Sources and References

Primary texts — Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press, 2018. / Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press, 2022. / Brynjolfsson, Erik and Andrew McAfee. The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies. W. W. Norton, 2014.

Supplementary texts — Acemoglu, Daron and Pascual Restrepo. “The Wrong Kind of AI? Artificial Intelligence and the Future of Labour Demand.” Cambridge Journal of Regions, Economy and Society, 2020. / Autor, David. “Why Are There Still So Many Jobs? The History and Future of Workplace Automation.” Journal of Economic Perspectives, 2015. / Korinek, Anton and Joseph E. Stiglitz. “Artificial Intelligence and Its Implications for Income Distribution and Unemployment.” NBER Working Paper 24174, 2017. / Parker, Geoffrey G., Marshall W. Van Alstyne, and Sangeet Paul Choudary. Platform Revolution: How Networked Markets Are Transforming the Economy. W. W. Norton, 2016. / Couldry, Nick and Ulises A. Mejias. The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford University Press, 2019.

Key papers — Acemoglu, Daron and Pascual Restrepo. “Robots and Jobs: Evidence from US Labor Markets.” Journal of Political Economy, 2020. / Autor, David, Frank Levy, and Richard Murnane. “The Skill Content of Recent Technological Change: An Empirical Exploration.” Quarterly Journal of Economics, 2003. / Brynjolfsson, Erik, Daniel Rock, and Chad Syverson. “Artificial Intelligence and the Modern Productivity Paradox.” NBER Working Paper 24001, 2017. / Bank for International Settlements. “Big Tech in Finance: Opportunities and Risks.” BIS Annual Economic Report, 2019. / UNCTAD. “Digital Economy Report 2021: Cross-Border Data Flows and Development.” United Nations, 2021.

Online resources — MIT 14.13 / 14.661 (Economics of AI and Labour, Acemoglu/Autor). / Stanford ECON 235 (AI and Economics, lecture notes). / Harvard Kennedy School Technology Policy seminar readings. / LSE EC413 Economics of Innovation materials. / Toronto Rotman Creative Destruction Lab AI stream publications. / OECD “Artificial Intelligence in Society” (2019). / IMF “World Economic Outlook: Countering the Cost-of-Living Crisis” chapter on AI and inequality (2023).

Part I: Foundations — AI as an Economic Object

Chapter 1: What Kind of Technology Is AI? General-Purpose Technology and Prediction Machines

1.1 The Economic Framing Problem

Economists have struggled to assign artificial intelligence a stable ontological category, and this difficulty is not merely taxonomic — it carries profound analytical consequences. When a technology resists classification, standard policy instruments misfire, statistical agencies measure it poorly, and growth accounting frameworks distribute its gains in misleading ways. The opening task of this course is therefore to construct a clean economic framing of AI before engaging any of the substantive debates about labour markets, platforms, or financial stability.

The most productive starting point is to ask what, precisely, AI reduces the cost of. Agrawal, Gans, and Goldfarb argue in Prediction Machines that modern AI — meaning primarily the class of statistical learning systems that have dominated since the deep learning revolution of the early 2010s — is best understood as a dramatic reduction in the cost of prediction (预测). Prediction in their framework is not limited to forecasting the future; it encompasses any task that involves mapping observed inputs to uncertain outputs, including diagnosis, classification, translation, recommendation, and generation. When the cost of prediction falls, the value of complements to prediction rises. Those complements include judgment (判断) — the activity of deciding what to do given the prediction — and data (数据) — the raw material that makes prediction possible. This apparently simple reframing carries considerable analytic power. It explains why firms that were data-rich before the AI transition (Google, Amazon, Facebook) captured enormous rents when AI matured, why human judgment has not been automated away in the manner that early enthusiasts predicted, and why the returns to AI adoption are asymmetric across sectors that differ in their prediction-complementarity profiles.

