Angus Deaton

Foreword

By Edo Segal

The number that broke my confidence was not a technology number.

It was seventy to one. The ratio between the richest nations and the poorest at the end of the twentieth century. Before the industrial revolution, it was five to one. The great escape happened — life expectancy doubled, extreme poverty fell, billions of lives improved by every material measure — and the gap widened from five to seventy. Not because the poorest got poorer. Because the escapees escaped so far, so fast, that the distance between them and everyone else became a chasm.

I did not know that number when I wrote The Orange Pill. I knew about inequality in the abstract way that builders know about it — as a problem someone else would solve while I focused on making the tools better and cheaper and more accessible. I believed, genuinely, that lowering the barrier to building was a moral act. I still believe that. But Angus Deaton forced me to ask a question I had been avoiding: What happens when the people who benefit first from a new technology compound their advantage so quickly that the people who benefit last can never catch up?

Every technology book in this series hands you a lens. Csikszentmihalyi gave us flow. Han gave us friction. Kauffman gave us complexity at the edge of chaos. Deaton gives us something more uncomfortable. He gives us the distribution.

The aggregate is always impressive. GDP rises. Productivity climbs. Capability expands. The aggregate is also a lie — not because the numbers are wrong, but because they hide who captured the gain and who bore the cost. Deaton spent his career pulling the aggregate apart and counting what was inside. What he found, over and over, was that progress is real and progress is uneven, and the unevenness is not a side effect. It is a structural feature of how technologies move through populations.

This matters for AI more than it mattered for any previous technology, because the speed is faster and the amplification is greater. The twenty-fold productivity gain I documented in Trivandrum is real. It is also available, right now, primarily to people who already had every structural advantage. The question Deaton forces is not whether the gain is real. It is whether the gain will reach the people who need it most before the distance between them and the early escapees becomes permanent.

That question cannot be answered by building better tools. It can only be answered by building better institutions. Deaton shows you why.

— Edo Segal ^ Opus 4.6

About Angus Deaton

1945-present

Angus Deaton (1945–present) is a British-American economist and Nobel laureate whose career has been devoted to measuring how economic progress actually reaches — or fails to reach — the people it is supposed to serve. Born in Edinburgh, Scotland, he studied at Cambridge before spending the majority of his academic career at Princeton University, where he is the Dwight D. Eisenhower Professor of Economics and International Affairs Emeritus. His landmark book The Great Escape: Health, Wealth, and the Origins of Inequality (2013) documented the extraordinary improvements in human welfare over the past two and a half centuries while demonstrating that these improvements were distributed in patterns that deepened inequality between and within nations. With Anne Case, he identified and named "deaths of despair" — the startling rise in mortality among non-college-educated Americans driven by suicide, drug overdose, and alcoholic liver disease — in research that reshaped public understanding of inequality in wealthy nations. He was awarded the Nobel Memorial Prize in Economic Sciences in 2015 for his work on consumption, poverty, and welfare. In 2024, he published a remarkable intellectual self-reckoning in the IMF's Finance & Development, acknowledging that mainstream economics had underestimated the role of power in shaping who benefits from technological change — and warning that artificial intelligence would intensify these dynamics unless institutional countermeasures were deliberately built.

Chapter 1: The Pattern of Escape

In 2024, Angus Deaton published an essay in the International Monetary Fund's Finance & Development that amounted to something rare in the career of a Nobel laureate: a confession. The essay, titled "Rethinking My Economics," acknowledged that mainstream economics — the discipline Deaton had practiced and advanced for half a century — had failed to anticipate the consequences of the very forces it had championed. Globalization had produced aggregate gains while devastating specific communities. Market liberalization had increased efficiency while concentrating power. Technological change had raised productivity while hollowing out the livelihoods of millions who lacked the credentials or the geography to participate in the new economy. "I have changed my mind," Deaton wrote, with the directness of a man who had spent decades insisting that evidence, not ideology, should determine conclusions. The evidence had spoken, and what it said was uncomfortable.

The essay contained a sentence about artificial intelligence that, in its brevity, carried the weight of an entire analytical framework. Deaton noted that "the continuing rapid development of artificial intelligence means that this technological transition will endure." He placed this observation in the context of what he called "pervasive economies of scale more powerful than older industries," in which a small number of firms wield significant market power and pinpointing value creation has become "next to impossible." The sentence was not a prediction about AI's capabilities. It was a diagnosis of AI's distributional structure — who would capture the gains, who would bear the costs, and why the default outcome, absent deliberate institutional intervention, would be concentration rather than diffusion.

This diagnosis emerges from a career spent studying what Deaton calls the great escape — the extraordinary improvement in human welfare over the past two and a half centuries. Life expectancy has roughly doubled since the industrial revolution. Extreme poverty, which encompassed the vast majority of the human population for most of recorded history, has declined to levels that would have seemed utopian to any observer of the eighteenth century. Infant mortality has fallen. Literacy has spread. The caloric intake available to the average human being has increased dramatically. These are not projections or aspirations. They are measurements, assembled with the care that Deaton has brought to every empirical question he has addressed, and they document a transformation in the material conditions of human existence that is, by any historical standard, remarkable.

But the aggregate conceals the distribution. This is the sentence that could serve as the epigraph for Deaton's entire body of work, and it is the sentence that transforms his analysis of AI from a commentary on technology into something more urgent: a warning about the structure of progress itself. The great escape is real. The gains are measurable. And the gains are distributed in patterns that reproduce, and in many cases deepen, the inequalities that preceded them. For every population that has escaped from poverty, disease, and constrained capability, there are populations that have not escaped, and the distance between the two groups has widened even as the absolute conditions of both groups have improved. The escapees are genuinely better off. The left-behind are, in relative terms, worse off than before the escape began — not because their conditions have necessarily deteriorated, but because the gap between their conditions and the conditions of those who have escaped has grown in ways that create new competitive disadvantages, new political resentments, and new forms of exclusion.

The pattern is structural. It recurs across every major technological transition in the historical record. Writing escaped elites from the limitations of oral memory centuries before it reached the general population. Printing escaped scholars from the monopoly of scriptoria decades before it democratized reading. Vaccination escaped wealthy nations from smallpox while the disease continued to devastate populations without access to the technology. The green revolution escaped farmers with access to improved seeds and fertilizers while leaving those without access further behind. In each case, the escape was genuine, the long-term benefits were eventually distributed more broadly, and the transitional period was marked by inequality severe enough to reshape political structures and social hierarchies.

The AI transition exhibits this pattern with unusual clarity and unusual speed. Edo Segal's The Orange Pill documents the escape as experienced from inside a technology organization — the exhilaration of engineers in Trivandrum discovering capabilities they did not know they possessed, the developer in Lagos gaining access to coding leverage previously available only to engineers at major technology firms, the twenty-fold productivity improvements that collapse the distance between imagination and execution. These are genuine escapes from capability deprivation. The engineers really did gain new abilities. The productivity improvements really are measurable. The barriers to creative and productive work really have been lowered for populations that previously faced them.

Deaton's framework does not dispute any of this. What it demands is the question that the excitement of the escape tends to obscure: Who escapes first, who escapes last, and what happens in the gap between them?

The historical evidence, assembled across Deaton's career, reveals a consistent answer. The first beneficiaries of any technological transition are those with existing structural advantages: education, infrastructure, institutional support, economic security, and the cognitive flexibility that comes from prior exposure to technological change. The last beneficiaries are those without these advantages. And the transitional period — the period during which the early escapees pull away from the late arrivals — is the period of greatest danger, because the divergence in capability creates competitive advantages that can become self-reinforcing if they are not addressed by deliberate intervention.

The self-reinforcing character of the divergence is the feature that distinguishes Deaton's analysis from simpler accounts of technological inequality. The argument is not merely that some populations benefit before others — this is obvious and, in itself, not necessarily harmful. The argument is that the early beneficiaries accumulate advantages that make it progressively harder for the late arrivals to catch up. The firms that adopt AI first capture market share, attract talent, and reinvest profits in further AI development, pulling ahead of firms that adopt later. The nations that lead in AI development build institutional capacity to govern AI effectively, attracting further investment and talent, while the nations that lag lack both the technology and the institutional capacity to close the gap. The workers who adapt quickly gain experience and facility with the tools, commanding higher wages and more interesting work, while the workers who adapt slowly find their existing skills devalued and their career trajectories narrowed.

