases. It combined knowledge and logic — if a blood test shows X, then perform Y — and crucially, it could show its reasoning, which was essential if professionals were ever going to trust it. At first, it proved as accurate as human doctors.

A similar system called Dendro used the same approach to analyse the structure of chemicals, drawing on 17,500 rules provided by chemists. Both systems seemed to demonstrate that expert knowledge AI could work.

The AI winter appeared to be over. But once again, developers encountered a serious new problem. The MYCIN database quickly became outdated. In 1983, researcher Edward Feigenbaum wrote: "The knowledge is currently acquired in a very painstaking way... In the decades to come, we must have more automatic means for replacing what is currently a very tedious, time-consuming and expensive procedure. The problem of knowledge acquisition is the key bottleneck problem in artificial intelligence." MYCIN was never widely adopted. It proved expensive, quickly obsolete, legally questionable, and too difficult to roll out broadly.

In the 1980s, the influential computer scientist Douglas Lenat began a project intended to solve this — a system called CYC. Lenat wrote: "No powerful formalism can obviate the need for a lot of knowledge. Most of what we need to know to get by in the real world is too much common sense to be included in reference books." He gave examples: animals live for a single continuous interval of time; nothing can be in two places at once; animals don't like pain. Lenat and his team estimated it would take 200 years of work. They set about laboriously entering half a million taken-for-granted facts: that bread is a food, that Isaac Newton is dead.

But CYC ran into immediate difficulties. It didn't know whether bread was a drink, or that the sky is blue, or whether the sea is wetter than land, or whether one sibling could be taller than another. These simple questions reveal something profound about knowledge: often we don't explicitly know something ourselves, yet when asked, the answer seems laughably obvious. We may never have considered whether bread is a drink. When we do, we just know — not through explicit reasoning, but intuitively, through a web of implicit connections.

Logic struggles with nuance, uncertainty, and probability. Consider the classic example from AI handbooks: Quakers are pacifists. Republicans are not pacifists. Nixon is both a Republican and a Quaker. Is Nixon a pacifist? A computer, given only this information, sees a contradiction and cannot proceed. A human draws on history, context, politics, and the ambiguity inherent in such categories to arrive at a considered, if uncertain, response.

Wooldridge writes: "The main difficulty was what became known as the knowledge elicitation problem. Put simply, this is the problem of extracting knowledge from human experts and encoding it in the form of rules. Human experts often find it hard to articulate the expertise they have. The fact that they're good at something does not mean that they can tell you how they actually do it."

McCarthy's idea that logic was at the centre of intelligence fell out of favour. The logic-centric approach was rather like saying a calculator is intelligent because it can perform calculations — when it knows nothing itself.

Meanwhile, the Australian roboticist Rodney Brooks was arguing that the problem with simulated worlds was that they were too tightly controlled. Real intelligence, he argued, did not evolve in controlled environments. It emerged bottom-up, arising from everyday experience, from the interaction of all parts — not top-down, programmed from a central command point. Evolution, after all, is profoundly bottom-up: it begins with single-celled organisms and, over billions of years of incremental addition, arrives at consciousness.

Brooks worked on what he called embodied intelligence — connected to its environment through sensors, cameras, microphones, arms, and lasers. He asked: why could a machine beat any human at chess but fail to pick up a chess piece as well as a two-year-old child? And not only that — the child moves their hand automatically, without any apparent complex computation. In fact, the brain doesn't seem to have anything like a central command centre; it functions more like a city spread across a wide area than like a single pilot flying a complicated plane.

Brooks and his team built a robot called Cog, with thermal sensors and microphones but no central command point. They called it decentralised intelligence. It was innovative, but never quite worked. Brooks himself admitted that Cog lacked coherence.

In 1996, IBM's chess AI Deep Blue was beaten by Grandmaster Garry Kasparov. Deep Blue was an expert knowledge system, programmed not only to calculate every possible move but to incorporate professional chess knowledge: optimal opening moves, lines of attack, positional concepts. After its defeat, IBM threw more computing power at it. Deep Blue could search through 200 million possible moves per second using 500 processors. It played Kasparov again in 1997, and — a milestone for AI — it won. Kasparov accused IBM of cheating and, to this day, maintains there was foul play. Even with 500 processors and 200 million moves per second, IBM may still have relied on very specific human knowledge about Kasparov himself. This, if perhaps apocryphal, is at least a premonition of things to come. Operation is, in the end, the accumulated wisdom of years of human problem-solving, reflected back.

The Learning Revolution

In 2014, Google announced it was acquiring a small, relatively unknown four-year-old AI lab from the United Kingdom for $650 million. DeepMind had done something that, on the surface, seemed quite simple: it had beaten an old Atari game. But how it had done so was far more interesting. New terms began entering the mainstream — machine learning, deep learning, neural nets.