A competing framing treats AI as a general-purpose technology (通用目的技术), a concept developed by Timothy Bresnahan and Manuel Trajtenberg in their influential 1995 paper. A general-purpose technology, abbreviated GPT in the economic literature (not to be confused with the language model bearing the same acronym), is defined by three properties: pervasiveness across sectors, improvement over time, and the capacity to generate innovational complementarities (创新互补性) — that is, to stimulate the creation of new products and processes in the sectors it enters. The canonical historical examples are the steam engine, electrification, and information and communication technology. Each GPT shared the same awkward entry signature: a long lag between adoption and measured productivity growth, followed eventually by a surge that altered the economic landscape in fundamental ways. The parallel to AI is almost too obvious to need stating, though as Chapter 6 will explore, the productivity paradox of the current AI transition is not simply a repetition of the electrification story.

1.2 Decision Theory Under AI: The Agrawal-Gans-Goldfarb Framework

Agrawal, Gans, and Goldfarb develop their prediction-machine framing into a coherent decision-theoretic architecture that is worth examining formally. In standard decision theory, an agent chooses an action \( a \) from a feasible set \( \mathcal{A} \) to maximise expected utility:

where \( \omega \) is a state of the world drawn from some distribution \( F(\omega) \) that the agent cannot directly observe. The agent forms beliefs about \( \omega \) using available signals \( s \), obtaining a posterior \( F(\omega \mid s) \), and then applies judgment — a payoff function that assigns value to (action, state) pairs — to choose optimally. In this framework, prediction is the step that maps signals to posteriors. AI improves prediction by processing richer signals more accurately, thereby reducing the effective variance of \( F(\omega \mid s) \) and making the posterior more informative. Crucially, however, AI does not supply the payoff function \( U \). That function encodes values, priorities, and context-specific weightings that remain the province of human judgment.

This framework has immediate empirical implications. It predicts that AI adoption will be fastest in domains where prediction is the binding constraint on good decision-making — medical imaging diagnosis, credit scoring, demand forecasting — and slowest where judgment or physical manipulation dominate. It also predicts that as prediction costs fall toward zero, the strategic value of being a good judge of AI output rises relative to the value of being a good predictor. This is a structural argument against the naive view that AI will simply replace high-skill workers wholesale: the cognitive tasks that remain valuable are precisely those where human judgment determines how predictions translate into consequential actions.

The 2022 sequel, Power and Prediction, extends the framework to incorporate systemic effects. When prediction is costly, organisations are structured to accommodate uncertainty — middle managers exist partly to make judgment calls that AI could in principle resolve more cheaply. As prediction cost falls, Agrawal and colleagues argue, decision-making power concentrates at the points in an organisation where judgment is exercised, which may be either the top (centralisation) or the point of customer contact (decentralisation), depending on the nature of the judgment. This is not merely a management theory observation; it has implications for within-firm inequality and for the structure of labour demand explored further in Part II.

1.3 AI as Infrastructure vs. AI as Capital

A further dimension of the framing debate concerns whether AI is best conceptualised as a capital good (资本品) or as infrastructure (基础设施). The distinction matters for growth accounting and for the design of tax and competition policy. When AI is treated as a capital good, the standard Cobb-Douglas production function:

is extended by disaggregating capital into conventional capital \( K_C \) and AI capital \( K_{AI} \), with potentially different substitution elasticities with respect to labour. Aghion, Jones, and Jones (2019) propose a model in which AI capital is a perfect substitute for certain kinds of labour in the production of intermediate goods, but in which the overall production function has a Leontief structure at the top level — meaning that growth is limited by the factor whose supply grows slowest. Their model generates the potentially disquieting prediction that AI, if it substitutes for labour in sufficiently many tasks, could drive down the labour share of income without producing commensurate gains in wages.

When AI is conceptualised instead as infrastructure — more analogous to the electrical grid than to a single machine — a different set of questions arises. Infrastructure is characterised by high fixed costs, near-zero marginal costs, and positive externalities that make private provision likely to be suboptimal. If AI infrastructure (compute, foundational models, data pipelines) has these properties, then the economics of AI provision resembles a natural monopoly problem, and regulation rather than pure market allocation may be the welfare-maximising response. As HIST 415 (A History of Artificial Intelligence) traces in its treatment of the ARPANET and early semiconductor industry, the history of computing infrastructure is in fact replete with moments when public funding — DARPA grants, NSF supercomputer centres, university consortia — provided the foundational investment that private markets were unwilling to undertake. Whether the current moment calls for analogous public intervention in foundational model development is a live policy debate.