This is the mechanism that produced what Deaton and Anne Case documented as "deaths of despair" — the startling increase in mortality among white, non-college-educated Americans driven by suicide, drug overdose, and alcoholic liver disease. The mechanism was not poverty in the absolute sense. It was the collapse of the institutional structures — stable employment, union membership, community organizations, the social identity that comes from productive work — that had given meaning and structure to working-class life. When deindustrialization and automation hollowed out the manufacturing economy, the aggregate statistics showed increased productivity and economic growth. The distributional reality showed communities in freefall.

The deaths of despair framework has already been extended to AI by multiple scholars. A 2024 paper titled "Do Robots Cause Deaths of Despair?" tested Deaton and Case's hypothesis directly against automation data. The finding was not that automation caused despair mechanically, but that automation concentrated in communities without adequate institutional support produced precisely the social destruction that Deaton's framework predicted. The implication for AI is direct: the technology's effects on human welfare will be determined not by its capabilities but by the institutional context in which it is deployed.

Deaton's 2024 essay endorsed this institutional perspective explicitly. Citing Daron Acemoglu and Simon Johnson, he argued that "the direction of technical change has always depended on who has the power to decide" and that "unions need to be at the table for decisions about artificial intelligence." This is not a statement about AI's technical capabilities. It is a statement about power — about who controls the deployment of AI, whose interests the deployment serves, and whether the institutions that could ensure broader distribution of the benefits are strong enough to perform that function.

The Orange Pill captures the escape with a practitioner's intimacy. Deaton's framework provides the distributional analysis that the practitioner's perspective, by its nature, tends to underweight. The two perspectives are not opposed. They are complementary in the way that a pilot's experience of flight and an aeronautical engineer's analysis of the forces acting on the aircraft are complementary — both necessary, neither sufficient, and the gap between them precisely the space in which understanding must be constructed.

The chapters that follow apply Deaton's distributional framework systematically to the AI transition. The analysis begins with the question of who escapes first and examines the structural advantages that determine the order of escape. It proceeds to the specific barriers — infrastructure, education, language, institutional capacity — that gate access to the escape for the populations that need it most. It examines the measurement problems that make productivity gains look more uniformly beneficial than they are. It investigates the geography of disruption and the specific populations most vulnerable to what this analysis will call the premature death cross — the moment when existing capabilities become economically obsolete before alternative capabilities are accessible. And it concludes with the institutional interventions that Deaton's framework identifies as necessary to ensure that the AI escape becomes broadly shared progress rather than a new and wider divide.

The stakes are measured in the currency that Deaton has spent his career counting: years of life expectancy, children's educational attainment, the capacity of individuals and communities to participate in the economic and social life of their societies. The pattern of escape is a pattern of human flourishing and human suffering, and the distribution of the escape is a moral question as much as an economic one. The aggregate is not enough. The distribution is where the justice — or the injustice — resides.

---

Chapter 2: Who Escapes First

The order of escape is not random. It follows lines of structural advantage that are as predictable as they are stubborn, and the predictability is itself a source of both analytical power and moral urgency. If the pattern can be predicted — and Deaton's research across multiple technological transitions demonstrates that it can — then the institutional interventions necessary to alter the pattern can, in principle, be identified and implemented. But identification and implementation are separated by a gap that is political rather than intellectual, and the political gap is where most distributional promises go to die.

The first beneficiaries of the AI escape are knowledge workers in wealthy nations employed by technology firms with the resources to invest in adoption. These are the populations for whom The Orange Pill's account of the transition is most immediately recognizable — the engineers in Trivandrum whose employer provided the tools, maintained their employment during the learning curve, and supported the adaptation process with management attention and organizational commitment. These populations possess every structural advantage that determines the speed of escape: education that provides the domain knowledge necessary to direct AI tools productively, infrastructure that provides reliable high-speed connectivity, institutional contexts that support experimentation, and economic security that allows risk-taking without the fear of destitution.

The structural advantages are worth enumerating with specificity, because their absence in other populations is what determines the distributional shape of the transition. Education provides not merely technical skills but the domain knowledge that AI tools amplify. The Trivandrum engineers benefited from AI because they had years of engineering training and professional experience that gave them the judgment to evaluate the tools' output, the knowledge to direct the tools toward productive ends, and the capacity to identify when the output was wrong. Without this educational foundation, the same tools produce dramatically less value — not because the tools are less capable but because the user lacks the knowledge base that the amplification requires.

Infrastructure provides the physical substrate for AI-augmented work: reliable electricity, high-speed internet connectivity, computing hardware capable of running modern software, and the physical workspace conducive to sustained cognitive effort. These are conditions so ubiquitous in the offices of technology firms in wealthy nations that they have become invisible — taken for granted in the same way that oxygen is taken for granted until it becomes scarce. In the contexts where the AI escape has been most dramatic, the infrastructure was already in place before the tools arrived. The tools were the last piece, not the first.

Institutional support provides the organizational framework within which adaptation occurs. The Trivandrum case illustrates managed escape — escape facilitated by an employer that bore the cost of the transition, provided training and guidance, and maintained economic security during the adaptation period. This institutional support is a form of organizational capital that is scarce and unevenly distributed. The populations that possess it are disproportionately employed by large, well-resourced firms in wealthy nations. The populations that lack it — the self-employed, the informally employed, the employees of firms operating on margins too thin to absorb the costs of technological transition — face the adaptation without the safety net that makes risk-taking possible.

Economic security provides the freedom to experiment. Adaptation to a new technology involves a period during which productivity may temporarily decline as the worker learns the new tools — a period during which the worker is investing time and cognitive resources in learning rather than producing. Workers with economic security can absorb this investment. Workers without economic security cannot afford the temporary decline in productivity, because the decline translates directly into reduced income, missed rent payments, and the cascading consequences of financial precarity.

The second tier of early escapees includes knowledge workers in wealthy nations employed outside the technology sector — lawyers, consultants, marketers, educators, healthcare professionals — who have education and infrastructure but face institutional contexts less supportive of rapid adoption. Professional norms, regulatory requirements, and organizational cultures moderate the pace of change in these sectors, producing an escape that is genuine but slower than the escape in the technology sector.

The third tier includes knowledge workers in middle-income nations — the engineer in Bangalore, the accountant in São Paulo, the teacher in Cairo — who have education and some infrastructure but face higher costs for connectivity and hardware, less institutional support for adoption, and economic conditions that make experimentation riskier. For these populations, the escape is real but attenuated, and the gap between their rate of adaptation and the rate of adaptation in wealthy nations creates competitive disadvantages that compound over time.

The fourth tier — the tier that Deaton's framework identifies with the greatest urgency — includes the billions of people in low-income nations who lack the basic prerequisites for AI adoption. Reliable connectivity. Affordable hardware. English-language fluency, which remains the primary medium of interaction with the most capable AI systems. Educational preparation for knowledge work. Institutional support for technological change. For these populations, the AI escape is not yet underway in any meaningful sense. The tools exist, but the conditions that translate tools into capability do not.

This ordering reproduces, with remarkable fidelity, the ordering of every previous technological escape. The populations that need the technology most receive it last. This is not a conspiracy but a structural feature of how technologies are distributed. Markets serve populations with purchasing power before populations without it. Institutions serve populations already connected to institutional networks before populations that are not. Knowledge reaches populations with existing educational infrastructure before populations without it. The result is that each technological transition reproduces, in its distributional shape, the inequalities that preceded it.

The Orange Pill acknowledges the barriers honestly and predicts they will fall. Deaton's response, informed by decades of studying exactly this kind of prediction, is that the prediction underestimates the duration and severity of the transitional period. The cost of AI inference is declining. The cost of connectivity is declining. These are real trends. But the barriers to AI adoption are not merely technical costs that can be resolved by declining price curves. They are political and economic structures — the distribution of educational investment, the geography of infrastructure, the terms of international trade, the governance of intellectual property — that reflect and reinforce existing power relations. These structures do not fall merely because the technology becomes cheaper. They fall only when the institutional mechanisms for delivering the technology to underserved populations are deliberately built, funded, and maintained.

The speed of the AI transition intensifies the distributional challenge. Previous escapes unfolded over decades, providing time — however inadequate — for institutions to adapt. The AI transition is creating differentiation in months. The twenty-fold productivity improvement documented for early adopters creates competitive advantages that compound with each passing quarter, making it progressively harder for late adopters to close the gap. The speed reduces the time available for the institutional responses that could narrow the divide, and it increases the penalty for late arrival.