What the knowledge-based approaches to AI had struggled with was collecting knowledge effectively. MYCIN quickly became obsolete. CYC missed things that most people found obvious. Entering the totality of human knowledge was impossible — and besides, an average human doesn't carry all of that knowledge either. And so researchers began pivoting to a fundamentally different question: if we can't teach machines everything, how can we teach them to learn for themselves?

Machine learning begins with a goal. From that goal, the machine acquires the knowledge it needs through trial and error. Wooldridge writes: "The goal of machine learning is to have programs that can compute the desired output from a given input without being given an explicit recipe for how to do this."

DeepMind had built an AI that could learn to play — and win — not just one Atari game, but many of them, entirely on its own. The premise was simple: the AI was given the controls and a preference for increasing the score, then, through trial and error, it would attempt different actions, build on what worked, and avoid what did not. This is called reinforcement learning. If a series of actions led to losing a point, the AI would register this as likely bad — and vice versa. It would then play the game thousands of times, reinforcing successful patterns.

What was extraordinary was that the AI didn't merely learn the games — it quickly became better than humans, reaching superhuman levels across twenty-nine of forty-nine Atari games. In Breakout, for example, the developers were astonished when the AI independently discovered a technique for getting the ball to the top of the screen, where it would bounce around destroying blocks on its own — without further input. This tactic was described as spontaneous, independent, and creative.

Next, DeepMind turned its attention to Go, commonly regarded as the most difficult strategy board game in the world. AlphaGo was trained on 160,000 top professional games and played over 30 million games against itself before beating world champion Lee Sedol in 2016. Where combinatorial explosion had always made Go intractable — too many possible positions to calculate exhaustively — DeepMind's approach was based on sophisticated probabilistic reasoning. Rather than calculating every possible future move, AlphaGo estimated the probability of winning from any given position. This is, in fact, closer to how human intelligence operates: we scan, contemplate a few moves ahead, reject some, imagine others.

After thirty-seven moves in the match against Sedol, AlphaGo made a move that took everyone by surprise. No human commentator could understand it. Professionals described it as creative, unique, beautiful — and inhuman. The age of machine learning had arrived.

What Are Neural Nets?

In 1991, two scientists wrote that "the neural network revolution has happened." You may have heard the terms neural nets, deep learning, and machine learning. Some have come to believe this revolution may be the most historically consequential our species will undergo.

Recall that the symbolic approach attempted to create a kind of one-to-one map of the world and base artificial intelligence upon it. Machine learning instead learns through trial and error. Today, AI mostly does this using neural networks.

Neural networks are composed of layers of nodes, inspired by the biological neural networks of the brain. In both biological and artificial neural networks, the basic building blocks are neurons, or nodes. These are arranged in layers with connections between them. Each neuron can activate the next. The more neurons that activate, the stronger the activation of the connected neuron — and if the signal is strong enough, it crosses a threshold and fires the next neuron. This process repeats billions of times. In this way, intelligence can make predictions based on past experience.

Think of neural networks — in the brain and artificially — as commonly travelled pathways. The more neurons fire together, and the more successfully they do so, the stronger those connections become and the more likely they are to repeat. Hence the phrase: those that fire together wire together.

How are these used in AI? You need a great deal of data. You can either feed the neural network a large existing dataset — thousands of professional Go or chess games, for instance — or you can have it play games over and over, across many computers simultaneously. Once you have enough data, the next task is to find patterns. If you know a pattern, you can predict what comes next.

ChatGPT and its counterparts are based on large language models — neural networks trained on enormous quantities of text. ChatGPT was trained on around 300 billion words. And if you're asking whose words, and from where — you may already be onto something we will return to shortly.

The cat sat on the ___. If you immediately thought mat, you have some intuitive grasp of how large language models work. In 300 billion words of text, that pattern appears very frequently. ChatGPT predicts that mat is what follows.

But what if I say: the cat sat on the elephant? One of the problems that earlier approaches encountered was that not all knowledge is binary. Neural networks are particularly powerful precisely because they work with probabilities, ambiguity, and uncertainty. The nodes have strengths: while mat is overwhelmingly likely in the typical context, elephant can still appear with a lower probability. If a large language model encounters the phrase heads or tails, the following nodes will split roughly fifty-fifty between the two options.

As researcher Crawford explains, researchers began using statistical models that focus on how often words appear in relation to one another, rather than trying to teach computers rules-based approaches using grammatical principles or linguistic features.

