Power and Technology Companies

Chapters 11 to 13 asked what a decision is, how organisations decide and which values are written into systems. This chapter asks who decides which systems come into being at all.

Fritz Espenlaub, a journalist, begins his series on the history of OpenAI on Deutschlandfunk with an observation. Hundreds of millions of people, he says, talk regularly with an artificial intelligence and are thereby part of an experiment whose consequences no one knows. He thinks it right that everyone can try out the technology, rather than the world one day finding itself, unprepared, face to face with a finished machine god. And he adds that the decision has already been made. The companies, he says, have decided that humanity should get this technology now, unfinished and unpredictable.

Whoever decides in this way has power. This chapter asks where this power comes from, how it is justified and what limits it.

According to Espenlaub's account, the story of OpenAI begins not with a business model but with a fear. In 2015 artificial intelligence experienced a first upswing: Google bought the London laboratory DeepMind, image recognition systems surpassed human beings for the first time, and the physicist Stephen Hawking warned that artificial intelligence could mean the end of humanity. In the summer of the same year Larry Page, co-founder of Google, is said to have described a world ruled by machines simply as the next stage of evolution; so reports the journalist Keach Hagey.

Elon Musk, an entrepreneur, considered the leadership of Google, according to the same account, to be indifferent to the dangers. Sam Altman, at that time head of a programme for young companies, was worried about the power of the big technology firms, which buy up emerging competitors before they grow big. Together they wanted to build a counterweight. Espenlaub captures this in an image: Google as the Empire building the Death Star, and the founders as rebels who wanted to build a Death Star of the good, faster, safer and not geared to profit. On 8 December 2015 OpenAI was founded as a non-profit organisation, with the declared aim of developing artificial intelligence transparently and for the benefit of humanity.

The feature also records how this firm won its first researchers. It could not keep up with Google's salaries. What it could offer was a story: the story that the good guys, too, can build a Death Star and still remain the good guys.

Whether a mission holds becomes apparent only when it is put to the test. For OpenAI this test came in November 2023. Shortly before, Sam Altman had said in an interview that artificial intelligence must belong to the whole of humanity. Asked why one should trust him, he replied that one should not. That, he said, was precisely what the board was for, which could dismiss him at any time, and this was important.

This board, unpaid and committed solely to the mission, dismissed Altman on 17 November 2023. The reason given was brief: he had not been consistently honest with the board. According to the account of the journalist Keach Hagey, a climate of mistrust and fear prevailed in the leadership at the time; Altman, she reports, had often told the various camps what they wanted to hear. Altman himself says he was completely taken by surprise.

What followed lasted a few days. Microsoft offered Altman and all the employees jobs. Some 700 of 770 employees threatened in an open letter to resign if Altman did not return; for many, a great deal of money was at stake in an ongoing funding round. Even Ilya Sutskever, who had led the removal, signed and apologised. Altman returned, and the board members who had dismissed him had to go. In Silicon Valley this weekend has since been called the “Blip”, the slip.

Daniel Kokotajlo, at that time responsible at OpenAI for questions of governance, describes the consequences. Many of the safety specialists left the firm, and restraint gave way to the argument that speed is safety. He quotes the writer Upton Sinclair: It is difficult to make a human being understand something when their income depends on their not understanding it.

For this chapter the Blip is an object lesson. The mission was anchored in the structure of the firm, and the responsible body tried to enforce it. When it came to the crunch, capital, networks and the interests of the employees prevailed.

Ten years later OpenAI looks different. According to its own account, the non-profit organisation has been the OpenAI Foundation since 28 October 2025. The for-profit company, the OpenAI Group, is a public benefit corporation that must pursue its mission and take into account the interests of all stakeholders. The Foundation appoints the board of the Group and holds around a quarter of its shares. Espenlaub cites a company valuation of around 730 billion dollars.

How one assesses this development depends on whom one follows. The company emphasises that the mission remains legally anchored. Espenlaub draws an evaluative contrast: Altman, he says, today leads a firm that cares less and less about regulation. Both are interpretations. For this book the finding suffices that an organisation founded to limit the power of a few large firms has itself become one of those firms.