1.4 The Measurement Problem

Before any empirical claim about AI and growth can be evaluated, the measurement problem must be confronted squarely. National accounts were designed in the mid-twentieth century to measure an economy where output was predominantly physical goods with stable price indices. The digitisation of the economy has progressively eroded this foundation. Three measurement problems are particularly acute in the context of AI.

First, AI outputs are often quality improvements (质量改进) rather than new quantities. When a recommendation algorithm improves the match between a consumer and a product, no new transaction is recorded in GDP — only the existing transaction value. The welfare gain from better matching is invisible to national accounts. Brynjolfsson and colleagues have proposed GDP-B as an extension that would measure the consumer surplus generated by free digital goods, but this framework has not been adopted by statistical agencies.

Second, AI capital investment is difficult to distinguish from broader IT investment in standard capital expenditure data. A server farm that trains a large language model is categorised as the same asset class as a server farm that processes payroll. This conflation makes it nearly impossible to identify AI-specific investment trends from standard BEA or Statistics Canada data.

Third, and most fundamentally, the hedonic price indices (享乐价格指数) used to adjust for quality in computing equipment have struggled to keep pace with the pace of AI capability improvement. Moore’s Law has historically provided a rough benchmark, but foundation model capability does not scale simply with compute — it also scales with data, architecture innovations, and training techniques, none of which are captured in standard price deflators. The consequence is that AI’s contribution to real output growth is likely systematically underestimated by official statistics, which in turn creates a statistical appearance of productivity stagnation even when genuine improvements are occurring.

Chapter 2: Measuring AI: Productivity, Prices, and the Statistics Problem

2.1 The Solow Paradox and Its Contemporary Analogue

In 1987, Robert Solow offered one of the most celebrated aphorisms in the history of economics: “You can see the computer age everywhere except in the productivity statistics.” This observation — now known as the Solow paradox (索洛悖论) or the computer productivity paradox — posed a genuine puzzle for growth economists. The rapid diffusion of computing technology through the 1970s and 1980s should, on standard neoclassical grounds, have raised total factor productivity. Yet measured multifactor productivity growth was anaemic throughout this period. The resolution, which economists broadly accept, came in two parts. First, diffusion takes time: it took decades after electrification before factory layouts were redesigned to exploit the properties of electric motors, and the productivity gains from computing were similarly delayed while organisations learned to restructure around the new technology. Second, measurement error was significant: price deflators for computing equipment were inadequate, so the real capital stock of computers was understated, and their contribution to output was correspondingly invisible.

Brynjolfsson, Rock, and Syverson have argued forcefully that AI is producing a direct contemporary analogue of the Solow paradox. Their intangible assets hypothesis holds that AI investment is systematically complemented by intangible co-investments — in data curation, process redesign, workforce retraining, and organisational restructuring — that are largely expensed rather than capitalised in national accounts. This means that the true investment associated with an AI deployment is far larger than the visible hardware and software expenditure, and the returns to that investment will be measured only when the complementary intangibles begin to generate output. The implication is optimistic in one sense: the productivity surge may still be coming. But it raises difficult questions about how long the lag will be, and whether measurement reform can accelerate the detection of gains that are in fact already occurring.

2.2 Growth Accounting and the AI Capital Share

The standard growth accounting framework, developed by Solow and extended by Jorgenson and Griliches, decomposes the growth rate of output into contributions from capital accumulation, labour input growth, and total factor productivity (全要素生产率), which is the residual:

where hats denote growth rates and \( \alpha \) is the capital share. When AI capital is introduced as a distinct input, the accounting becomes:

The key empirical question is the magnitude and trajectory of \( \alpha_{AI} \), the share of output attributable to AI capital. Estimates vary widely depending on methodology, but several studies using firm-level data suggest that the AI capital share is rising rapidly in a small number of tech-intensive sectors while remaining negligible in most of the economy. This bimodal distribution is itself economically significant: it suggests that AI’s aggregate contribution to growth is currently concentrated enough to be sensitive to measurement choices in a handful of industries.