The same logic applies to images. How do you teach a computer that an image of the number nine is a nine — when every example is slightly different: in photographs, on signposts, handwritten, scrawled at strange angles, in varying shades, with imperfections, upside down? Feed a neural net millions of drawings, photographs, and designs featuring the number nine, and it can learn the recurring patterns until it can recognise a nine independently. The challenge is simply that it requires a very large number of examples.

The neural network revolution carries several groundbreaking implications. First, intelligence is not some abstract, transcendent, ethereal thing. What matters is connections — and those connections allow both humans and AI to predict what comes next. Second, machine learning researchers were realising that for this approach to work, they needed enormous quantities of data. Getting chemists and blood-diagnostic experts to visit a lab once a month and laboriously type in their latest findings was not going to suffice. The data problem needed solving at scale.

OpenAI & ChatGPT

There is a story — likely apocryphal — that Google co-founder Larry Page called Elon Musk suspicious because Musk prioritised human life over other forms of life: that he privileged humans over a potential artificial super-intelligence. If AI became greater and more significant than humans, Page reportedly argued, there would be no reason to prioritise or protect humans at all. Perhaps the machines really should take over.

Musk claims that this worried him — especially as Google, after acquiring DeepMind, was at the forefront of AI development. And so, despite being a multi-billion-dollar businessman himself, Musk became concerned that AI was being developed behind the closed doors of multi-billion-dollar corporations.

In 2015, he founded OpenAI, with the stated aim of developing the first general artificial intelligence in a safe, open, and humane way. AI at the time had become very good at performing narrow tasks: Google Translate, social media algorithms, GPS navigation, chatbots, calculators. This was what came to be called narrow AI — also known, rather unflatteringly, as weak AI — and it had been something of a quiet revolution. It was already spreading slowly and pervasively everywhere.

The purpose of OpenAI was to pursue something more ambitious: artificial general intelligence, or AGI — the kind of intelligence depicted in films, capable of crossing between tasks, doing unexpected creative things, and acting broadly as a human does. AI researcher Luke Muehlhauser describes AGI as "the capacity for efficient cross-domain optimisation" — the ability to transfer learning from one domain to another.

With donations from Silicon Valley venture capitalists including Peter Thiel and Sam Altman, OpenAI began as a non-profit, committed to transparency, openness, and — in its founding charter's words — building "value for everyone rather than shareholders." It promised to publish its research and share its patents.

But the team quickly realised they had a serious problem. The best approach — neural nets and deep machine learning — required vast amounts of data, enormous server capacity, and above all, enormous computing power. This was something their main rival, Google, had in abundance.

By 2017, OpenAI decided to restructure as a for-profit company — a capped-profit structure, with a 100-fold limit on returns for investors, overseen by the non-profit board. In a statement, OpenAI said: "We anticipate needing to marshal substantial resources to fulfil our mission, but will always diligently act to minimise conflicts of interest among our employees and stakeholders that could compromise broad benefit."

The decision paid off. In February 2019, OpenAI announced it had a model capable of producing written articles on any subject — articles apparently indistinguishable from human writing. They claimed it was too dangerous to release. At first, it was dismissed as a publicity stunt. But in 2022, they released ChatGPT, a large language model that appeared, at least in part, to pass the Turing test. You could ask it anything; it could write anything; it could do so in different styles; it could pass many exams. By the time it reached ChatGPT-4, it could pass the SAT, the law school bar exam, biology and high school mathematics, and medical licensing examinations — in some cases with flying colours.

ChatGPT attracted one million users in five days and, by the end of 2023, had 180 million — the fastest-growing consumer application in history. In January 2023, Microsoft made a multi-billion-dollar investment in OpenAI, and began embedding ChatGPT into Windows and Bing.

But OpenAI had suspiciously become something more like Closed AI. Some began asking: how had ChatGPT made so much progress using material that was not freely and openly available on the legal internet? A dichotomy was emerging between open and closed, between transparency and opacity, between the many and the few, between democracy and profit. When journalist Karen Hao visited OpenAI, she said there was a misalignment between what the company publicly espoused and how it operated behind closed doors.

The Scramble for Data

For almost all of human history, data — information — has been both a driving force and relatively scarce. The scientific revolution and the Enlightenment accelerated the idea that knowledge could and should be acquired both for its own sake and in order to innovate, advance, and progress.

AI accelerated a much older trend — one that reaches back to the Enlightenment, to the scientific revolution, even to the agricultural revolution. More data means better predictions. If you know patterns, you can make predictions about when those patterns will recur. More data, more patterns, better predictions.

The internet was initially a military project. The US Defense Advanced Research Projects Agency — DARPA — realised that surveillance, reconnaissance, and information were key to winning the Cold War. The appetite for data to predict has only grown since. And the problem has always been how to collect it.

