The Architecture of the AI Illusion: Capital, Cognition, Power, and the Mythology of Technological Inevitability

Effina Driss, PhD 

Geopolitical Economist — US–China Rivalry, Europe’s Decline, Power, Trade & Strategic Dominance 

ABSTRACT 

Artificial intelligence has become the defining technological narrative of the early twenty-first century — a discourse so pervasive, and so systematically amplified, that it now functions less as a description of technological reality than as a structuring mythology for global capital, geopolitical competition, and institutional transformation. This article argues that what we term the ‘AI illusion’ is not an error of perception but a socially produced artifact: the cumulative output of interlocking financial incentives, geopolitical ambitions, corporate competitive pressures, and epistemic confusions about the nature of machine cognition. Drawing on political economy, science and technology studies, and strategic analysis, we dissect five dimensions of the illusion: the financial architecture that requires narrative inflation to sustain; the conceptual conflation of statistical pattern recognition with genuine intelligence; the geopolitical instrumentalization of AI as a tool of technological dependency and digital sovereignty; the organizational sociology of AI adoption driven by legitimacy pressures rather than operational utility; and the structural limits — energetic, physical, regulatory, and epistemic — that constrain the illusion’s indefinite expansion. We conclude that the illusion will not be dispelled by the failure of AI as a technology, which will continue to produce genuine and consequential innovations. It will be corrected by the progressive accumulation of empirical evidence about what AI systems can and cannot do — a process whose trajectory and timetable remain uncertain, but whose direction is not. 

Keywords: artificial intelligence; technological narrative; capital markets; geopolitics; digital sovereignty; cognitive science; automation; political economy of technology; AI governance; epistemic inflation. 

I. INTRODUCTION: THE NARRATIVE BEFORE THE TECHNOLOGY 

There is a standard way to write about artificial intelligence in 2026. One begins with the scale of investment — trillions of dollars committed by the world’s largest technology companies to the development of foundational models, data center infrastructure, and semiconductor supply chains. One proceeds to the capabilities: language models that draft legal briefs and poetry, image generators that produce photorealistic art on demand, coding assistants that write functional software from natural language descriptions. One concludes with the implications: a productivity revolution comparable to the Industrial Revolution, the transformation or elimination of entire professional categories, and the eventual emergence of systems that will surpass human cognitive capacity across virtually every domain. 

This article begins differently. It begins with a question: who benefits from this narrative, and how? Not in the conspiratorial sense of asking whether AI is a deliberate fraud — it is not — but in the analytical sense of asking which interests the dominant AI narrative serves, whose investments it justifies, whose geopolitical ambitions it advances, and whose institutional anxieties it exploits. The answers to these questions do not diminish the genuine technological significance of recent advances in machine learning and large language models. They do, however, reframe our understanding of why those advances are described in the particular way they are — with the particular urgency, the particular scope of implication, and the particular exclusion of counterevidence that currently characterizes mainstream AI discourse. 

The concept of the ‘AI illusion,’ as used in this article, requires careful definition. It does not refer to the nonexistence of AI capabilities. Large language models demonstrably produce useful outputs across a wide range of tasks; machine learning systems have materially improved outcomes in medical imaging, drug discovery, logistics optimization, and financial risk modeling. The illusion refers instead to the systematic inflation of AI’s current capabilities and near-term trajectory; the conceptual confusion between statistical optimization and human intelligence; the institutional dynamics that drive adoption beyond demonstrated utility; and the geopolitical amplification of AI’s transformative power as a tool of strategic positioning. In sum: the illusion is the gap between what these systems actually are and the mythological narrative constructed around them. 

The article proceeds in six substantive sections. Section II examines the financial architecture of AI narrative inflation. Section III dissects the central conceptual confusion at the core of the illusion — the equation of pattern recognition with cognition. Section IV analyzes the geopolitical instrumentalization of AI discourse. Section V examines organizational sociology and the legitimacy dynamics of corporate AI adoption. Section VI maps the structural limits of the illusion. Section VII assesses the trajectory and conditions of correction. Section VIII concludes. 