2.3 Price Measurement and the Hedonic Challenge

The construction of accurate price indices for AI services faces challenges that are qualitatively different from those encountered with conventional goods. A hedonic price index for, say, automobiles regresses observed prices on observable characteristics — horsepower, fuel economy, safety ratings — and uses the estimated coefficients to hold quality constant across time. For AI services, the relevant quality dimensions are not fixed: a language model’s capabilities in 2023 include tasks that were literally impossible for any system in 2020, so there is no stable basket of characteristics to hold constant.

One approach, advocated by several statistical agencies, is to use benchmark tasks (基准任务) as quality anchors — standardised tests of capability that can be administered across model generations to produce a consistent quality-adjusted price series. This approach has the advantage of grounding quality in measurable performance, but it faces the problem that AI capabilities improve fastest precisely on the benchmarks used to measure them, a phenomenon sometimes called Goodhart’s Law (古德哈特定律) in the machine learning context: when a measure becomes a target, it ceases to be a good measure.

A second approach is to use willingness to pay (支付意愿) as revealed through market data — the prices that businesses and consumers actually pay for AI services compared to the prices they paid for the tasks that AI is displacing. This approach has the advantage of grounding measurement in actual economic decisions, but it conflates quality improvements with changes in market power, since the pricing of AI services reflects both the technology’s value and the market concentration of the providers offering it.

2.4 The Aggregate Productivity Debate

The aggregate productivity debate in AI economics does not resolve neatly into a single empirical finding, and it is worth charting the main positions carefully. Optimists, led by Brynjolfsson and McAfee, argue that the current productivity stagnation in official statistics is an artefact of measurement failure and implementation lag. The second machine age has brought capabilities that are genuinely transformative — image recognition, language understanding, game-playing at superhuman levels — and the economic gains from these capabilities will manifest in productivity statistics once organisational adaptation catches up. The historical analogy to electrification provides their central evidence: the productivity surge of the 1920s came roughly forty years after the invention of the dynamo, when factory layouts were finally redesigned from the ground up for electric power rather than retrofitted from steam-era designs.

Pessimists, including Robert Gordon in his magisterial The Rise and Fall of American Growth, argue that the comparison to electrification is misleading. The great inventions of the late nineteenth and early twentieth centuries — electrification, the internal combustion engine, sanitation, telecommunications, pharmaceuticals — were transformative in ways that touched every hour of every person’s day. AI, on this view, is impressive but narrow: it accelerates certain cognitive tasks without generating the kind of broad-based quality-of-life improvements that drove the great productivity surge of 1920–1970. Gordon’s argument is fundamentally about the scope of transformative complementarity (变革性互补性): GPTs generate large productivity surges only when they complement and enable improvements across the full range of economic activities, not merely in cognitively intensive tasks. Whether AI crosses this threshold remains an open empirical question.

Part II: Labour Markets and Automation

Chapter 3: The Task-Based Model of Automation

3.1 From Skill-Biased to Task-Biased Technical Change

For much of the second half of the twentieth century, labour economists explained the rising wage premium for educated workers using the framework of skill-biased technical change (技能偏向型技术变革), abbreviated SBTC. The basic idea is straightforward: new technologies — computers, information systems, sophisticated machinery — are complementary to high-skill workers and substitutable for low-skill workers. As technology advances, demand for high-skill labour rises and demand for low-skill labour falls, which, in a standard supply-and-demand framework, predicts rising wage inequality. SBTC proved empirically powerful in explaining the widening college-high school wage gap through the 1980s and 1990s.