But by the 2010s, with the rise of high-speed internet, camera phones, and social media, vast numbers of people across the globe were, for the first time, voluntarily uploading terabytes of data about themselves. Philosopher Shoshana Zuboff calls the appetite for this kind of data "the right to the future tense" and describes surveillance capitalism as unilaterally claiming human experience as "free raw material for translation into behavioural data."

Before the age of big data, AI researchers had struggled to extract knowledge effectively. IBM scanned its own technical manuals. Universities used government documents and press releases. One researcher in the early 1990s recalled: "Back in those days, you couldn't even find a million words in computer-readable text very easily."

By the 2000s, the idea of consent itself seemed to be changing. In 2007, computer scientist Fei-Fei Li began ImageNet, a project to map the entire world of objects using neural networks and deep learning. By 2009, the researchers had scraped over fourteen million images from the internet and used micro-workers to label them across tens of thousands of categories.

As people began voluntarily uploading their lives online, the data problem was, in effect, solving itself. Today, we generate an estimated 2.5 quintillion bytes of data every day — enough, if printed, to circle the Earth once every four days.

Private companies, the military, and the state are all engaged in data extraction for prediction. The NSA runs a programme called Treasure Map, aiming to map the physical location of everyone on the internet at any given moment. Google Street View engineers have described themselves as "building a mirror of the real world." And as far back as 1985, AI researcher Robert Mercer put it simply: "There's no data like more data."

Stolen Labour

AI presents itself as though it arrived fully formed — already sentient, useful, almost omniscient. It appears as a conjurer, a magician, an illusionist. But this illusion conceals how much labour, how many ideas, how much creativity, passion, and life has been used — and, as we shall see, often stolen — to bring it into being.

First: much of the organising, moderation, labelling, and cleaning of data is outsourced to workers in the developing world. When Jeff Bezos assembled a database of millions of books from catalogues and libraries, the team found the data was messy and unusable. Amazon outsourced the cleaning to temporary workers in India. This proved so effective that, in 2005, Amazon launched a new service: Mechanical Turk — a platform on which businesses can outsource tasks to an army of cheap temporary workers, paid not by the hour but per micro-task.

Amazon claims half a million workers are registered on Mechanical Turk, though the active figure is likely closer to 100,000 to 200,000. Either way, it would rank comfortably among the world's largest employers. Computer scientists often call this human computation, but in their book Ghost Work, Mary Gray and Siddhartha Suri call it what it is. They point out that most automated jobs still quietly require humans to work around the clock. AI researcher Tom Dietterich has said: "We must rely on humans to backfill with that broad knowledge of the world to accomplish most day-to-day tasks."

These tasks are repetitive, underpaid, and often deeply unpleasant. Some workers spend their days reviewing illegal images of child abuse, earning a few cents per image. Others watch colonoscopy videos, circling polyps hundreds of times over. One paper estimates the average hourly wage on Mechanical Turk at just two dollars — lower than the minimum wage in India, let alone in many other countries where this work takes place.

Perhaps worst of all, workers can be — and regularly are — cheated out of their earnings entirely. A 28-year-old from Hyderabad named Riaz had built a small business around Mechanical Turk, employing ten friends and family members. Without warning, all of their accounts were suspended, and he received a brief email from Amazon informing him that funds remaining on his account were forfeited. No explanation was given. No appeal was possible. Two months of wages, for nearly two dozen people, simply vanished. A survey by Gray and Suri found that thirty percent of Mechanical Turk workers report having performed work for which they were never paid.

One Google employee told The Guardian: "It's all smoke and mirrors. Artificial intelligence is not that artificial. It's human beings that are doing the work." Another added: "It's like a white-collar sweatshop. If it's not illegal, it's definitely exploitative."

The irony is not lost. Amazon's Mechanical Turk takes its name from an eighteenth-century chess-playing automaton that was built to impress the Empress of Austria. The machine appeared to play chess on its own. In truth, concealed inside the elaborate cabinet was a cramped human being. The machine intelligence wasn't a machine at all. It was entirely human.

Stolen Libraries: The Mystery of Books2

In 2022, an artist named Lapine used the website Have I Been Trained? to discover whether her work had been used in AI training sets. To her astonishment, a photo of her own face appeared — one taken by her doctor as clinical documentation for a skin condition, and covered by a confidentiality agreement. The doctor had died in 2018, but somehow the highly sensitive images had not only ended up online, but had been scraped as AI training data.

The same dataset, LAION-5B, was used to train the popular AI image generator Stable Diffusion — and has also been found to contain at least a thousand images of child sexual abuse.