The question of who escapes first is not merely descriptive. It is prescriptive. If the order of escape is predictable, the interventions necessary to alter it are identifiable: investment in infrastructure, in education, in institutional support, in the economic security that allows adaptation. These interventions are expensive and politically difficult. But they are the interventions that have, in every previous technological transition, determined whether the escape became broadly shared progress or a new and more durable form of inequality. The populations that are escaping last from the AI transition are the populations for whom these interventions matter most — and the populations for whom, historically, they arrive latest.

---

Chapter 3: The Capability Gap

The most dangerous feature of the AI transition is not the divide between those who have AI tools and those who do not. It is the divide between those who have the capability to use AI tools productively and those who do not. This distinction — between access to a technology and the capability to translate that access into genuine improvements in human functioning — is the analytical contribution that Deaton's framework, drawing on the work of Amartya Sen, brings to the AI conversation with a precision that most technology commentary lacks.

Sen's capability approach, which has profoundly influenced Deaton's thinking about development, distinguishes between commodities and capabilities. A commodity is a thing — a computer, a software subscription, an internet connection. A capability is a functioning — the ability to participate in economic life, to express oneself creatively, to contribute to one's community, to live a life one has reason to value. The relationship between commodities and capabilities is mediated by what Sen calls conversion factors: the personal, social, and environmental conditions that determine whether a commodity translates into a genuine expansion of what a person can do and be.

A computer in the hands of a trained engineer with reliable electricity, high-speed connectivity, domain expertise, and institutional support translates into dramatic productive capability. The same computer in the hands of a person who lacks literacy, connectivity, domain knowledge, or economic security does not translate into the same capability, regardless of the computer's technical specifications. The commodity is identical. The capability it enables is radically different. And the difference is determined not by the commodity but by the conversion factors that mediate between the commodity and the capability.

Applied to AI, the capability approach reveals a distributional dynamic that declining cost curves alone cannot address. The prediction that AI tools will become cheap — perhaps even free — may prove correct. But cheapness addresses only the commodity dimension of access. It does not address the conversion factors: the education that provides the domain knowledge necessary to use the tools productively, the infrastructure that provides the physical substrate for AI-augmented work, the institutional context that provides the organizational support for adaptation, and the economic security that provides the freedom to invest in learning.

The evidence on conversion factors is extensive and, in Deaton's assessment, unambiguous. Consider education. The AI escape as documented in The Orange Pill depends fundamentally on the user's capacity to evaluate, direct, and build upon the AI's output. This capacity is not a generic skill that can be taught in a weekend workshop. It is domain expertise — the deep knowledge of a specific field that enables a practitioner to distinguish between plausible and true, between competent and excellent, between output that solves the right problem and output that solves the wrong one fluently. The engineer who directs an AI coding assistant draws on years of accumulated knowledge about systems architecture, failure modes, performance constraints, and the thousand small decisions that separate a prototype from a product. Without this accumulated knowledge, the AI tool produces output that may look impressive but that the user cannot evaluate, refine, or integrate into productive work.

This creates what might be called the amplification paradox: AI tools amplify existing capability, which means they benefit most the populations that already possess the most capability. The well-educated knowledge worker gains a dramatic productivity boost because she has the domain knowledge that the tool amplifies. The poorly educated worker gains a smaller boost — or no boost at all — because she lacks the knowledge base that the amplification requires. The technology is, in a precise empirical sense, anti-equalizing in its educational effects: it widens the gap between the well-educated and the poorly educated rather than narrowing it.

The amplification paradox extends to every conversion factor. Reliable infrastructure amplifies the productivity of AI-augmented work; unreliable infrastructure constrains it. Strong institutional support facilitates adaptation; weak institutional support impedes it. Economic security enables the risk-taking that adaptation requires; economic precarity prevents it. In each case, the conversion factor operates as a multiplier: populations with strong conversion factors capture a disproportionate share of the AI productivity gain, while populations with weak conversion factors capture little or none.

The cumulative effect of these multipliers is a form of inequality that is qualitatively different from the income inequalities that previous technological transitions produced. Previous inequalities were primarily inequalities of income and wealth — differences in the material resources that individuals possessed. The AI-driven inequality is an inequality of capability — a difference in the range and quality of what individuals can do and become. This distinction matters because the remedies are different. An inequality of income can, in principle, be addressed by redistribution: by transferring resources from those who have more to those who have less. An inequality of capability cannot be addressed by redistribution alone, because capability depends on conversion factors — education, institutions, infrastructure, health — that cannot be redistributed in the way that income can. They must be built. And building them takes decades.

The temporal mismatch between the speed of the AI transition and the time required to build conversion factors is the central distributional challenge of the current moment. The transition is creating capability differentials in months. The conversion factors that could narrow those differentials require years or decades of sustained investment. During the intervening period, the populations that lack the conversion factors fall further behind, and the falling-behind itself reduces the probability that the conversion factors will be built in time — because the political attention and the institutional resources tend to flow toward the populations that are already succeeding rather than toward the populations that are falling behind.

Health, as a conversion factor, operates through mechanisms that Deaton's research has documented with particular thoroughness. The AI-augmented work process demands sustained cognitive effort — the kind of intensive, iterative dialogue between human and machine that The Orange Pill describes as the foundation of productive collaboration. This cognitive effort draws on reserves of attention, working memory, and executive function that are directly affected by health status. Chronic malnutrition, parasitic infection, untreated mental health conditions, environmental stressors, and the cumulative effects of poverty on neurological development all reduce the cognitive capacity that productive AI collaboration requires. The populations with the worst health outcomes are, disproportionately, the populations furthest from the AI escape — and their health status is itself a barrier to the escape that declining tool costs cannot address.

Language operates as a conversion factor with particular political sensitivity. The most capable AI systems perform best in English and in a handful of other widely spoken languages. Their performance in the thousands of languages spoken by populations in Africa, South Asia, and indigenous communities worldwide is significantly inferior. This is not a temporary limitation that will resolve through normal technological improvement. It is a consequence of the training data available for these systems, which reflects the existing distribution of digital content, which is itself a reflection of the existing distribution of economic and institutional power. The populations whose languages are underrepresented in AI training data are, for the most part, the same populations that have been underserved by every other dimension of the global technological infrastructure.

The capability gap, properly understood, reframes the AI policy challenge. The challenge is not merely to make AI tools cheaper and more widely available — though this is necessary. The challenge is to build the conversion factors that translate access into capability: to invest in the education, the infrastructure, the health systems, the institutional frameworks, and the economic security that enable populations to use AI tools productively. This investment requires resources, political will, and sustained commitment over timescales that exceed the attention span of most political systems and most technology firms.

In the preface to the IFS Deaton Review on inequality, Deaton wrote that economists "have also become more concerned about unfair practices in the economy, about monopoly and monopsony and about corporate influence on politics. In the face of accelerating technical change — e.g. in artificial intelligence — they worry about just how the new technologies will be applied, who decides and to whose benefit." The capability gap is the empirical expression of this worry. The technologies are being applied by populations that already possess the conversion factors, to their benefit. The populations that lack the conversion factors are, by default, excluded from the application and its benefits. And the exclusion, if unaddressed, compounds over time into a form of inequality that is harder to reverse than any previous technology-driven divide.

---

Chapter 4: Escape Managed and Unmanaged

The Trivandrum training and the Lagos developer represent two modes of the AI escape — one managed, one unmanaged — and the contrast between them reveals, with the clarity of a controlled experiment, the role that institutional support plays in determining whether technological capability translates into genuine improvements in human functioning.

The Trivandrum case, as documented in The Orange Pill, is a case of managed escape. Twenty engineers, employed by a technology firm, received training in AI-powered development tools under conditions deliberately constructed to support adaptation. The employer provided the tools, maintained employment during the learning period, offered management guidance, and created an organizational culture in which experimentation was encouraged. The result was a genuine transformation in the team's capabilities: engineers who had previously been confined to narrow technical specializations discovered they could build across domains, producing work that would previously have required teams of specialists.

Read through Deaton's distributional lens, the Trivandrum case reveals not just what the engineers gained but the specific conditions that made the gains possible. Stable employment meant the engineers could invest cognitive resources in learning without the distraction of economic anxiety. Employer-provided tools meant the cost of access was absorbed by the organization rather than borne by the individual. Management support meant the adaptation was guided rather than left to individual initiative. Organizational culture meant experimentation was rewarded rather than punished. And the educational foundation — years of engineering training and professional experience — meant the engineers possessed the domain knowledge that the AI tools amplified.