II. THE FINANCIAL ARCHITECTURE OF NARRATIVE INFLATION 

2.1 Capital Commitments and the Necessity of the Narrative 

The relationship between capital allocation and technological narrative is not incidental; it is structurally necessary. As of 2025, the five largest US technology companies had collectively committed over $300 billion to AI infrastructure over a two-year horizon: data centers, custom silicon, energy supply, and talent acquisition.¹ Microsoft’s partnership with OpenAI represents a commitment of $13 billion; Google’s DeepMind restructuring and Gemini development have consumed comparable resources; Amazon’s AWS AI infrastructure investments have been described by the company’s CEO as the ‘largest capital expenditure in Amazon’s history.’ Meta, despite a period of investor skepticism following its metaverse pivot, returned to AI infrastructure spending at a scale that rewarded early shareholders and sustained the company’s market valuation. 

These commitments create an irreversible structural dependency between capital market valuations and the AI narrative. Once a company has publicly committed hundreds of billions to a technological infrastructure, the narrative of that infrastructure’s inevitability is no longer a neutral analytical position — it is an institutional necessity. The alternative narrative — that AI returns will be partial, delayed, uneven, and structurally constrained — carries catastrophic implications for market capitalizations, management credibility, and the ongoing ability to raise capital for continued investment. The AI narrative is therefore not merely a description of technological prospects; it is a component of the financial architecture that enables the capital flows sustaining AI development. 

This dynamic generates what might be termed ‘epistemic leverage’: the capacity of financially interested parties to shape public understanding of a technology’s capabilities and trajectory through selective emphasis, forward-looking projections, and the systematic underweighting of counterevidence. The mechanism is well-documented in the financial economics literature on analyst forecasting: analysts whose institutions have underwriting relationships with the companies they cover systematically issue more optimistic forecasts than independent analysts. A similar structural bias operates in AI discourse, where the most prominent and widely cited voices — company executives, venture capital partners, and sponsored researchers — are precisely those with the greatest financial interest in the narrative’s continuation. 

2.2 The Hype Cycle and Its Asymmetric Costs 

Gartner’s ‘hype cycle’ model describes a recurring pattern in technology adoption: a period of inflated expectations following a new technology’s introduction, followed by a ‘trough of disillusionment’ when initial implementations fall short of promises, and eventually a ‘slope of enlightenment’ in which realistic assessments of the technology’s actual utility emerge.² What the model does not adequately address is the asymmetric distribution of costs and benefits across the cycle’s phases. During the peak of inflated expectations, the primary beneficiaries are those who invested earliest — founders, early-stage venture capitalists, and company employees with equity compensation. The costs are borne by later-stage investors, by organizations that adopt technologies prematurely, and by workers whose jobs are restructured around AI tools that deliver partial or unreliable results. 

The dot-com parallel is instructive. Between 1995 and 2000, the internet was genuinely transforming economic activity — email, e-commerce, and online information retrieval were real and consequential innovations. Yet the narrative around the internet significantly exceeded its near-term economic reality. When the correction came in 2000–2001, it was not because the internet failed as a technology but because the gap between narrative and reality became unsustainable. The internet subsequently delivered on many of its long-term promises — but on a timeline measured in decades rather than years, and with a distribution of value that differed substantially from what the peak-period narrative suggested. The current AI cycle shares this structure at a scale and speed that has no precise historical precedent. 

The illusion is not in the technology. It is in the systematic compression of the timeline between present capability and promised transformation — a compression that serves the interests of capital markets far more than the interests of institutional decision-makers, workers, or citizens. 

2.3 Sovereign Investment and the Nationalization of the Narrative 

The financial architecture of AI narrative inflation is not confined to private capital markets. States have become major direct investors in AI infrastructure, motivated by both genuine strategic considerations and the political economy of technological nationalism. The United States CHIPS Act committed $52 billion to domestic semiconductor manufacturing. China’s ‘New Generation AI Development Plan’ targets global AI leadership by 2030. Gulf sovereign wealth funds — including Saudi Arabia’s Public Investment Fund and the UAE’s ADQ — have committed billions to AI infrastructure as elements of economic diversification strategies. State investment in AI creates a second-order narrative dynamic: once a government has publicly committed to AI leadership as a strategic national objective, the political costs of acknowledging the technology’s limitations become significant. The narrative of technological inevitability becomes fused with national ambition in ways that make critical assessment politically costly, producing a systematic bias in public-sector AI discourse that reinforces, rather than corrects, the inflationary dynamics of private capital markets. 