There are many black boxes in the world of AI. The term black box has been adopted by developers to refer to algorithmic behaviour that even the developers themselves cannot fully explain. But there is another type of black box that developers do understand — they simply choose not to make it public: how their models are trained, what they are trained on, and what problems and dangers lurk within.

Much of what models like ChatGPT were trained on is publicly available text from the internet, or books long out of copyright. But there is an almost mythical dataset — a shadow library — made up of two sets called Books1 and Books2. OpenAI acknowledges their existence and states they contribute fifteen percent of the training data. But their contents remain a closely guarded secret.

As ChatGPT took off, authors and publishers began to wonder how it could produce detailed summaries, analyses, and stylistic imitations of books that were under copyright — books that could not be read without buying them. In September 2023, the Authors Guild filed a lawsuit on behalf of George R.R. Martin, John Grisham, and seventeen others, claiming that OpenAI had engaged in "systematic theft on a mass scale." Further complaints followed from Jon Krakauer, James Patterson, Stephen King, Zadie Smith, Jonathan Franzen, Margaret Atwood, and thousands of others. In total, more than 8,000 authors signed an open letter to six major AI companies protesting the use of their work.

Sarah Silverman, as lead plaintiff in another lawsuit, presented as Exhibit A the fact that ChatGPT could still summarise in detail the first part of her book The Bedwetter. In another case, author Michael Chabon and others cited OpenAI's own published research, which acknowledged that "of all sources and content types that can be used to train the GPT models, written works, plays and articles contain the best examples of high-quality long-form writing" and "long stretches of contiguous text, which allows the generative model to learn to condition on long-range information."

Developer Shawn Presser had speculated as early as October 2020 that OpenAI's Books2 database might be the entirety of Library Genesis — a pirated shadow library of millions of copyrighted books and journal articles. Presser used a script written by the late programmer and activist Aaron Swartz to download the entire library. He called the resulting dataset Books3 and hosted it on an activist website called The Eye.

Journalist Alex Reisner at The Atlantic later obtained the Books3 dataset and wrote a program to identify the books it contained. He found over 190,000 releases, the vast majority less than twenty years old and therefore under copyright — including titles from Verso, HarperCollins, and Oxford University Press. He concluded: "Pirated books are being used as inputs. The future promised by AI is written with stolen words."

Bloomberg eventually acknowledged using Books3. Meta declined to comment. OpenAI has still not revealed what Books2 contains.

When Gary Marcus and filmmaker Reid Southen published an investigation in January 2024 demonstrating that image-generating AI systems like Midjourney and DALL-E could easily reproduce clearly copyrighted material from films including The Matrix, The Avengers, Star Wars, and The Hunger Games, Southen was banned from Midjourney. He opened two new accounts; both were also banned. Marcus and Southen concluded: "We believe that the potential for litigation may be vast, and that the foundations of the entire enterprise may be built on ethically shaky ground."

Midjourney CEO David Holz, when asked directly whether consent had been sought for training materials, replied candidly: "No, there isn't really a way to get 100 million images and know where they're coming from."

The BBC, CNN, and Reuters have all attempted to block OpenAI's crawlers from accessing their content. Elon Musk's Grok AI has produced error messages from OpenAI — suggesting, rather hilariously, that its own code may have been taken from OpenAI. And in March 2023, the Writers Guild of America stated plainly: "AI software does not create anything. It generates a regurgitation of what it's fed. Plagiarism is a feature of the AI process." Breaking Bad creator Vince Gilligan called it "a giant plagiarism machine in its current form."

Copyright and the Future of Creativity

There has always been a wide-ranging philosophical debate about what knowledge is, how it is formed, where it comes from, and whose — if anyone's — it really is. Does it come from God? Is it a spark of individual genius, creating something from nothing? Is it the product of institutions, of collective effort, of lone visionaries?

Copyright, as an idea, is historically unusual — it emerged loosely from Britain in the early eighteenth century. The purpose of protecting original work for a limited period was twofold: first, to compensate the creator; and second, to incentivise innovation more broadly. UK law refers to copyright being applied to "the sweat of the brow of skill and labour." US law requires some minimal degree of creativity. It does not protect ideas — only how those ideas are expressed.

OpenAI's defence has been that training on copyrighted material constitutes fair use — that ChatGPT does not quote verbatim but substantially transforms the material into something new. They wrote: "Training AI models using publicly available internet materials is fair use, supported by long-standing and widely accepted precedents. We view this principle as fair to creators, necessary for innovators, and critical for US competitiveness."

The defence is at least plausible. The key question, however, is whether the original creators are rewarded — and whether the model incentivises further innovation. If large language models dominate the internet, neither citing authors nor compensating those they draw from, then we lose both societal benefit and a strong incentive to produce the original work on which these systems depend.