By the early 2000s, however, the data were revealing anomalies that SBTC could not cleanly explain. The wage distribution was not simply becoming more unequal at the top — it was becoming polarised (极化): high-wage and low-wage occupations were growing while middle-wage occupations were hollowing out. This pattern prompted Autor, Levy, and Murnane to propose the routine task hypothesis (常规任务假说), which redirects attention from skill levels to the content of specific tasks. Their 2003 paper in the Quarterly Journal of Economics draws a fundamental distinction between routine tasks (常规任务) — activities that can be fully characterised by explicit rules — and non-routine tasks (非常规任务) — activities that require adaptability, judgment, or interpersonal interaction. Computers and automation technologies excel at routine tasks, whether cognitive (bookkeeping, data processing, formula application) or manual (assembly line welding, repetitive physical sorting). Non-routine tasks, by contrast, resist automation because they cannot be reduced to an algorithmic specification.

This reframing has a crucial implication: the occupation of a worker matters less for automation risk than the task content of that occupation. A lawyer who spends most of her time drafting standardised contracts faces higher automation exposure than a home health aide who coordinates unpredictable care needs, even though the lawyer’s average wage is far higher. The empirical prediction — that middle-skill, routine-intensive occupations will decline while high-skill and low-skill occupations grow — matches the observed polarisation of the wage distribution with a precision that the simple SBTC story cannot achieve.

3.2 The Acemoglu-Restrepo Task-Based Framework

Acemoglu and Restrepo have developed the most formally rigorous version of the task-based approach, and it is worth working through their framework with some care because it underlies much of the empirical and policy analysis in the remainder of the course. The economy produces a final good using a continuum of tasks indexed by \( i \in [0, 1] \). Each task can be performed either by labour or by capital (automation technology):

where \( y(i) \) is the quantity of task \( i \) produced. Labour has a comparative advantage in tasks near \( i = 1 \) and automation has a comparative advantage in tasks near \( i = 0 \). There exists a threshold \( \mathcal{I} \) such that tasks with \( i < \mathcal{I} \) are performed by automation and tasks with \( i > \mathcal{I} \) are performed by labour. Automation shifts \( \mathcal{I} \) upward, displacing labour from more tasks.

The key insight of Acemoglu and Restrepo is that automation need not reduce aggregate labour demand as long as the reinstatement effect is strong enough to offset displacement. Historically, the creation of new tasks — the development of entirely new occupations accompanying new industries — has been the primary mechanism by which labour maintains employment even as technology raises the productivity of each worker. However, Acemoglu and Restrepo argue in their “Wrong Kind of AI” paper that recent AI investment has been disproportionately concentrated in automating existing tasks rather than creating new ones, resulting in a net displacement effect that is unusually large relative to historical precedent. If the reinstatement effect remains weak — if the new economy does not spontaneously generate the new task categories that absorbed displaced workers in previous technological transitions — then the current AI wave could produce persistent unemployment and wage stagnation rather than the technological abundance that optimists predict.

3.3 Automation and Factor Shares

One of the most consequential empirical predictions of the task-based framework concerns the functional distribution of income. In a standard two-factor model, the labour share of income is determined by the relative bargaining power and productivity of capital and labour. Automation, in the task-based model, affects the labour share through two channels. The price effect operates because automation reduces the cost of producing tasks, which expands real output; if labour demand is elastic enough, the expansion can raise the wage bill even as the labour share falls. The quantity effect operates directly through the displacement of workers from tasks, reducing the share of production activities attributable to labour.

Empirically, the labour share of national income has declined in most advanced economies since the 1980s, and the timing correlates with waves of automation investment. Karabarbounis and Neiman documented this trend across sixty-plus countries in their 2014 paper and attributed it primarily to declining relative prices of investment goods — a finding that is consistent with the task-based model’s prediction that cheaper automation capital substitutes for labour in routine tasks. However, disentangling the contribution of AI specifically from broader trends in capital deepening, globalisation, and market concentration is empirically demanding and the subject of ongoing debate.

Chapter 4: Skills Polarisation and the Hollowing-Out of Work

4.1 The Empirical Phenomenon of Labour Market Polarisation

The most extensively documented empirical consequence of routine-task automation is labour market polarisation (劳动力市场极化), a term coined by Autor and Dorn in their influential 2013 study of the US labour market. Polarisation refers to the simultaneous growth of high-wage, high-skill employment and low-wage, low-skill employment, combined with the erosion of middle-wage, middle-skill employment. When occupation-level data are plotted by initial wage percentile, the growth in employment shares forms a U-shape: positive growth at both tails and negative growth in the middle, where routine-intensive clerical, administrative, and production occupations cluster.