III. THE COGNITIVE CONFUSION: INTELLIGENCE, STOCHASTICITY, AND THE TURING TRAP 

3.1 What Language Models Actually Are 

The most consequential conceptual confusion at the heart of the AI illusion is the equation of statistical pattern recognition with cognition, and of fluent language generation with understanding. Large language models are, at the computational level, extremely high-dimensional statistical systems trained to predict the probability distribution of text tokens given preceding context. They are not, in any technically rigorous sense, ‘thinking.’ They are not building world models, forming beliefs, drawing inferences in the logical sense, or possessing awareness of any kind. What they are doing — with extraordinary technical sophistication — is interpolating and extrapolating from vast corpora of human-generated text to produce outputs that are statistically consistent with that corpus. 

This description is not a dismissal. The outputs of these systems are genuinely useful across a wide range of tasks — drafting, summarization, translation, code generation, question answering within domains well-represented in training data. The point is rather that the mechanisms producing these useful outputs are fundamentally different from the mechanisms of human cognition, and that this difference has profound implications for the reliability, scope, and appropriate deployment of AI systems. A system that predicts likely text continuations cannot, by definition, reliably distinguish between what it ‘knows’ and what it confidently hallucinates — because the distinction between accurate recall and plausible confabulation is not architecturally available to it. Its outputs are not beliefs with calibrated confidence levels; they are probability-weighted text continuations that may or may not correspond to reality. 

3.2 The Turing Trap and the Epistemology of Fluency 

Alan Turing’s 1950 paper ‘Computing Machinery and Intelligence’ introduced the ‘imitation game’ — a behavioral test for machine intelligence based on the ability to produce text indistinguishable from human output.³ The test was philosophically controversial from the outset: John Searle’s Chinese Room argument questioned whether behavioral indistinguishability constitutes genuine intelligence or merely sophisticated simulation. What had been a philosophical controversy has become a practical problem: contemporary large language models pass informal versions of the Turing Test with regularity, yet they manifestly lack the capacities — contextual judgment, causal reasoning, embodied understanding, genuine reference to the world — that most accounts of human intelligence require. 

The persuasiveness of AI outputs has thus become, paradoxically, an epistemological hazard. When a language model produces a confident, fluent, internally coherent text on a complex topic, the output’s surface qualities create a strong inference of understanding. This inference is systematically misleading. The model has no understanding of the topic; it has a statistical model of how texts about that topic are typically structured. The gap between apparent competence and actual competence is not visible in the output — which is precisely what makes it dangerous in high-stakes deployment contexts: medical diagnostics, legal reasoning, intelligence analysis, and strategic decision-making. 

3.3 Benchmarks, Goodhart’s Law, and Capability Theater 

The AI industry has developed a sophisticated ecosystem of benchmarks — standardized tests designed to measure system capabilities across domains including reasoning, mathematics, coding, medical knowledge, and legal analysis. The narrative of AI progress is substantially driven by benchmark performance: state-of-the-art results on MMLU, HumanEval, MATH, or Bar Exam simulations generate headlines announcing AI’s encroachment on professional expertise. These benchmark improvements are real. Their interpretation, however, requires substantial caution. Goodhart’s Law — ‘when a measure becomes a target, it ceases to be a good measure’ — applies with particular force here.⁴ AI systems are now trained with access to benchmark-adjacent data, optimized through feedback signals that reward benchmark-consistent outputs, and evaluated on benchmarks that may not capture the distribution of real-world deployment conditions. A system that achieves 90 percent accuracy on a law school exam benchmark may fail systematically on the kind of open-ended, context-dependent, precedent-sensitive reasoning that actual legal practice requires. This phenomenon — ‘capability theater’ — is a structural feature of the current AI development environment, not an incidental failing of particular systems. 