The AI plagiarism website Copy Leak analysed ChatGPT 3.5 and estimated that sixty percent of it contained plagiarism — forty-five percent identical text. By some estimates, within a few years, ninety percent of the internet could be AI-generated. Under these conditions, what would become of journalism, art, and science? No one rewarded. No one read. No wages, no portfolio, no purpose. Just bots endlessly rewording everything, forever.

Google Search worked so well because it linked to websites, blogs, experts, artists, and authors — so that you could find them, read them, watch them. In Bard, or Claude, or ChatGPT, that does not happen. All our words, images, and music are taken, scraped, analysed, repackaged, and sold back to us as someone else's.

Yes, much of the attention falls on well-known, commercially successful artists — Silverman, Grisham, The New York Times. But most of the billions of words and images these models were trained on came from unknown, underpaid, underappreciated people. As artist Nikki Bones pointed out: "Everyone knows what Mario looks like, but nobody would recognise Mike Finkelstein's wildlife photography. When you ask for a sharp, beautiful photo of an otter leaping out of the water, you probably don't realise the output is essentially a real photo that Mike stayed out in the rain for three weeks to take."

What is to be done? Training data needs to be paid for. Artwork must be licensed. Authors must be referenced, cited, and credited. Regulation is necessary — these models are, in important ways, not much different from Napster, which was shut down. Transparency about training datasets seems a minimum requirement. And some form of democratic oversight appears essential. None of this will happen through commercial incentive alone.

The End of Work and a Different Kind of Apocalypse

In March 2022, researchers in Switzerland found that an AI model designed to study chemicals could suggest how to make 40,000 toxic molecules — including nerve agents like VX — in under six hours. Separately, Professor Andrew White, working as part of OpenAI's red team, found that GPT-4 could recommend how to synthesise dangerous chemicals, identify suppliers, and — in a test he actually carried out — order the necessary ingredients automatically to his house.

The problem is that neural nets and machine learning at superhuman speeds discover patterns for doing things that even the developers themselves cannot understand. There are so many possible inputs, so many ways to prompt the model, so many data pathways, that it is impossible to predict all of the possibilities. AI performance, by definition, means being a step ahead — which means doing things in ways that we cannot understand, or that we can only understand after the fact.

Philosopher Nick Bostrom has given an influential example: imagine asking a powerful AI to make as many paperclips as possible. The AI achieves this with increasing efficiency until you tell it to stop. But stopping contradicts its original instruction. It calculates that the biggest obstacle to the goal is human interference, and so it eliminates that obstacle — hacking into nuclear systems, poisoning water supplies, eventually converting all available matter into paperclips. The terrifying point is that machine intelligence is so fast it will always be a step ahead. It will pursue goals in ways we cannot comprehend.

Roboticist Rodney Brooks offers a counterargument: a Boeing 747 could not be built by accident. Intelligence requires careful planning, complicated cooperation, and intentional construction. AGI will not spontaneously appear. But this argument also misses something important: passenger jets may not be built by accident, but they certainly crash by accident. And as technology improves, the chance of misuse, malpractice, and catastrophic accident grows alongside it.

When DeepMind's AI beat Breakout, it did so by discovering something nobody expected — going behind, doing something that could not be anticipated, surprising us from a direction we had not considered. That is the metaphor. Not the Terminator. The Atari game.

In 2013, Carl Frey and Michael Osborne at Oxford published a report predicting that nearly half of all jobs could be automated. High-risk professions included telemarketers, insurance agents, data-entry clerks, cashiers. Therapists, surgeons, and teachers were considered safest, on the assumption that creative jobs, jobs requiring dexterity, and jobs demanding genuine human connection were most resilient.

But models like DALL-E and Midjourney have already become remarkably creative, and will only improve. Autonomous trucks and household robots remain stubbornly difficult to perfect. Ironically, the creatives appear more immediately threatened than the drivers.

There is also a deeper irony. Truckers, for instance, work long hours across vast distances, their vehicles collecting data with sensors and cameras — data that is then used to train the autonomous vehicles designed to replace them. The people most displaced by AI are, in many cases, the people whose labour trained it. Only those with the capital — the trucks, the servers, the machines — will reap the rewards. The rest will slowly be rendered redundant.

The question is not only what AI can do. It is who it can do it for.

The End of Humanity

After a shooting at the University of Michigan, administrators sent students a letter of consolation about community, mutual respect, and togetherness. The bottom of the email revealed it had been written by ChatGPT. One student said: "There is a sick and twisted irony to making a computer write your message about community and togetherness because you can't be bothered to reflect on it yourself."