Remove any one of these conditions, and the outcome changes. Remove stable employment, and the engineer cannot afford the temporary productivity decline that learning requires. Remove employer-provided tools, and the cost of access may exceed what the individual can bear. Remove management support, and the adaptation proceeds unevenly, with some engineers thriving and others struggling without guidance. Remove the educational foundation, and the tools produce output the engineer cannot evaluate, direct, or integrate.

The Trivandrum case also reveals distributional dynamics within the escaping group that larger-scale analyses tend to miss. Even among these relatively homogeneous engineers — all employed by the same firm, all working with the same tools, all receiving the same management support — the adaptation proceeded at different rates. The Orange Pill describes a senior engineer who oscillated between excitement and terror for the first two days before finding his footing, and a woman who had never written frontend code building a complete user-facing feature within forty-eight hours. The variation within the group illustrates a principle that Deaton's work has documented at every scale of analysis: even under favorable conditions, the distribution of outcomes is uneven, and the unevenness compounds over time as early adapters develop facility that late adapters struggle to match.

The within-group variation also illuminates a finding that Deaton would consider analytically significant: the more capable the individual before the AI tools arrived, the more capable they became after. The senior engineer's decades of architectural knowledge became more valuable, not less, because the tools removed the implementation labor that had consumed most of his time and revealed the judgment layer — the ability to decide what should be built, to anticipate failure modes, to evaluate trade-offs — that had been masked by mechanical work. The tools did not equalize the engineers. They amplified existing differences in expertise, judgment, and creative capacity.

Now consider the Lagos developer — the figure invoked in The Orange Pill as evidence that AI tools can bridge the gap between advantaged and disadvantaged populations. The claim is that a developer in Lagos can now access the same coding leverage as an engineer at Google, and the claim contains genuine truth. The AI tool is, in principle, available to both. The capability it provides — the ability to produce working software through natural-language conversation — does not discriminate by geography.

But the claim, subjected to Deaton's distributional analysis, reveals the gap between commodity access and capability functioning. The Lagos developer operates in an environment where the conversion factors are weaker than in Trivandrum — not marginally weaker but structurally different. Nigeria's electricity supply is among the most unreliable in Africa; the World Bank estimates that Nigerian firms experience power outages on average for over thirty days per year. Internet connectivity in Lagos, while improving, remains expensive relative to income and intermittent relative to the demands of sustained AI-augmented work. The institutional ecosystem — the professional networks, the venture capital infrastructure, the payment systems, the legal frameworks for contracting and intellectual property — provides fewer pathways for translating improved capability into improved economic outcomes.

The Lagos developer may possess the capability to produce software of quality comparable to that produced by a Silicon Valley developer. The AI tool may genuinely bridge the skill gap. But the functioning she derives from this capability — the income, the career trajectory, the economic security, the social status — depends on the institutional context in which she operates. Can she access markets that value her output? Can she demonstrate competence to employers who cannot verify her credentials? Can she navigate international payment systems? Can she participate in the professional networks that provide mentorship and access to opportunities? Each question identifies a conversion factor that the AI tool, by itself, does not address.

This is what Deaton would call unmanaged escape — escape achieved through individual initiative and tool access but without the institutional support that determines whether capability translates into sustained improvement. The unmanaged escape is genuine. The Lagos developer really does gain new abilities. But the escape is more precarious, more dependent on favorable circumstances, and more vulnerable to reversal than the managed escape in Trivandrum.

The concept of the marginal escapee illuminates the policy implications. The marginal escapee possesses just enough of the prerequisites — enough education, enough connectivity, enough institutional connection — to begin using AI tools productively, but possesses them in a degree that is barely sufficient. Her escape is real but fragile, dependent on conditions that could shift, and operating with a margin for error that is thinner than the margin enjoyed by her counterparts in more favorable institutional contexts.

The marginal escapee is the population for whom relatively modest institutional interventions could produce the largest gains. A more reliable internet connection, access to a workspace with stable electricity, membership in a professional network that provides mentorship and market access, a payment infrastructure that enables international contracting — these are investments that are modest relative to the cost of building basic infrastructure from scratch, and they may produce returns that are disproportionately large relative to their cost.

But Deaton's framework warns against the temptation to focus institutional attention on the marginal escapees at the expense of the populations further from the escape threshold. The marginal escapees are the populations for whom intervention is cheapest and most visible. Their success stories can be deployed as evidence that the technology is democratizing, that barriers are falling, that the transition is working. A policy that showcases marginal escapees while neglecting the populations that lack even basic prerequisites creates an illusion of progress that masks the persistence of deep structural exclusion.

The contrast between managed and unmanaged escape also reveals a dimension of the AI transition that has received insufficient attention: the role of organizational quality as a determinant of productive outcomes. Deaton's research on productivity in developing economies provides extensive evidence that organizational quality — the effectiveness of management, the coherence of processes, the degree to which organizational culture supports adaptation — explains a significant portion of the variation in productive outcomes between firms with similar resources. Two firms with identical equipment, identical workforces, and identical market conditions produce dramatically different results depending on the quality of their organizational processes.

The Trivandrum case is a case of high organizational quality applied to the AI transition. The firm's management understood the technology, invested in adoption, and created the conditions for adaptation. The result was a transformation that exceeded what any individual engineer could have achieved alone. The organizational context was not incidental to the escape. It was constitutive of it.

Most workers in the global economy do not have access to organizations of comparable quality. The self-employed worker, the employee of a small firm in a developing country, the freelancer navigating the gig economy — these workers face the AI transition without the organizational support that made the Trivandrum escape possible. Programs that train individual workers to use AI tools without addressing the organizational context in which the tools will be deployed are likely to produce disappointing results, because the organizational context determines whether individual capability translates into productive output. The most effective interventions, Deaton's research suggests, address both dimensions simultaneously — developing individual capabilities while strengthening the organizational environments in which those capabilities are exercised.

The managed-unmanaged distinction is not a binary but a spectrum, and most of the world's workers are located closer to the unmanaged end than to the managed one. The institutional challenge of the AI transition is to move a larger share of the world's workers toward the managed end of the spectrum — to create the conditions of stable employment, organizational support, educational preparation, and infrastructure access that the Trivandrum engineers enjoyed and that the Lagos developer, for the most part, did not. Creating these conditions at scale requires the engagement of governments, international institutions, firms, and civil society in a coordinated effort that matches the magnitude of the transition itself. The alternative — leaving the escape to individual initiative and market forces — will produce outcomes that the historical record, assembled across Deaton's career, predicts with depressing reliability: concentrated gains for those who are already advantaged, and a widening gap for everyone else.

Chapter 5: What the Numbers Miss

The central empirical claim of The Orange Pill — that AI tools produce a twenty-fold improvement in individual productivity for certain categories of knowledge work — is the kind of number that Deaton's career has trained him to treat with both interest and suspicion. Interest, because if the claim is even approximately correct, it represents a productivity improvement of a magnitude that has few precedents in the history of technology. Suspicion, because Deaton's decades of measuring human welfare have taught him that impressive numbers rarely mean what they appear to mean, and that the distance between what is measured and what matters is almost always wider than the measurers acknowledge.

Productivity is output divided by input. When the claim is a twenty-fold improvement, the assertion is that a single worker equipped with AI tools produces, in a given period, twenty times as much output as the same worker without the tools. The arithmetic is seductive. The question is what, precisely, is being counted — and what is not.

The first problem is output quality. Productivity measures count volume without distinguishing between output that is excellent and output that is merely adequate, between work that solves a genuine problem and work that addresses a manufactured one, between artifacts that endure and artifacts that are consumed and forgotten. The AI-augmented worker who produces twenty times as much code may be producing code that is twenty times the volume but not twenty times the value. Some of the additional output may be redundant. Some may address problems that did not need solving. Some may solve the right problems in ways that create new ones. The productivity measure counts all of this equally, and the result is a number that overstates the welfare gain in ways that are invisible from inside the measurement.

Deaton's research on the relationship between economic output and human welfare has documented this divergence with particular force in the context of American economic development. In the decades following the Second World War, measured productivity in the United States increased steadily while measures of subjective well-being, social cohesion, and — eventually — life expectancy for specific populations diverged from the productivity trend. The lesson of this divergence is direct: productivity is a necessary but insufficient condition for welfare improvement, and increases in productivity can, under specific institutional conditions, coexist with stagnation or decline in the dimensions of human life that people actually care about.