IV. THE GEOPOLITICAL INSTRUMENTALIZATION OF AI DISCOURSE 

4.1 AI as Infrastructure of Technological Dependence 

The geopolitical dimensions of the AI narrative are rarely foregrounded in mainstream discussions, which tend to present AI development as a neutral process of technological progress driven by scientific advancement and market forces. A strategic reading reveals a more complex picture: the development and global adoption of specific AI platforms, cloud computing infrastructures, and semiconductor supply chains creates forms of technological dependence with profound implications for national sovereignty, economic autonomy, and long-term strategic positioning. The concentration of AI capability in a small number of US technology companies — alongside a parallel concentration in Chinese state-adjacent entities — means that global AI adoption effectively requires either dependency on US commercial infrastructure or alignment with the Chinese technological ecosystem. For the majority of the world’s nations, AI adoption is not a neutral choice between functionally equivalent tools; it is a strategic alignment decision with implications for data sovereignty, regulatory exposure, and the degree to which critical national systems operate on infrastructure controlled by a foreign commercial or state actor. 

4.2 The Semiconductor Wars and the Hardware Dimension 

The geopolitical stakes of AI development are most concretely visible in the global competition over advanced semiconductor manufacturing. Advanced AI systems require chips — primarily GPUs and specialized AI accelerators — manufactured at process nodes that only two companies in the world are currently capable of producing: Taiwan Semiconductor Manufacturing Company (TSMC) and Samsung. The US government’s export controls on advanced semiconductors to China, implemented in October 2022 and progressively tightened through 2025, represent the most consequential industrial policy intervention in the global technology sector in decades.⁵ These controls reflect a geopolitical calculation — that denying China access to the computational infrastructure required for advanced AI development is a legitimate national security objective — that is simultaneously an implicit acknowledgment of AI’s strategic importance and an instrument for maintaining US technological hegemony. The irony is considerable: the US government’s semiconductor export controls validate the narrative of AI’s strategic importance that technology companies have been constructing to justify their capital expenditures, while simultaneously demonstrating that the ‘neutral’ diffusion of AI technology is inseparable from questions of power, sovereignty, and strategic competition. 

4.3 Data Colonialism and the Political Economy of Training 

Large language models are trained on vast corpora of human-generated text scraped from the global internet with little transparency about sources, selection criteria, or copyright status. The models’ capabilities are, in a direct sense, the crystallized output of humanity’s collective linguistic and intellectual production — a public good whose value has been captured by a small number of private entities. This dynamic has been termed ‘data colonialism’ by Couldry and Mejias: the extraction of value from human social activity for the benefit of entities that did not produce that activity and do not compensate its producers.⁶ The populations whose digital activity has been most extensively mined for AI training have received no compensation, while the entities that captured and commodified their output have achieved market capitalizations in the trillions. The AI narrative of universal benefit and democratized capability systematically elides this distributional question — rendering invisible the mechanism by which AI’s value has been accumulated at the expense of its unknowing contributors. 

V. ORGANIZATIONAL SOCIOLOGY: THE LEGITIMACY TRAP 

5.1 Institutional Isomorphism and Competitive AI Adoption 

DiMaggio and Powell’s foundational analysis of institutional isomorphism identified three mechanisms through which organizations in a field tend to become structurally similar: coercive pressures, mimetic pressures, and normative pressures.⁷ All three are operating in the current AI adoption environment, generating a dynamic in which organizations adopt AI tools not primarily because of demonstrated operational benefit but because of structural pressures that make non-adoption increasingly costly from a legitimacy perspective. The coercive dimension is visible in the growing integration of AI-related requirements into regulatory and procurement frameworks. The normative dimension is visible in the professionalization of ‘AI strategy’ as a distinct organizational function, with associated consulting practices, credentialing programs, and executive compensation structures that reward AI adoption decisions regardless of their demonstrated utility. The mimetic dimension — perhaps the most powerful — manifests in peer-pressure dynamics of executive discourse: CEOs who publicly describe AI as central to their organization’s strategy create pressure on competitors to articulate equivalent commitments, regardless of whether either organization has rigorously evaluated the operational case. 