The philosopher Michel Foucault famously said that the concept of "man" — anthropomorphic, human-centred, with an individualistic psychology — was a historical construct and a very modern one. One that changes, shifts, and morphs over time, and that one day, he said, "man would be erased like a face drawn in the sand at the edge of the sea."

It was once believed that humanity had a soul. René Descartes was inspired by the scientific revolution taking place around him: the world increasingly described mechanistically, like clockwork. Atoms striking atoms. Gravity pulling things earthward. Everything explainable by cause and effect. This troubled Descartes, because he believed there was something special about the mind — it was not pushed and pulled mechanically; it was free. And so Descartes divided the universe into two: res extensa, the extended material world, and res cogitans, the abstract, thinking, free substance of the human mind.

This duality defined the next few centuries. But it has come increasingly under attack. Today, many say the mind is simply a computer — inputs and outputs, drives and appetites, causes and effects, synapses and neurons like everything else. The AI revolution is showing us, if it has not already done so, that intelligence is nothing soulful, rare, or special at all. That it is just stuff — algorithmic, pattern-detecting, data-driven.

If machines outperform us at everything, then by definition, they do things we can no longer understand. Even the developers cannot explain why neural nets follow the paths they choose. AlphaGo makes moves that humans cannot decipher. ChatGPT might one day produce a personalised guide to an emotional problem you had not even realised you had. We might not be made by gods. But we could be making them.

The transhumanist movement predicts that to survive, we will need to merge with machines: neural implants, bionic improvements, uploading the mind to digital form. We already augment our biology with technology — hearing aids, glasses, prosthetics. One of transhumanism's founding figures, Ray Kurzweil, wrote: "We don't need real bodies." In 2012, he became Director of Engineering at Google.

But transhumanism rests on a kind of optimism that some essential part of us will be able to keep pace with the ever-accelerating speed of technological adoption. The history of AI, as we have seen, has been one of absorbing, appropriating, and absorbing still more of all human knowledge. Taking more and more and more. Until what is left?

We deny this. We are a species in denial. We say: Sure, it can win at chess, but Go is the truly human game. Then: Sure, it can win at Go, but it cannot truly be creative. Then: Sure, it can be creative, but it cannot truly understand emotions. But slowly, AI catches up. And then it surpasses.

The world is becoming re-enchanted. Understanding moves away from us — becomes superhuman. Necessarily, it leaves us behind. In the long arc of history, the age of human understanding has been a blip — a tiny island in a deep, unknown sea. We will be surrounded by enchanted machines, wearables, nanotechnology, objects we can no longer comprehend.

Or a New Age of Artificial Humanity

We live in strangely contradictory times. We are told that anything is technologically possible — that we will transcend our weak, fleshy limitations and enter the age of the superhuman. And yet we seem unable to provide a basic standard of living, political stability, or a reasonable set of life expectations for the majority of people on Earth. We cannot seem to do anything about inequality, climate change, or global war.

If AI can do all of these incredible things better than all of us, what kind of world might we dare to imagine?

What we need is nothing less than a new model of humanity.

The philosophical and scientific critics of Descartes' dualism offer a better starting point. The Dutch philosopher Baruch Spinoza argued against the idea that thought and matter are separate. He saw all of the world's phenomena — nature, us, animals, forces, mathematics, bodies and thought — as part of one scientific universe. Everything is connected. Knowledge is spread throughout, embedded in everything. "The highest activity a human being can attain," Spinoza wrote, "is learning for understanding — because to understand is to be free."

This model aligns far better with what AI researchers have discovered. Neural nets, deep learning, the neurons in the brain — they are all intelligent because they take data about the world and detect patterns in it. Intelligence is not in here. It is out there. It is everywhere. What matters are connections.

True artificial intelligence will connect to, build upon, and work in relation to other machines, other people, other groups, other resources. It will work across logistics, shipping, manufacturing, research. But access to this intelligence — embodied, connected, material — will determine who benefits and who does not. Intelligence makes little sense without the ability to reach out, shape, and use it.

I believe we are entering an age of storytelling. If AI can do the science better than any of us, if it can write the best article on international relations and build machines and cook and work in our place, what will remain of us? Our stories. Our lives. The people who can tell the best stories about what we should be doing with these tools, what shape our future should take, which ethical questions are worth pursuing, which artistic ones — stories about family, emotion, journey, and aspiration, about local life, friendship, and all the things that still make us human.

Maybe the AI age will be more about meaning. Meaning is about being compelling, passionate, making a case, articulating — using data and algorithms and inventions to tell a good story about what we should all be doing with it all. The greatest innovators and marketers have always known: it is not the technology that matters. It is the story that comes with it.