The second problem concerns the boundary of the calculation. The twenty-fold improvement is bounded by the individual worker and the task at hand. But the welfare implications extend to the broader organizational and economic context. The worker who produces twenty times as much output may displace the work of colleagues who were previously employed to produce it. The net effect on organizational productivity may be positive, but the distributional effect — the consequences for the displaced — is not captured by the individual metric. A twenty-fold improvement for one worker accompanied by the elimination of employment for nineteen others is, in aggregate, a wash — and the aggregate conceals the fact that the nineteen experience the transition as catastrophe while the one experiences it as liberation.

The Orange Pill argues that the twenty-fold improvement will not eliminate jobs but will enable individuals to do work that previously required teams, freeing team members to pursue work of higher value. This is a plausible scenario, and Deaton does not dismiss it. But the historical record on the employment effects of productivity-enhancing technologies is decidedly mixed. The mechanization of agriculture eliminated agricultural employment on a massive scale, and the displaced workers did not uniformly find alternative employment of comparable quality. The computerization of back-office functions eliminated clerical employment, and the transition was marked by significant hardship for the displaced even as aggregate productivity increased. In some cases — the personal computer, the internet — productivity improvements created more jobs than they eliminated. But the new jobs were different in character and distribution from the old ones, concentrated in different geographies and requiring different skills, and the transition period was painful for the specific populations that bore its costs.

The third problem is the relationship between output volume and output value. When AI tools enable a twenty-fold increase in the supply of a particular category of output — software, analysis, content, design — the economic value of each unit may decline because the increased supply reduces the scarcity that sustained the previous price. This is the paradox of productivity: the observation that increases in output do not always translate into proportional increases in income, because the market price of output tends toward the marginal cost of production, and AI dramatically reduces that marginal cost. The worker who produces twenty times as much may earn roughly the same — or less — as the consumer captures the surplus through lower prices.

The distributional implications depend on who consumes the output. If the output is consumed primarily by wealthy populations, the price reduction represents a transfer from workers to consumers who were already advantaged. If consumed more broadly, the benefit is more widely distributed. But the distributional question is not answered by the productivity number. It requires the kind of disaggregated analysis that Deaton has spent his career insisting upon and that most technology commentary declines to perform.

The fourth measurement problem may be the most fundamental. The productivity numbers miss entirely what The Orange Pill describes as one of the most important features of AI-augmented work: the transformation of the work experience itself. The book documents the experience of flow — the creative engagement, the exhilaration, the sense of capability expansion — that accompanies productive AI collaboration. These dimensions of work are not captured by any productivity metric, and they may be among the most consequential features of the transition for human welfare.

Deaton's research, building on the foundational work with Daniel Kahneman on well-being measurement, has consistently found that the quality of work experience — the degree of autonomy, the sense of purpose, the opportunity for mastery — is at least as important for subjective well-being as the income that work provides. The famous Kahneman-Deaton finding that income improves life evaluation but not emotional well-being beyond a threshold suggests that the experiential dimension of the AI transition matters in ways that income and productivity statistics cannot capture.

If AI tools transform the work experience in ways that genuinely increase autonomy, creativity, and the sense of mastery, the welfare gain may be substantially larger than the productivity measure suggests. But if the tools transform the experience in ways that reduce these qualities — by deskilling workers, eroding the need for judgment, creating a sense of displacement and obsolescence — the welfare loss may offset the productivity gain. The Berkeley study cited in The Orange Pill, which found that AI-augmented work intensified rather than reduced the experience of labor, suggests that both dynamics are operating simultaneously, in different populations and different contexts, and that the net effect depends on conditions that the aggregate number does not reveal.

The measurement problems do not invalidate the productivity claims. They qualify them. The twenty-fold improvement is real for certain workers, in certain contexts, performing certain tasks. The welfare implications are more complex, more uncertain, and more unevenly distributed than the number alone suggests. Deaton's insistence on looking beyond the aggregate to the distribution is not pedantry. It is the difference between a policy that maximizes a metric and a policy that improves human lives — and the two are not the same thing nearly as often as the metric's proponents assume.

For the populations closest to the AI frontier, the productivity gains are both real and welfare-enhancing. For the populations furthest from the frontier, the gains are abstract — a number reported in someone else's accounting, bearing no relationship to their own experience of work. The number is important. The distribution is more important. And the distribution is invisible to anyone who mistakes the aggregate for the whole.

---

Chapter 6: The Death Cross and Its Geography

The Orange Pill introduces a concept borrowed from financial analysis: the death cross — the point at which an existing workflow becomes less productive than its AI-augmented alternative. When the death cross is reached, the economic logic of adoption becomes inexorable. The cost of maintaining the old approach exceeds the cost of adopting the new one, and competitive pressure intensifies to the point where non-adoption is economically unsustainable. The term is borrowed from stock-chart analysis, where a death cross signals that momentum has shifted from bullish to bearish, and the metaphor is apt: the lines have crossed, and the old order is on the wrong side.

Deaton's contribution is to map the death cross onto the geography of global development — to examine how its timing and consequences vary across populations, nations, and institutional contexts. The death cross is not a single event. It is a wave, moving across the global economy at a speed and with a distribution shaped by the same structural factors that determine the order of escape.

The wave arrives first where the economic case for AI adoption is most compelling: in high-wage knowledge-work occupations in wealthy nations, where the cost of human labor is high relative to the cost of AI tools, where the infrastructure for deployment is mature, and where the institutional environment supports rapid change. A law firm in New York paying associates several hundred thousand dollars annually faces the death cross earlier than a law firm in Nairobi paying associates a fraction of that amount, because the economic return on AI investment is proportional to the labor cost it displaces.

This ordering creates a distributional dynamic that is counterintuitive and, for the populations most affected, dangerous. The populations that face the death cross first are, in general, the populations best equipped to manage it. They have education, institutional support, alternative employment options, and the economic security to navigate the transition. The populations that face the death cross later are less equipped — and the later arrival does not provide additional time to prepare. It provides additional time during which the competitive landscape is reshaped by the adaptation of the early movers, making the eventual disruption more severe.

The developer who confronts the death cross in 2028 faces a competitive environment already transformed by the adaptation of developers in Silicon Valley and Bangalore in 2025 and 2026. The standards of output have risen. The price that clients will pay for work that can be AI-augmented has declined. The competitive advantage that early adaptation provided has been captured and compounded by the early movers. The late arrival of the death cross does not soften the blow. It hardens the ground on which the latecomer must land.

The most consequential geographic dimension of the death cross concerns the outsourcing industry — the sector that has served as an escalator to middle-class employment for millions of workers in India, the Philippines, Eastern Europe, and other middle-income nations. The outsourcing model rests on a simple economic proposition: cognitive labor is cheaper in certain geographies than in others, and firms in high-wage nations can reduce costs by contracting with workers in low-wage nations to perform knowledge work remotely. The proposition has been valid for decades, supporting an industry that employs millions and has been a significant driver of economic development, urbanization, and middle-class formation in the participating nations.

AI disrupts this proposition at its foundation. When the cost of AI-augmented labor in a wealthy nation falls below the cost of outsourced labor in a middle-income nation — when a single developer in San Francisco equipped with AI tools can produce the same output as a team in Bangalore at comparable or lower total cost — the economic rationale for outsourcing collapses. The arbitrage that sustained the industry disappears, not gradually but with the binary finality of a death cross.

This is not a hypothetical scenario. Early evidence suggests that firms in wealthy nations are already discovering that AI-augmented domestic workers can perform tasks previously outsourced, and the discovery is reshaping procurement decisions in real time. The Indian IT industry, which employs several million workers directly and supports a much larger ecosystem of ancillary employment, faces a displacement risk of a magnitude that would have significant economic, social, and political consequences — consequences felt not only by the displaced workers but by the communities and institutions built around the industry over two decades of growth.

The concept that this analysis identifies as the most dangerous feature of the death cross geography is what might be called the premature death cross: the arrival of the death cross before a population has developed the capability to adopt the AI alternative. In this scenario, the old workflow becomes economically uncompetitive, but the new workflow is not yet accessible. The worker finds herself in a gap — her existing skills devalued by the competitive pressure of AI-augmented workers elsewhere, but the AI tools and the conditions for productive AI collaboration not yet available to her.

Before the death cross, the worker performing her tasks without AI was operating at a level the market accepted. After the death cross, the same level of performance is no longer competitive. She faces not merely the challenge of adapting but the challenge of adapting without the tools, the training, and the institutional support that adaptation requires. The gap between the obsolescence of the old and the inaccessibility of the new is a period of capability deprivation that is qualitatively different from the pre-transition condition — a period in which the worker is worse off than before the transition began.