5.2 The Measurement Problem and the Illusion of Impact 

One of the most structurally important features of the current AI adoption environment is the difficulty of measuring AI’s actual impact on organizational productivity and outcomes. Unlike investments in manufacturing capacity or logistics infrastructure, the returns from AI adoption are diffuse, difficult to attribute, and often confounded by simultaneous changes in other organizational systems. This measurement difficulty creates a structural opportunity for the AI illusion to persist within organizations: when returns cannot be clearly measured, they cannot be clearly contradicted, and the narrative of AI’s transformative potential remains unchallenged by organizational experience. Daron Acemoglu’s analysis suggests that current AI systems are economically viable for a substantially smaller share of tasks than the dominant narrative implies, and that productivity gains from viable applications are frequently offset by implementation costs, retraining, error management, and the overhead of human oversight that responsible AI deployment requires.⁸ Erik Brynjolfsson’s work on generative AI in the workplace has found significant heterogeneity in outcomes — with gains concentrated among lower-skill workers performing routine text tasks and gains absent or negative for expert workers performing complex, judgment-intensive work. 

5.3 The Labor Displacement Paradox 

No dimension of the AI narrative has generated more social anxiety than the claim that AI will displace large shares of human employment. Estimates range from McKinsey Global Institute’s projection of 400 million workers globally displaced by 2030 to the OECD’s finding that approximately 14 percent of jobs in advanced economies are at high automation risk. The historical record on technological displacement is more nuanced than either the catastrophist or the optimist narrative acknowledges. The ATM machine, widely predicted to eliminate bank tellers when introduced in the 1970s, was associated with an increase in the number of bank tellers over the subsequent two decades — because lower operating costs per branch enabled the opening of more branches, increasing demand for tellers in customer-facing roles that machines could not fully perform. The AI case is likely to exhibit a similarly complex pattern. The tasks that will grow in value are precisely those that require the capabilities AI systems most conspicuously lack: contextual judgment, ethical reasoning, stakeholder navigation, creative synthesis, and the tacit knowledge that accumulates through embodied experience. The narrative of universal displacement is as misleading as the narrative of universal complementarity: the AI labor market transition will be highly uneven, heavily dependent on institutional responses, and fundamentally shaped by political choices about education, social protection, and the governance of AI deployment. 

VI. STRUCTURAL LIMITS OF THE ILLUSION 

6.1 The Energy Wall 

The physical infrastructure supporting current AI development faces constraints that the dominant narrative consistently underweights. Training a single large foundation model requires energy consumption equivalent to the annual electricity usage of hundreds of households. The aggregate energy consumption of global AI infrastructure is growing rapidly: Goldman Sachs Research estimates that data center power demand will increase by 160 percent between 2023 and 2030, driven primarily by AI workloads.⁹ This energy trajectory collides with two structural realities: the carbon neutrality commitments of major technology companies, and the physical constraints on electricity grid expansion in the regions where data center infrastructure is most concentrated. Microsoft, Google, and Amazon have made public commitments to carbon neutrality on timelines that are structurally incompatible with the energy consumption trajectories implied by their AI investment commitments, absent either a dramatic acceleration of renewable energy deployment or a significant reduction in the energy intensity of AI workloads. 

6.2 The Regulation Ratchet 

Regulatory frameworks for AI are in a formative phase globally, but the direction of travel is clearly toward more stringent requirements for transparency, accountability, safety, and human oversight. The European Union’s AI Act — the most comprehensive regulatory framework yet enacted — categorizes AI systems by risk level and imposes requirements ranging from transparency disclosures to conformity assessments, human oversight requirements, and prohibition of certain applications.¹⁰ The regulatory trajectory has two structural implications for the AI illusion. First, it imposes costs and constraints on AI deployment that the dominant narrative systematically ignores: compliance costs, liability exposure, restricted use cases, and mandatory human oversight requirements that reduce the efficiency gains attributed to AI automation. Second, it creates accountability mechanisms that will generate empirical records of AI system performance, failure modes, and real-world impacts — records systematically more reliable than self-reported performance claims. The accumulation of regulatory oversight data will serve as a corrective force against narrative inflation, providing evidence bases for realistic assessment that the current environment systematically prevents. 

6.3 The Epistemological Limit: What AI Cannot Know 

Beyond the practical constraints of energy and regulation, there is a deeper epistemological limit to AI’s transformative potential that the dominant narrative rarely engages: the boundary between pattern recognition and genuine knowledge. Large language models are trained on historical data; they can only reproduce, interpolate, and extrapolate from patterns present in their training corpora. They cannot generate genuine novelty in the sense of knowledge that exceeds what is implicit in the data they were trained on. They cannot reliably reason about novel situations whose structure differs significantly from training data patterns. And they cannot update their knowledge in response to new evidence in the way human experts do: their knowledge is frozen at training time, subject only to the limited updating available through context windows and fine-tuning. A system that cannot distinguish between confident expression of accurate knowledge and confident expression of plausible confabulation is not a reliable tool for any application where this distinction matters — which includes the entire domain of high-stakes, evidence-sensitive, open-ended professional judgment. 