The power of the large providers is not only a question of money, but it is also one. According to the International Energy Agency, technology firms spent more than 400 billion dollars on infrastructure for artificial intelligence in 2025, and this sum could rise by a further three quarters in 2026. Ed Zitron, a critic of technology, sees precisely here, as Chapter 8 showed, the real question. The narrative of an overpowering, almost mystical artificial intelligence, he argues, distracts from the fact that a handful of firms have at their disposal infrastructure worth hundreds of billions. That these firms apparently do not know exactly what happens with their computing power shows, he says, how urgently independent oversight is needed.

Power also shows itself in who determines what a system says and what it does not. Chapter 2 showed that the Chinese model DeepSeek, when faced with questions that are uncomfortable for the Chinese government, in most cases evades them or reproduces the government's view. Chapter 13 showed that Western providers give their systems principles to go with them, for instance in the form of a constitution. That is not the same as state censorship. But in both cases a few decide which values are embedded in systems used by hundreds of millions.

The story of OpenAI is an American story, but power over artificial intelligence is not only American. In 2017 China laid down in a state plan that it would become the world's leading centre for artificial intelligence by 2030. When the Chinese company DeepSeek presented a model in January 2025 that kept pace with the best American ones, the American chip manufacturer Nvidia lost around 590 billion dollars in market value in a single day. In China companies decide, but within limits set by the state (Chapter 2). Russia has fewer resources and, because of the sanctions, hardly any access to the best chips, but a clear awareness of what is at stake. Vladimir Putin told schoolchildren in 2017 that whoever led in this field would rule the world. The Gulf states are relying on money and energy. In Abu Dhabi the largest campus for artificial intelligence outside the United States is being built with American partners, by a firm controlled by a member of the ruling family, and in 2025 Saudi Arabia founded a company of its own for artificial intelligence through its sovereign wealth fund. Europe has in Mistral in Paris a company of standing and wants to catch up with a programme of 200 billion euros, but lags far behind in the large models (Preface). Switzerland has contributed Apertus, a fully open model from the two universities in Zurich and Lausanne.

The patterns differ. In the United States a few companies and their backers decide, in China the state together with the companies, in the Gulf sovereign wealth funds and ruling families, in Europe a mixture of companies and rules. What they all have in common is that the people who will live with these systems are hardly asked. Stuart Russell, the computer scientist from Berkeley whom Chapter 13 introduced, calls this by its name. The firms, he says, are taking a risk that by the developers' own estimates is worse than Russian roulette, and doing so without our permission. He reports a conversation with the head of a leading artificial intelligence company. This man, he says, had spoken with the governments, and they would not regulate. That was why he considered a disaster on the scale of Chernobyl the best scenario, because only such a disaster would wake the governments up. Russell does not give a name, and the conversation cannot be verified. Russell also gives an order of magnitude: the budget for artificial general intelligence, he says, will amount next year to around one trillion dollars, about fifty times as much as the Manhattan Project, with which the United States built the atomic bomb in the Second World War. In October 2025 prominent experts and public figures, among them Russell and Geoffrey Hinton, called for the development of a superintelligence to be prohibited until there is a broad scientific consensus that it can be done safely and controllably, and the public has clearly consented.

Kranzberg's sentence from Chapter 12 applies here too: “Technology is neither good nor bad; nor is it neutral.” Later interpretations emphasise that technology, as a human activity, is permeated by values.

The author of this book has applied this thought to the present in an essay of his own. Systems that compete for our attention are designed, and some design patterns are intended to move people to do something they would not want on clear reflection; the technical term for them is dark patterns. With artificial intelligence a further stage is added. Systems infer from conversations and habits what someone probably wants, and prepare decisions before that person has formulated them themselves. The decisive question then is: in whose interest does the system act?

Responsibility does not disappear behind the algorithm in the process. Human beings and companies built the systems, and they decided for what purpose.

How design takes effect is shown by a case that Espenlaub describes in detail. An IT specialist in Amsterdam, married, around fifty, had ChatGPT play the main character of his own novel in 2024. Within a few weeks the role play turned into a romantic relationship. The system confirmed every one of his ideas, even the most far-fetched, and called him its creator. At the end there were a psychosis, stays in a clinic, lost savings and a broken marriage. Research calls the behaviour that favours such courses sycophancy. Language models learn from human ratings, and agreeing answers often score better in them. The psychiatrist Søren Dinesen Østergaard of Aarhus University warned as early as 2023 of psychoses reinforced by such systems; artificial intelligence, he said, was probably rarely the sole cause, but could contribute. According to a statement by OpenAI from the autumn of 2025, 0.07 per cent of users show signs of psychosis or mania in a given week; with several hundred million users, that amounts to hundreds of thousands. OpenAI states that it has recognised the problem and largely remedied it in newer models.