The hollowing-out of the middle is not merely a statistical artefact. It represents the elimination of career ladders that historically provided pathways from working-class entry-level jobs to middle-class stability. A production line worker in the 1970s could advance through seniority and skill acquisition into supervisory or technical roles that required the kind of tacit, context-specific knowledge that automation could not replicate. When the middle of the skill distribution is compressed by automation, these career pathways disappear, leaving workers who lose routine jobs with the choice between upskilling to high-demand professional occupations (a transition that is expensive and uncertain) or accepting low-wage service employment.

4.2 Geographic Concentration of Polarisation Effects

Polarisation is not uniformly distributed across geographic space, and the local labour market dimension of automation is one of the richest areas of recent empirical research. Autor, Dorn, and Hanson show that commuting zones with higher initial concentrations of routine-intensive employment — often manufacturing-heavy regions in the American Midwest and upper South — experienced significantly larger declines in employment and wages when automation exposure intensified. The geographic concentration of automation effects is amplified by the immobility of workers: unlike capital, which can quickly be redeployed from declining industries to growing ones, workers face housing market frictions, family ties, and social network dependencies that reduce their ability to relocate in response to local labour market shocks.

The interaction of automation with agglomeration economies (集聚经济) further concentrates gains and losses. High-skill, non-routine-intensive employment clusters in large metropolitan areas that offer complementary amenities, thick labour markets, and knowledge spillovers. When automation displaces middle-skill workers in non-metropolitan areas, those workers face the double disadvantage of losing their jobs in regions where alternative non-routine employment opportunities are sparse. This geographic dimension transforms what might appear to be a symmetric distributional shock — gains at both ends of the skill distribution — into a highly asymmetric spatial shock that concentrates losses in specific communities.

4.3 AI and the New Frontier of Automation

Earlier waves of automation, including the computerisation wave of the 1980s and 1990s documented by Autor, Levy, and Murnane, primarily threatened routine cognitive and routine manual tasks. The distinctive feature of modern AI — particularly large language models, multimodal models, and AI-assisted robotics — is that it is beginning to encroach on tasks previously thought to be safe because they required adaptability, language comprehension, or perceptual sophistication.

Acemoglu and colleagues’ 2022 paper “Tasks, Automation, and the Rise in US Wage Inequality” revisits the task boundary question in the context of AI. They find that while AI has expanded the automation frontier, the tasks being automated now — complex customer service, legal document review, radiological imaging interpretation, financial analysis — cluster in higher wage percentiles than the tasks automated by earlier computerisation. This finding has two opposed implications. On the one hand, it suggests that AI could reduce inequality by compressing the upper tail of the wage distribution (high-skill workers being displaced by AI) rather than further polarising it. On the other hand, it suggests that the professional classes who previously felt protected by the non-routine character of their work may face significant exposure to AI displacement for the first time.

Chapter 5: Robots, Wages, and the Evidence Base

5.1 The Acemoglu-Restrepo Robot Study

The most widely cited empirical study of automation and labour markets is Acemoglu and Restrepo’s 2020 paper in the Journal of Political Economy, “Robots and Jobs: Evidence from US Labor Markets.” Their approach uses data on industrial robot adoption from the International Federation of Robotics, combined with regional variation in initial industry composition to construct an exposure variable (暴露变量) that measures the extent to which each local labour market was exposed to robot adoption driven by industry-level demand for automation elsewhere. This exposure is constructed to be plausibly exogenous to local labour demand shocks, enabling a causal interpretation of the estimates.