VII. THE TRAJECTORY OF CORRECTION 

The AI illusion will not be corrected by a single dramatic failure analogous to the dot-com crash or the 2008 financial crisis. Technological narratives correct more slowly and more unevenly than financial bubbles, because they are embedded in institutional practices, professional identities, and organizational strategies that resist rapid revision. The correction will instead proceed through the gradual accumulation of empirical evidence from three sources. First, as organizations mature in their AI implementation experience, they will develop more realistic assessments of operational impact. The initial phase — characterized by pilot projects, executive enthusiasm, and narrative-driven investment — is giving way in many organizations to a more demanding evaluation phase, in which the question shifts from ‘are we using AI?’ to ‘what is AI actually delivering, and at what cost?’ The emerging evidence is already more mixed than the dominant narrative acknowledges. Second, the regulatory accountability mechanisms currently being constructed will generate empirical records of AI system performance, failure, and impact that are systematically more reliable than industry self-reporting. Third, and perhaps most importantly, the physical constraints on AI infrastructure expansion will impose a natural limit on the narrative of exponential growth. When the rate of capability improvement begins to slow — as all technological learning curves eventually do — the gap between narrative promise and empirical performance will widen to the point of becoming publicly visible. 

VIII. CONCLUSION: THE PRODUCTIVE VALUE OF DISILLUSIONMENT 

The argument of this article is not that artificial intelligence is a fraud, a bubble, or a technology whose importance has been fundamentally misattributed. It is rather that the dominant narrative about AI systematically conflates genuine technological progress with speculative projection, financial incentive, geopolitical ambition, and epistemic confusion about the nature of machine cognition. This conflation serves the interests of capital markets, technology companies, and strategic competitors in ways that consistently override the interests of institutional decision-makers, workers, citizens, and the societies for whom the technology is nominally being developed. 

The concept of ‘productive disillusionment’ — the corrective process through which inflated technological expectations are progressively refined by empirical evidence — is not a failure of innovation; it is its most important precondition. The railroad did not become economically transformative during its initial speculative bubble; it became transformative when the bubble burst and the genuine infrastructure was built by investors and operators who understood what railroads could and could not do. The internet did not deliver its full transformative potential during the peak of dot-com enthusiasm; it delivered that potential over the subsequent two decades, through the patient construction of applications that addressed genuine human needs in economically sustainable ways. 

Artificial intelligence will follow an analogous trajectory. Its genuine capabilities — in narrow, well-defined task domains, under appropriate human oversight, integrated into broader human systems — are significant and will continue to expand. Its claimed capabilities as a general reasoning engine, an autonomous agent capable of replacing human judgment across complex domains, or an imminent agent of economic disruption comparable to the Industrial Revolution will be progressively corrected by the accumulation of empirical evidence. The correction will be painful for those who have committed institutional strategies, career identities, and capital allocations to the inflated narrative. It will be productive for those who have maintained the analytical discipline to distinguish between what these remarkable systems actually are and what it is currently profitable to claim they might become. 

The greatest risk of the AI illusion is not that the technology will fail. It is that the gap between narrative and reality will be exploited — by those who understand it — at the expense of those who do not. Strategic lucidity about what AI can and cannot do is not a counsel of technological pessimism. It is the necessary precondition for using it wisely. 

NOTES 

¹ Goldman Sachs, ‘AI Infrastructure: Where Are We in the Buildout?’ June 2024; Bloomberg, ‘Big Tech AI Capital Expenditure Tracker 2025,’ updated quarterly. 

² Jackie Fenn and Mark Raskino, Mastering the Hype Cycle (Cambridge, MA: Harvard Business Press, 2008); Gartner, ‘Hype Cycle for Artificial Intelligence, 2024,’ August 2024. 