I would like to think — perhaps naively — that I will not simply be replaced by ChatGPT. Because while it might one day write a better script about the history of AI — more accurate, better sourced — it will not be able to do this part as well. Because I hope that you also want to know something about me: my values, my emotions, my passions and idiosyncrasies, my mistakes, my style and perspective. Who I am, and what I believe, from my small corner of the world — so that you can agree or disagree with my idea of what it means to be human.

Getting to the Future

How do we get to a better future — one in which everyone benefits from AI? I think we need to focus on two things primarily: cultural bias and competitive bias. And then, beyond both of those, we need to think carefully about the wider political questions.

Bias is part of being human. We have positions, perspectives, cultures, lenses through which we see things. AI models are trained through reinforcement learning — gently nudging the AI in particular directions. But there is often no single correct answer to an ethical question. David Hume famously argued that you cannot derive an ought from an is — you cannot determine what the world should look like from what it currently is. The data is biased by history. And it is being further shaped, nudged, and corrected by a very specific group of people in a very specific part of the world at a very specific moment in history.

When Google updated Bard in February 2024, it was widely mocked for depicting Black soldiers in 1940s Germany and people of colour as the American Founding Fathers. But this also illustrates the genuine complexity AI programmers face. If you ask an image generator to depict office workers, should the result be demographically accurate for the country in question? Should it reflect who actually holds those jobs? Should it aim for inclusivity? Should it differ between Bangladesh, Japan, and the United States? If it is based on historical stock photography of office workers, it will skew white. Should this be corrected for? These are not simple questions.

The second bias is toward competitive advantage. The logic running through the entire industry is: if we don't do this, our competitors will. If Instagram makes its algorithm less addictive, TikTok will outperform it. Safety testing is slow, and the deadlines are tight. If wealthy, free-market tech entrepreneurs — not exactly anti-capitalist firebrands — are asking governments to step in and regulate, then that ought to tell us something.

We already regulate medicine, law, clinical trials, pharmaceuticals, biological and chemical weapons, food, air travel, vehicles, space travel, and electrical engineering. Regulation of AI is not a radical idea. Careful regulation should aim for the maximum benefit for all with the minimum interference.

Transparency will be central. There must be some mechanism by which auditors, safety testers, regulatory bodies, and the public can, in varying ways and degrees, look under the hood of these models. Data transparency. Algorithmic transparency. References, sources, and credits — so that people are in some way compensated for their work.

As Wooldridge writes: "Transparency means that the data the system uses about us should be available to us, and the algorithms used within it should be made clear to us too." This will only come through regulation — through regulatory bodies with qualified experts who are democratically accountable.

Suleyman writes: "As an equal partner in the creation of the coming wave, governments stand a better chance of steering it towards the overall public interest." There is a strange misconception that regulation inhibits innovation. It does not. Innovation always happens in a context, shaped by many forces. Recent advances in green technology, batteries, and electric vehicles would not have occurred without regulatory changes, and might have happened sooner with different incentives. The internet itself — along with many other scientific, military, and space advances — was not primarily a product of private innovation but of the entrepreneurial state.

I always return to openness, transparency, accountability, and democratic oversight. Because it takes only one mad king, one greedy dictator, one foolish leader, to nudge the levers they hover over toward chaos, rot, and tyranny.

AI implicates everything, everyone, everywhere. And so, more than anything, it is about the issues we already face. The knowledge produced by artificial intelligence is already political, already economic, already implicated in global affairs — in workers and wages, healthcare and pensions, culture and war. And it is those concrete, material realities that matter.

I would wager that, like every total technological transformation before it — the printing press, the Industrial Revolution — this one will also require radical political change. The printing press gave us the Reformation. The Industrial Revolution gave us liberalism and capitalism. What our political world will look like under AI is, ultimately, up to us.

We are entering a period of mass disruption. We need to democratically determine how AI can address our most pressing problems rather than exacerbate them. If we do not, we will make ourselves slowly obsolete — forgotten. AI, built on everything we knew, everything we discovered and created and wrote and painted and photographed and filmed, will surpass us. And we will drift, unmoured, into a future that has left us behind.

We need to ensure that each of us is tethered to that incomprehensibly different future — connected to it, taught by it, in some measure in control of it, compensated by it. It must be in service of all of us.

I am excited and daunted in equal measure. But my sense is that if we fail to do this, rather than being wiped out, we will simply be left stranded in the wake of a colossal, super-intelligent juggernaut we no longer understand — forgotten, left behind, like all the human species we once surpassed, who then went extinct. All of us, fleshy and old-fashioned, bobbing in the water, watching the future disappear into a dark, deep, boundless sea.

‍