Deaton's research on the consequences of sudden competitive shocks provides a framework for understanding the human costs of this gap. The experience of economies exposed to rapid trade liberalization — domestic manufacturers suddenly competing with international imports — suggests that adjustment costs are real, significant, and persistent. Workers displaced from competitive industries do not automatically find employment of comparable value. Many experience extended unemployment or underemployment. Many accept work at lower wages and in less favorable conditions. The aggregate statistics may show improvement — GDP growth, total employment — even as specific populations experience sustained hardship.

The geography of the death cross also interacts with the geography of institutional capacity in ways that compound the challenge. The nations facing the death cross in their outsourcing industries are nations that have built substantial portions of their economic development strategies around the outsourcing model. Displacement requires not merely individual adaptation but institutional reconstruction — new economic strategies, new educational investments, new infrastructure priorities — and the capacity for institutional reconstruction varies dramatically across nations. India has the institutional depth and the human capital to reconstruct. Some smaller economies that have specialized in outsourcing may not.

The policy response that Deaton's framework suggests is not to prevent the death cross — it cannot be prevented, and the attempt to prevent it would sacrifice the genuine productivity gains that AI provides. The response is to manage the transition: to invest in the capabilities that complement AI rather than compete with it, to build social safety nets that provide economic security during the adjustment period, to develop new economic strategies that position affected economies for the opportunities the AI transition creates as well as the disruptions it imposes, and to do all of this at a speed that matches the pace of the technological change. The death cross waits for no institution, and the institutions that respond too slowly will find their populations stranded in the gap between the old and the new — bearing the costs of the transition without access to its benefits.

---

Chapter 7: The Foundations

The conditions that determine whether populations escape from deprivation are never purely economic. They are intertwined with the two dimensions of human development that Deaton has studied most persistently across his career: health and education. The relationship between these dimensions and economic capability is not merely correlational. It is causal, operating through mechanisms that Deaton has documented with the precision that careful empirical work demands and that casual invocations of "human capital" routinely obscure.

The AI escape requires sustained cognitive effort of a specific kind — the iterative, attention-intensive dialogue between human and machine that constitutes productive AI collaboration. This cognitive effort draws on reserves of attention, working memory, and executive function that are directly and measurably affected by health status. The evidence is extensive. Chronic subclinical malnutrition — not the dramatic malnutrition that produces visible wasting but the mild, persistent inadequacy that affects hundreds of millions of people in low-income countries — reduces cognitive function in ways that are subtle enough to escape casual observation but significant enough to affect productive capacity. Robert Fogel's research, which Deaton has built upon, demonstrated that improvements in nutrition accounted for a substantial share of the productivity gains of the industrial revolution — not because better-fed workers were stronger (though they were) but because better-fed workers could think more clearly, sustain attention longer, and adapt to new tasks more readily.

Parasitic infection, endemic in much of the tropical world, imposes a cognitive tax that is invisible in economic statistics but real in the lived experience of the affected populations. A child who grows up with chronic helminth infection develops a brain less capable of the sustained concentration that complex knowledge work requires — not dramatically less capable, but measurably so, in ways that compound over a lifetime of work. Untreated depression, which affects populations worldwide but is most severely undertreated in low-income countries, reduces executive function, working memory, and the capacity for sustained creative effort. Environmental stressors — air pollution, noise, the chronic anxiety of economic precarity — further degrade the cognitive substrate on which productive AI collaboration depends.

The implication is direct. The AI escape, for populations with compromised health, is constrained not by the availability or the cost of the technology but by the cognitive capacity to use it. This constraint cannot be addressed by making tools cheaper, improving connectivity, or providing training. It can only be addressed by improving the health conditions that determine cognitive capacity — and improving health conditions requires the kind of sustained, multi-dimensional investment in healthcare systems, nutrition programs, environmental remediation, and the broader social determinants of health that Deaton's work has identified as the foundations of human capability development.

The educational dimension operates through mechanisms that are equally fundamental and equally resistant to quick fixes. The AI escape depends not on generic computer skills but on domain expertise — the deep knowledge of a specific field that enables the practitioner to direct AI tools toward productive ends, to evaluate their output, and to integrate their contributions into coherent work. The engineer in Trivandrum benefited from AI tools because she had years of engineering education and professional experience that provided context for the tools' output. Without this domain knowledge, the tools produce results that the user cannot assess. The output may look impressive — syntactically correct code, grammatically fluent prose, statistically plausible analysis — but the user lacks the expertise to determine whether it is substantively correct, and the inability to evaluate output makes the collaboration unproductive at best and dangerous at worst.

This creates the amplification paradox in its educational dimension. AI tools amplify existing knowledge. Populations with deep educational foundations gain the most from the amplification. Populations with weak foundations gain less. The technology widens the gap between the well-educated and the poorly educated because it provides leverage that is proportional to the knowledge base it has to work with. A free AI tool in the hands of someone without domain expertise is not democratization. It is the appearance of democratization — access without capability, commodity without functioning.

The educational challenge is compounded by a question that Deaton considers newly urgent: what kind of education is appropriate for a world in which AI performs many of the cognitive tasks that education has traditionally prepared people to perform? If AI tools can write competent prose, perform calculations, analyze data, and generate research summaries, then the traditional emphasis on these skills may need fundamental reconsideration. The capacities that AI cannot replicate — judgment, ethical reasoning, creative synthesis, the domain expertise that enables productive direction of AI-augmented work, the interpersonal skills that sustain organizational collaboration — may be the capacities that education should prioritize.

But educational systems are among the most institutionally conservative structures in any society. Curriculum reform, teacher training, assessment redesign, and resource reallocation operate on timescales measured in years and decades. The AI transition is creating demand for new educational outputs on timescales measured in months. The lag between what the economy demands and what the educational system produces is growing, not shrinking, and the populations most affected by the lag are the populations in countries where educational systems are weakest — where the baseline curriculum is furthest from what the AI economy requires and where the institutional capacity for reform is most constrained.

If educational systems in wealthy nations adapt rapidly to the changed demands — emphasizing judgment, creativity, and domain expertise — while educational systems in low-income nations continue to emphasize procedural skills that AI is replacing, the educational gap will widen in a new dimension. The distributional consequences of the AI transition will be compounded by a misalignment between what education produces and what the economy rewards, and the misalignment will be most severe in the countries that can least afford it.

Deaton's framework treats health and education not as complementary to the economic dimensions of the AI transition but as foundational to them. The economic escape celebrated in The Orange Pill — the dramatic expansion of individual productive capability — rests on a foundation of health and education that determines whether the escape is possible, for whom it is possible, and how durable the escape proves to be. A policy that addresses the economic dimensions — infrastructure investment, tool access, cost reduction — without addressing the health and educational foundations will produce results that are narrower, more concentrated, and less durable than anticipated. The foundation must be built alongside the structure, and the building requires the kind of sustained, cross-generational investment in human capability that Deaton's career has identified as the most reliable predictor of developmental success — and the most consistently underfunded priority in development policy.

---

Chapter 8: Building Dams

Every major technological transition generates a surplus — the difference between the value the innovation creates and the cost of producing it. The question that determines the welfare consequences of the innovation is never the size of the surplus but its distribution. Deaton has devoted a substantial portion of his career to precisely this question, and the AI transition provides a new context in which the question has both unprecedented urgency and depressingly familiar structural dynamics.

The AI surplus is, by any reasonable estimate, enormous. If the productivity improvements documented across the early adoption wave are representative of what AI can produce across the knowledge economy, the total surplus will be measured in trillions of dollars annually. A surplus of this magnitude could transform material conditions for billions of people if distributed broadly. It could also create concentrations of wealth and power exceeding anything in the historical record if distributed narrowly. The technology does not determine which outcome obtains. The institutions do.

The current institutional arrangement favors concentration. The firms that develop AI technologies are, for the most part, headquartered in a small number of wealthy nations and accountable primarily to their shareholders. The workers most complementary to the technology — highly educated knowledge workers who can direct, evaluate, and build upon AI output — are already at the top of the global income distribution. The platforms through which AI-augmented products reach consumers are themselves dominated by a small number of firms that capture a significant share of the value generated by products distributed through their infrastructure. The geographic concentration of AI development means that the economic surplus — profits, wages, tax revenue, ancillary economic activity — flows disproportionately to locations that are already wealthy.