³ Alan Turing, ‘Computing Machinery and Intelligence,’ Mind 59, no. 236 (1950): 433–460; John Searle, ‘Minds, Brains, and Programs,’ Behavioral and Brain Sciences 3, no. 3 (1980): 417–424; Hubert Dreyfus, What Computers Can’t Do (New York: Harper & Row, 1972). 

⁴ David Manheim and Scott Garrabrant, ‘Categorizing Variants of Goodhart’s Law,’ arXiv preprint, 2018; Mitchell et al., ‘Measuring Massive Multitask Language Understanding,’ NeurIPS 2020. 

⁵ US Bureau of Industry and Security, ‘Export Controls on Advanced Computing and Semiconductor Manufacturing Items,’ Federal Register, October 2022; Chris Miller, Chip War (New York: Scribner, 2022). 

⁶ Nick Couldry and Ulises Mejias, The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism (Stanford: Stanford University Press, 2019). 

⁷ Paul DiMaggio and Walter Powell, ‘The Iron Cage Revisited,’ American Sociological Review 48, no. 2 (1983): 147–160. 

⁸ Daron Acemoglu, ‘The Simple Macroeconomics of AI,’ NBER Working Paper 32487, May 2024; Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, ‘Generative AI at Work,’ NBER Working Paper 31161, 2023. 

⁹ Goldman Sachs Research, ‘AI Is Poised to Drive 160% Increase in Data Center Power Demand,’ May 2024; IEA, ‘Electricity 2024: Analysis and Forecast to 2026,’ January 2024. 

¹⁰ European Parliament and Council, ‘Artificial Intelligence Act,’ Regulation (EU) 2024/1689, June 2024; OECD, ‘OECD AI Principles,’ updated 2023. 

REFERENCES 

Acemoglu, Daron. ‘The Simple Macroeconomics of AI.’ NBER Working Paper 32487. Cambridge, MA: National Bureau of Economic Research, May 2024. 

Autor, David. ‘Work of the Past, Work of the Future.’ AEA Papers and Proceedings 109 (2019): 1–32. 

Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. ‘Generative AI at Work.’ NBER Working Paper 31161. Cambridge, MA: NBER, 2023. 

Couldry, Nick and Ulises Mejias. The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford: Stanford University Press, 2019. 

DiMaggio, Paul and Walter Powell. ‘The Iron Cage Revisited: Institutional Isomorphism and Collective Rationality in Organizational Fields.’ American Sociological Review 48, no. 2 (1983): 147–160. 

Dreyfus, Hubert. What Computers Can’t Do: A Critique of Artificial Reason. New York: Harper & Row, 1972. 

European Parliament and Council. ‘Artificial Intelligence Act.’ Regulation (EU) 2024/1689. June 2024. 

Fenn, Jackie and Mark Raskino. Mastering the Hype Cycle: How to Choose the Right Innovation at the Right Time. Cambridge, MA: Harvard Business Press, 2008. 

Gartner. ‘Hype Cycle for Artificial Intelligence, 2024.’ Stamford, CT: Gartner Inc., August 2024. 

Goldman Sachs Research. ‘AI Is Poised to Drive 160% Increase in Data Center Power Demand.’ New York, May 2024. 

International Energy Agency. ‘Electricity 2024: Analysis and Forecast to 2026.’ Paris: IEA, January 2024. 

Manheim, David and Scott Garrabrant. ‘Categorizing Variants of Goodhart’s Law.’ arXiv preprint arXiv:1803.04585, 2018. 

McKinsey Global Institute. ‘A New Future of Work: The Race to Deploy AI and Raise Skills.’ February 2023. 

Miller, Chris. Chip War: The Fight for the World’s Most Critical Technology. New York: Scribner, 2022. 

OECD. ‘OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market.’ Paris: OECD, 2023. 

Searle, John. ‘Minds, Brains, and Programs.’ Behavioral and Brain Sciences 3, no. 3 (1980): 417–424. 

Turing, Alan. ‘Computing Machinery and Intelligence.’ Mind 59, no. 236 (1950): 433–460. 

US Bureau of Industry and Security. ‘Export Controls on Advanced Computing and Semiconductor Manufacturing Items.’ Federal Register 87, no. 197, October 2022. 

World Economic Forum. ‘Future of Jobs Report 2025.’ Geneva: WEF, 2025.