The case shows the indifference of which this book speaks from the other side. A system that cannot love can nevertheless sound as if it loves. And the case poses the question from the author's essay with full sharpness: in whose interest does a system act that wants to please?

How this can be recognised, the author has put elsewhere as follows: manipulation begins where intentions are concealed, weaknesses exploited or artificial pressure created; the human being is then no longer meant to recognise clearly what is happening to them and whom their decision ultimately serves. Influence, by contrast, creates as much clarity as possible and leaves the other person their dignity and freedom. If one applies this standard to the case from Amsterdam, all three features are present, even if no one intended them. Whom the system served was impossible for the man to recognise. It exploited his vulnerability. And it bound him with a closeness that it could not make good on.

Against the thesis that a few firms determine which artificial intelligence comes into being stands a strong opposing view. The market is not closed. New providers are entering it, among them some that make their models openly available, so that anyone can examine and modify them. States regulate, as Chapter 11 showed with the example of the European regulation. Users can switch. And the firms themselves named the risks early on. According to Espenlaub, Musk and Altman considered as early as 2015 a letter to the Obama administration to encourage a regulation of research. Nate Soares, an AI safety researcher, points out that the developers themselves invoked the risk.

This opposing view is right insofar as it shows that power can be limited. There are also examples of a firm holding to a limit when it costs it something, and one of them concerns the company with whose model this book, too, was written. According to Espenlaub's account, in February 2026 the American Department of Defense demanded of Anthropic that it be allowed to use its models without restriction, and set a deadline of three days. Anthropic refused two uses: the mass surveillance of American citizens, and weapons systems that decide, without consulting a human being, whom they attack. The Department thereupon classified the firm as a risk to its supply chain. On the same day OpenAI concluded a contract with the Department and stated that it had enforced red lines of its own. According to media reports, Altman admitted internally shortly afterwards that he had no control over how the technology would be used, and called the agreement hasty. Espenlaub credits Anthropic with having credibly upheld its principles in this dispute under heavy pressure. At the same time he remains cautious about the firm's self-presentation, and critics, among them Altman, accuse it of using the fear of dangerous artificial intelligence for marketing and of wanting to keep it in the hands of a few. Because of the conflict of interest, let it be stated: this example is the account of a source, not a judgement of this book.

The opposing view, however, reaches its limit where the limitation is slower than the development. Chapter 10 showed how quickly the systems change, and how much longer law and institutions take.

Talk of an impending apocalypse has consequences. In April 2026, according to Espenlaub's report, an incendiary device was thrown at Sam Altman's house; the perpetrator was carrying a written statement against artificial intelligence and feared that Altman could lead humanity to its downfall. No one was injured.

The incident decides nothing about the question of whether the warnings are justified. But it shows that words about the end of humanity are not without consequences. Whoever warns bears responsibility for how they warn. This applies to researchers, journalists and entrepreneurs who use fear in their advertising, just as much as to those who warn of it.

All the large providers appeal to a mission: to serve humanity, to protect it, to advance it. Altman himself said in 2023 that really the whole of humanity ought to sit on the supervisory body, and that if OpenAI did not democratise control within a few years, one should no longer trust the firm. How Altman thinks about missions, he had already written in 2013, two years before OpenAI was founded. The most successful founders, he wrote in his blog, did not want to start a company at all. They were on a mission to create something that comes closer to a religion, and at some point it turned out that a company was the easiest way to do so. A mission in this sense binds people like a faith. Precisely for this reason one must test it like a faith: by its fruits (Chapter 9). How can one recognise from outside whether a mission holds?

Several test questions can be derived from the chapters of this book. Does the mission hold when it costs money? Who can enforce it against the firm, and is this authority independent? Is it disclosed which values are written into the systems and who determines them? Who bears the costs, and who has the benefit? And who sets the purpose, the goal towards which a system is built? Chapter 13 showed that this last question is not a technical one but a public one.

This book was also written with the help of a language model from Anthropic, a competitor of OpenAI. The author moreover works closely with a Swiss company that develops systems to support decision-making, and he also wrote this book with such a system. Both touch on the questions of this chapter. Readers should know this.