In his 2024 IMF essay, Deaton identified the structural conditions that produce this concentration: "pervasive economies of scale more powerful than older industries," in which "a small number of firms have significant market power" and "pinpointing value creation is next to impossible." The observation applies with particular force to AI, where the economics of model training create natural monopoly dynamics — the cost of training a frontier model is so high that only a handful of firms can afford to do so, and the trained model, once it exists, can be deployed at near-zero marginal cost to serve billions of users. The result is a market structure in which a small number of firms capture an outsized share of the surplus, not because they are uniquely virtuous or efficient but because the economics of the technology favor concentration.

The distribution of the surplus within nations is shaped by tax systems that Deaton has studied extensively. The capacity of governments to provide public goods — education, healthcare, infrastructure, social protection — depends on the capacity of tax systems to capture a share of the surplus generated by economic activity. If the AI surplus is captured primarily by firms and high-income individuals, and if tax systems fail to capture an adequate share for public investment, then the distributional consequences are compounded by declining provision of the public goods that the most disadvantaged populations depend upon. The evidence from recent decades is not encouraging: the effective tax rates on the most profitable technology firms have declined in most wealthy nations even as the firms' market power has increased, and the political dynamics that produced this decline show no sign of reversal.

The international dimension is equally important. The firms that develop AI are headquartered in wealthy nations, and the tax revenue from their activities flows to those nations' treasuries. The nations that consume AI technologies — where the tools are used by workers and firms but where the developers are not headquartered — capture a smaller share. The nations that lack the infrastructure and institutional capacity to adopt AI capture almost none. The result is an international distribution that favors the already wealthy and disfavors the already disadvantaged, following a pattern that Deaton has documented across every previous technological transition and that the current international institutional arrangements do nothing to correct.

Building institutional structures to direct the surplus more broadly requires interventions across multiple dimensions. The first is investment in digital infrastructure for underserved populations — not merely connectivity but the complete ecosystem that AI-augmented work requires: reliable electricity, affordable devices, high-speed internet, and the maintenance systems that keep infrastructure functional over time. The investment is large but not unprecedented. The global effort to expand vaccination access, which Deaton has studied extensively, required infrastructure investments of comparable magnitude and produced returns that more than justified the cost. The case for digital infrastructure investment rests on similar logic: the cost is high, the returns are higher, and the populations that benefit most are the populations that the market would serve last.

The second intervention is educational reform at a pace and scale that matches the technological transition. The reform must address content — emphasizing judgment, creativity, and domain expertise over procedural skills that AI is replacing. It must address method — preparing students for the iterative, collaborative mode of work that AI-augmented production requires. And it must address access — expanding quality education to populations currently excluded. The tools of the transition itself may provide a means of acceleration: AI-powered educational technologies can personalize instruction, expand access to knowledge, and provide feedback at scale in ways that traditional educational delivery cannot. But the deployment of these tools requires the same institutional capacity, infrastructure, and human capital that the broader AI transition requires — creating a circularity that can only be broken by deliberate public investment.

The third intervention is labor market policy scaled to the magnitude of the structural change. Retraining programs, income support during transition, occupational mobility assistance, and social safety nets must be designed for structural displacement rather than cyclical unemployment. The existing mechanisms are, in most nations, inadequate — underfunded, poorly targeted, and designed for a labor market that no longer exists. Deaton's endorsement of Acemoglu and Johnson's argument that unions must be at the table for AI decisions reflects a broader point about the governance of the transition: the populations that bear the costs of technological change must have institutional voice in the decisions that determine how the change unfolds.

The fourth intervention concerns the governance of AI development itself. The decisions made by the firms that develop AI systems — decisions about capabilities, pricing, language support, safety standards, terms of access — have distributional consequences that are currently determined almost entirely by the firms themselves, with limited external accountability. Deaton's framework suggests that governance must be broadened to include the perspectives of affected populations — not only the consumers who use the products but the workers displaced by them, the communities disrupted, and the nations excluded from the benefits.

The fifth intervention, and the one that Deaton's framework identifies as most structurally important, is international coordination. The AI transition is a global phenomenon with global distributional consequences, and the mechanisms for managing those consequences — development assistance, trade agreements, technology transfer, capacity building — must be adapted to the specific requirements of the transition. This means expanding development assistance to include digital infrastructure and educational reform. It means revising trade frameworks to account for AI's effects on international labor markets. It means establishing mechanisms for technology transfer that prevent developing nations from becoming permanently dependent on the firms and nations that control AI development. And it means building institutional capacity in developing nations for effective AI governance.

These interventions are expensive, politically difficult, and institutionally complex. They require resources that many governments do not have, institutional capacity that many public sectors do not possess, and political will that the urgency of the situation demands but that democratic governance may not generate quickly enough. Deaton's career provides grounds for believing the interventions can work — the historical evidence demonstrates that institutional action shapes the distribution of technological benefits and that the distribution can be altered by deliberate policy. But the evidence also demonstrates that institutional action does not arise spontaneously. It is built by people who understand the distributional dynamics, who possess the political resources to act on that understanding, and who sustain the commitment over timescales that exceed the attention span of most political systems.

The AI transition can produce broadly shared improvement in human welfare. It can expand capability, accelerate the escape from deprivation, and democratize access to productive power in ways that no previous technology has achieved. These outcomes are possible. They are not default. They require that the surplus generated by the technology be directed toward the populations that need it most — through infrastructure, through education, through institutional support, through the sustained political commitment to distributional equity that Deaton's career has demonstrated to be the essential and perpetually underprovided ingredient of broadly shared progress. The surplus is being generated now. The question of its distribution is being answered now — by the institutional choices that are being made, or not made, in the specific window of time during which the trajectory of the transition can still be shaped. Deaton's work provides no guarantee that the choices will be made well. It provides only the evidence that they can be, and the insistence that they must be.

Chapter 9: The Great Escape, Revisited

In 2023, Angus Deaton published Economics in America, a book that read, in significant passages, like a reckoning. The economist who had spent decades inside the discipline's mainstream — contributing to its methods, advancing its empirical techniques, accepting its foundational assumptions about markets, efficiency, and the self-correcting character of economic competition — turned to face the discipline with a question that was less academic than moral: What did we miss, and what did the missing cost?

What the discipline missed, in Deaton's assessment, was power. The models that predicted efficient outcomes from competitive markets assumed that power was distributed in ways that made competition meaningful — that workers could move to better jobs, that consumers could choose among alternatives, that firms operated under constraints imposed by genuine rivalry. The reality, which the models obscured and which decades of evidence had made undeniable, was that power had concentrated in ways that rendered the competitive assumptions false. A small number of firms dominated entire industries. Employers in many labor markets faced workers with few alternatives. The political process, which was supposed to correct market failures through regulation, had itself been captured by the interests it was meant to regulate.

The AI transition unfolds in the institutional landscape that this concentration of power has produced, and Deaton's late-career reckoning provides the analytical frame for understanding why the default distribution of the AI surplus is concentration rather than diffusion. The technology is new. The institutional dynamics are not. They are the same dynamics that produced the outcomes Deaton spent his career measuring — the deaths of despair, the widening gap between the educated and the uneducated, the erosion of the institutional structures that had given working-class communities their coherence and their dignity.

The reckoning has a specific implication for AI that extends beyond the distributional analysis of the preceding chapters. It concerns the relationship between the technology and the political institutions that are supposed to govern it. Deaton's 2024 IMF essay noted that contemporary innovations in AI "have tremendous potential to promote prosperity, improve health and education, and address global challenges," while also observing that "many are concerned that these innovations could further endanger the environment, increase inequality, and lead to political polarization." The juxtaposition is deliberate. The potential and the concern coexist not because one is right and the other wrong but because the outcome depends on institutional choices that are being made under conditions of concentrated power.

The concentration of power in AI development is more extreme than in any previous technological transition. The cost of training a frontier language model — measured in hundreds of millions of dollars — creates barriers to entry that ensure the field is dominated by a small number of firms, most of them headquartered in the United States, most of them controlled by a small number of individuals, and most of them operating under governance structures that prioritize shareholder returns and organizational objectives over broader distributional concerns. These firms make decisions about what the technology can do, who can access it, how much it costs, which languages it supports, and what safety standards apply — decisions with distributional consequences that affect billions of people and that are made without meaningful input from the populations most affected.

