Miguel Ángel Presno Linera

In May 2025, Karen Hao published the book Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI (Penguin Press). The book was translated into Spanish as El imperio de la IA. Sam Altman y su carrera por dominar el mundo (Península, 2025, translated by Jorge Paredes). Karen Hao is a well-known journalist and mechanical engineer who specialises in the social, economic and political impact of artificial intelligence (hereafter, AI). She has worked as a technology journalist at The Wall Street Journal and The Atlantic, and as an editor at MIT Technology Review.

Against this backdrop, this review examines some of the main contributions of Hao’s remarkable book in its Spanish edition. Although the work offers a detailed account of the origins and extraordinary growth of OpenAI, and particularly of the role played by its chief executive, Sam Altman, it goes much further, functioning as a genuine study of the development of generative artificial intelligence (GAI). The discussion focuses more on the latter than on Altman, although the two are not easily separable.

In its opening pages, the book goes back to the summer of 2015, specifically to a private dinner organised by Sam Altman and attended, among others, by Elon Musk, Dario Amodei, Ilya Sutskever and Greg Brockman, ostensibly to “discuss the future of AI and humanity”. A few months later, several of those present at the dinner founded OpenAI as a non-profit organisation, choosing a name for the new entity that reflected “the spirit of the mission they shared” (Hao, 54).

During these years, Altman in particular benefited from the support of his main mentors: Paul Graham and Peter Thiel, the former a promoter of emerging technology companies, and the latter one of the principal backers of PayPal, Palantir and Facebook, as well as co-founder of the first two.

OpenAI’s leadership team quickly realised that, in order to achieve GAI, they would need vast numbers of highly expensive chips—graphics processing units (GPUs)—sold almost exclusively by Nvidia and requiring enormous amounts of energy. They clearly needed substantial funding, which triggered an internal debate about transforming the organisation into a company and, in parallel, a power struggle between Altman and Musk over its leadership. Although no change in OpenAI’s structure took place at that point—in 2018—Musk did step down from the co-chair position.

Shortly afterwards, in an effort to ease its financial pressures, OpenAI began approaching Microsoft. This culminated in April 2019, when the GPT-2 language model presented by OpenAI’s researchers convinced Bill Gates of the benefits of an agreement with the organisation, which they had failed to achieve when they had previously shown him a robotic hand. As a result, what had not initially been OpenAI’s flagship product moved to centre stage, securing, through Microsoft, the funding required to continue research into GAI.

In August 2019, Hao visited OpenAI’s offices to prepare a report based on a series of interviews and access to internal documents. The piece was published in February 2020 and, among other things, questioned the organisation’s lack of transparency and the discrepancy between what was said publicly and what took place internally, which resulted in her being denied access to the organisation in the following years.

The book revisits 1956 (Hao, 138), tracing the well-known origins of the term AI and the context in which it emerged, emphasising the significance of a label that carried clear marketing value and reinforced the anthropomorphisation of technology, a tendency that continues to this day and was strongly promoted by Hollywood cinema. In Hao’s words (139), speaking of machines as if they learn, read or write creates the impression that they possess greater capabilities than they actually do and helps the companies that create them evade responsibility.

Two lines of AI research became established in the following years: the symbolic approach (Minsky and others), according to which intelligence derives from knowledge favouring expert systems, and the connectionist approach (Rosenblatt and others), which linked knowledge to learning and AI development to machine-learning systems. Until the 1990s, expert systems remained the dominant option, but gradually the connectionists (Hinton, Nobel Prize in Physics 2024 and his collaborators) came to prevail, thanks to the creation and development of increasingly deep neural networks, that is, statistical calculators that identify patterns from past data and apply them to new data (Hao, 149).

Corporate investment in AI surged in the second decade of the twenty-first century, a trend that has intensified dramatically over the past five years. This shift drew research talent towards neural networks and, ultimately, moved researchers from universities into companies, consolidating deep-learning models without eliminating their flaws, since these models absorb and amplify even the smallest imbalances present in the vast quantities of data on which they are trained (Hao, 164).

OpenAI’s success with language models began to take shape with the aforementioned GPT-2, followed by GPT-3 in June 2020. By then, as Hao recalls (235 ff.), studies such as those by Emma Strubell, Ananya Ganesh and Andrew McCallum had already warned about the growing environmental impact of the data centres used to train these models. Other research, including work by Joy Buolamwini and Timnit Gebru, highlighted failures in facial-recognition systems, particularly in relation to dark-skinned women.