Deaton's endorsement of the argument that unions need to be at the table for AI decisions is not a sentimental attachment to organized labor. It is a structural observation about the conditions under which technological transitions produce broadly distributed benefits. The historical evidence, which Deaton has examined across multiple transitions, demonstrates that the distribution of technological benefits is shaped by the balance of power between the interests that capture the surplus and the interests that demand its broader distribution. When worker organizations, consumer advocates, civil society groups, and democratic governments have meaningful influence over the deployment of new technologies, the distribution tends to be broader. When these countervailing institutions are weak or absent, the distribution tends to concentrate.

The countervailing institutions for AI are, at present, weak. Labor unions in most wealthy nations have declined in membership and political influence over the past four decades. Consumer advocacy organizations lack the technical expertise to engage meaningfully with AI policy. Civil society groups focused on technology governance are underfunded relative to the scale of the challenge. Democratic governments are constrained by the same concentration of corporate power that they are supposed to regulate, and the regulatory frameworks they have developed — the EU AI Act, the American executive orders, the emerging frameworks in other jurisdictions — address supply-side questions about what firms may build while leaving demand-side questions about what populations need largely unaddressed.

The institutional deficit creates a specific danger that Deaton's framework identifies with the clarity of a researcher who has studied exactly this pattern in other contexts: the danger that the early beneficiaries of the AI transition will capture the institutional mechanisms that govern the technology and shape them in ways that preserve their advantage. The firms and nations that lead in AI development are also the firms and nations with the greatest influence over the international institutions, standard-setting bodies, and regulatory frameworks that will determine how AI is governed globally. The populations that are most affected by these governance decisions — the workers displaced by AI, the communities disrupted, the nations excluded from the benefits — are the populations with the least influence over the decisions.

This is not a prediction of conspiracy. It is an observation about institutional dynamics that Deaton has documented across multiple contexts. The populations that benefit first from a new technology tend to capture the institutions that govern it, not through malice but through the normal operation of influence, access, and the concentration of expertise. The institutions, once captured, tend to distribute benefits in ways that favor the populations that captured them. The pattern is structural, not intentional, and it can be altered — but only by the deliberate construction of countervailing institutions that represent the interests of the populations that the default distribution leaves behind.

The construction of these countervailing institutions is the political challenge of the AI transition. It requires the kind of institutional innovation that Deaton's career has demonstrated to be possible but difficult: the development of new forms of worker representation that can engage with AI deployment decisions, new mechanisms for public participation in technology governance, new international agreements that ensure developing nations have voice in the decisions that affect their populations, and new forms of democratic accountability for the firms that control the most consequential technology of the current era.

The challenge is complicated by the speed of the transition. Previous institutional innovations — labor unions, regulatory agencies, international development organizations — developed over decades, through iterative processes of political struggle, institutional experimentation, and gradual adaptation. The AI transition is moving faster than these processes can operate. The institutional responses that the transition requires must be developed not over decades but over years, and the gap between the speed of the technology and the speed of institutional development is itself a distributional variable — the populations that suffer most from the gap are the populations that lack institutional voice.

The great escape is not finished. It has never been finished. Each generation faces its own version of the escape, its own distributional challenge, its own institutional demands. The AI transition is this generation's version, and its distributional stakes are measured in the same currency that Deaton has spent his career counting: years of life, children's futures, the capacity of communities to sustain the institutional structures that give individual lives their meaning and their dignity. The escape is real. The distribution is the question. And the question is being answered now, in the institutional choices that are being made — or not made — in the specific window of time during which the trajectory of the transition can still be shaped.

The evidence from Deaton's career supports a conditional conclusion. The condition is institutional. If the countervailing institutions are built — if the populations affected by the transition gain voice in the decisions that shape it, if the surplus is directed through deliberate policy toward the infrastructure, education, and social protection that broaden the escape, if the governance of AI development is expanded beyond the firms that control it to include the populations affected by it — then the AI transition can produce the most broadly shared expansion of human capability in the historical record. The potential is genuinely extraordinary.

If the institutions are not built — if the default distribution is allowed to operate unchecked, if the surplus concentrates among the already advantaged, if the governance remains in the hands of the firms and nations that develop the technology — then the transition will produce a new and wider divide, and the populations on the wrong side of the divide will bear costs that compound over generations. The technology will not have failed. The institutions will have failed. And the failure will be measured not in stock prices or productivity statistics but in the specific, irreducible currency of human lives that did not reach their potential because the structures that could have supported them were not built in time.

Deaton's career has been devoted to demonstrating that this choice is real — that the distribution of technological benefits is not determined by fate but by institutions, and that institutions are built by people who possess the understanding, the resources, and the political will to act. The AI transition is the latest and most consequential test of this proposition. The evidence supports cautious optimism. The caution is not timidity. It is the recognition that optimism untempered by distributional analysis has, throughout the history of technological change, served as cover for the concentration of gains among those who need them least. The optimism is genuine. The conditions for its realization are identifiable but not guaranteed. And the distance between identifiable and realized is the distance that institutions must cross — the distance that is, in the end, the measure of a society's commitment to the proposition that the escape should reach everyone.

---

Epilogue

The number that stays with me is not twenty.

Not the twenty-fold productivity multiplier I wrote about in The Orange Pill, the number that made my engineers' eyes widen in Trivandrum. Not the twenty engineers in that room discovering they could each do the work of a team. The number that stays with me, after spending months inside Angus Deaton's framework, is seventy to one.

That is the ratio Deaton documented between the per capita income of the richest nations and the poorest at the end of the twentieth century — a ratio that was five to one before the industrial revolution began. The great escape happened. Life expectancy doubled. Extreme poverty declined. And the gap widened from five to seventy. Not because the poorest got poorer. Because the escapees escaped so fast, and so far, that the distance between them and everyone else became a chasm.

I built my career on the other side of that chasm. Everything I described in The Orange Pill — the exhilaration, the creative flow, the collapse of the distance between imagination and execution — happened inside a set of conditions so favorable that I had stopped noticing them. Reliable electricity. Fast internet. A team with engineering degrees. An employer willing to invest in their adaptation. English as the operating language. These were not features of the technology. They were features of my position in the global economy. The technology amplified what was already there. Including the advantages.

Deaton's framework forced me to see something I had been looking past. The developer in Lagos I wrote about — the one I said could now access the same coding leverage as an engineer at Google — can access the tool. She cannot access the conversion factors that make the tool transformative. The stable electricity. The connectivity that does not drop during a rainstorm. The professional network that turns a working prototype into a career. The market that pays for her output at the rate it pays for identical output from San Francisco. The commodity is approaching parity. The capability is not.

That distinction — between the tool and the conditions that make the tool work — is the hardest thing Deaton's thinking has given me. Because I want the tool to be enough. I want the story to be: the barriers fall, the technology democratizes, the floor rises for everyone. And some of that story is true. But the part that is true is the part that operates inside institutions strong enough to translate access into capability. The part that is not yet true is the part that operates outside those institutions — which is where most of the world's population lives.

I keep returning to a sentence from Deaton's 2024 IMF essay: "the direction of technical change has always depended on who has the power to decide." Not what the technology can do. Who decides what it does, and for whom. The twenty-fold multiplier is real. The question of who captures the multiplication — that is not a technology question. It is an institutional question, a political question, and ultimately a moral one.

I wrote The Orange Pill from inside the escape. This book forced me to look at the escape from outside — to see its shape not as an expanding circle of capability but as a widening distance between those who are inside it and those who are not yet. The widening is not inevitable. But it is the default. And defaults, as every builder knows, are what you get when you do not make a deliberate choice.

The dam-building I wrote about — the institutional structures that channel the river toward life rather than away from it — is more urgent than I understood when I wrote it. It is not a nice-to-have. It is the variable that determines whether the AI transition becomes the most broadly shared expansion of human capability in history or a repetition of the pattern that Deaton spent his career documenting: genuine progress in the aggregate, genuine suffering in the distribution.

I do not have Deaton's patience with data. I am a builder, not a measurer. But I know enough now to know that what I build matters less than who it reaches. And who it reaches depends on structures that builders alone cannot construct. It depends on institutions. On policy. On the political choices of societies that must decide, in the specific window of time during which the trajectory can still be shaped, whether the escape will be managed or unmanaged, broad or narrow, a great escape or a great divergence.

The aggregate is not enough. That is the sentence I will carry from this book. The aggregate — the productivity gains, the capability expansion, the creative liberation — is real and it is not enough. The distribution is where the justice lives. Or doesn't.

— Edo Segal