In 2021, Gebru herself, together with Emily Bender, Angelina McMillan-Major and Meg Mitchell, published the well-known article On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, addressing the ethics of AI research, a publication that ultimately led to the dismissal of Gebru and Mitchell from Google. Since then, OpenAI, Google and Anthropic, among other organisations in the sector, have restricted access to significant technical details of their models, effectively preventing external audits.

To avoid conflicts arising from the content generated by these models, OpenAI implemented automated filters. These, in turn, relied on the prior work of people who reviewed and categorised hundreds of thousands of examples of content that the models were intended not to produce. To carry out this demanding task, the company turned to the practice of outsourcing it to countries where labour was as cheap as possible (Kenya, Venezuela, Colombia…), thereby creating what Hao describes as “disaster capitalism” (275–322).

OpenAI reached its peak in October 2022 with the launch of ChatGPT. Within two months, it had reached one hundred million users, becoming the fastest-growing application in history up to that point. This catapulted OpenAI into global prominence, surprising even its own leadership and its Microsoft partners, and ultimately helped Altman and other executives push forward the commercialisation of the company and the abandonment of its non-profit status.

As a result, in February 2023, a paid and improved version of ChatGPT was released. This, in turn, drove the use of more powerful supercomputers and increased the need to raise further funding to build the so-called “megacampuses” (Hao, 393), vast data centres that consume enormous amounts of energy and water and are often relocated to countries in the Global South.

OpenAI’s rise greatly increased Altman’s social and political visibility. In May 2023, he appeared before the US Senate to advocate regulatory leniency and consolidate his position by invoking the risks that AI development in China posed to the United States. His statements later found resonance in Congress, to the point that several lawmakers acknowledged consulting Altman and other executives about the scope that any future regulation of AI systems should have. Rather than the familiar idea of regulatory capture, it almost seemed as though the regulator had yielded from the very beginning.

The success achieved did not prevent Altman’s leadership from being questioned among OpenAI’s executives, due to his decision-making style, the withholding of information, his resistance to accountability, and the emergence of allegations made by Annie Altman, his sister, according to which he had sexually abused her when she was a child.

The final chapters of Hao’s book focus on Altman’s struggle to remain at the head of OpenAI, something he ultimately achieved, allowing him to carry out what Hao describes as a “reckoning” throughout 2024 (Hao, 533–563). Altman began 2025 by calling for more capital, arguing that “we now know with certainty how to build AGI” (Hao, 575).

Hao ends the book by advocating, first, for a redistribution of power over AI, which would require funding to conduct research “outside the existing empire”; second, for greater transparency and external oversight of these products, which would in turn contribute to such a redistribution of power; and third, although she does not use this expression, for “digital literacy” to enable a better understanding of how AI works, its strengths and its limitations.

She closes with a quotation from Joseph Weizenbaum, inventor of the first chatbot, ELIZA, in the 1960s: “Once a program is unmasked and its inner workings are explained in sufficiently simple terms to make them understandable, its magic crumbles.”

Postscript: I briefly summarise here the conclusions of a recent paper—The European Artificial Intelligence Regulation: Market Guarantee or Defence of Rights?—in which I express my scepticism regarding the merits of the recent European Artificial Intelligence Regulation (AIR). In my view, the European Union has chosen to prioritise both the internal and the external market at the expense of weaker protection of health, safety and fundamental rights. In short, several systems that should, in principle, fall under the prohibitions and limitations of the AIR remain outside its scope of application. The Regulation establishes only a limited prohibition of certain AI practices; its governance approach, based on the acceptance of industrial risk, does not appear to provide adequate guarantees for fundamental rights; and reliance is placed on harmonised standards and common specifications that remain underdeveloped.

Conformity assessments for high-risk systems are entrusted to the providers themselves, while impact assessments of such systems on fundamental rights appear insufficient. In the name of expanding the European AI industry in external markets, the export of AI systems prohibited within the European Union for infringing fundamental rights is nevertheless permitted. The recognition of rights to lodge complaints and to receive explanations is more rhetorical than effective, and there is an evident absence of binding principles common to all AI systems. Finally, an excessive delay has been allowed in the entry into force of certain requirements concerning systems already placed on the market or put into service.

Incidentally, under Article 51(2) of the AIR, a general-purpose AI model is presumed to have systemic impact when the cumulative amount of computation used for its training, measured in floating-point operations, exceeds 10²⁵. This threshold is almost identical to the 10²⁶ level established by President Biden’s 2023 Executive Order on AI, a figure widely regarded by critical researchers in the United States as clearly insufficient.

This translation has been revised by María Amparo González Rúa from the original Spanish version, which can be consulted here.