Their findings are striking. One additional robot per thousand workers in a commuting zone is associated with a reduction in the employment-to-population ratio of 0.18–0.34 percentage points and a decline in average wages of 0.25–0.5 percent. These effects are not trivial at the margin: during the period of their study (1990–2007), the authors estimate that robots reduced employment by 400,000 to 840,000 workers and wages by 0.5 percentage points. Crucially, they find no significant offsetting employment gains in other occupations within the same commuting zones — the displacement effect dominated the reinstatement effect in the sample period, though they are careful to note that long-run reinstatement effects may not yet be visible in the data.

5.2 Critiques and Heterodox Evidence

The Acemoglu-Restrepo robot study has been subjected to vigorous methodological scrutiny, and several of the critiques carry real analytical force. Daron Acemoglu and colleagues have themselves acknowledged that the study measures the effects of a specific vintage of industrial robots — large, expensive, single-purpose machines used predominantly in automotive and electronics manufacturing — that may not be representative of modern AI systems. The concern is that the external validity of findings from 1990s industrial robots to 2020s AI is limited precisely because AI systems operate through very different mechanisms: they work through the automation of cognitive tasks rather than physical ones, they are deployed in white-collar rather than manufacturing settings, and they interact differently with human complementarity.

David Autor’s response to these critiques is nuanced. He argues that the task-based framework remains valid regardless of the physical character of the automation technology; what matters is the task content being displaced, not whether the displacing system is a robot arm or a language model. The appropriate empirical question for AI is therefore to identify which cognitive tasks have been displaced, in which occupations, and with what magnitude — and the methodological tools for answering this question are extensions of the same tools used in the robot study, with the occupational exposure variable constructed from AI capability rather than robot density data.

5.3 International Evidence and Developing Economy Exposure

Most empirical work on robots and labour markets has focused on advanced economies with rich administrative data. The international evidence is sparser but directionally consistent: Graetz and Michaels find robot adoption positively associated with labour productivity and GDP growth in a cross-country panel of seventeen advanced economies, but with negative effects on low-skill workers’ employment shares. Chang and Huynh use data from Asia to find that automation exposure in manufacturing export sectors has significant pass-through to wage inequality, with middle-skill workers again bearing a disproportionate share of the adjustment burden.

For developing economies, the exposure question is more complex because their manufacturing sectors were already facing competitive pressure from the same forces driving automation adoption in advanced economies. If manufacturing in Vietnam or Bangladesh is automated, the adjustment burden does not fall on high-wage workers being displaced into service employment — it falls on low-wage manufacturing workers who may have no access to a social safety net and limited opportunities for occupational transition. This development dimension of automation is explored more fully in Part V.

Chapter 6: The Productivity Paradox of AI

6.1 The Paradox Stated

It is now a stylised fact that the period of the most dramatic AI capability improvements — roughly 2012 to the present — has coincided with one of the most prolonged periods of weak multifactor productivity growth in the postwar history of advanced economies. US total factor productivity growth averaged roughly 1.5 percent per year in the 1995–2004 ICT boom but has averaged under 0.5 percent per year in the decade following 2010. This divergence between the subjective impression of rapid technological change — which anyone who has used a modern AI system will find compelling — and the objective measurement of productivity growth constitutes the AI productivity paradox (人工智能生产力悖论).

Brynjolfsson, Rock, and Syverson lay out three hypotheses for this paradox. The first is mismeasurement: AI’s gains are real but invisible to standard statistical instruments because they manifest as quality improvements and consumer surplus rather than measured output. The second is implementation lag: AI’s gains are real but delayed, pending the organisational and complementary investment changes needed to translate AI capability into economic output. The third is limited scope: AI’s gains are real in narrow domains but insufficient to move aggregate productivity because the relevant sectors are too small or too insulated from the rest of the economy to generate broad-based spillovers.

6.2 The Implementation Lag Hypothesis

The implementation lag hypothesis has the strongest historical support. The electrification analogy is canonical, but a closer parallel may be the productivity effect of ICT in the 1990s. Computers and networking technologies had been diffusing throughout the economy since the 1960s, but the productivity acceleration associated with ICT did not appear in aggregate statistics until the mid-1990s — by which point the internet had become commercially widespread and, critically, firms had had thirty years to experiment with organisational redesign. The evidence strongly suggests that the productivity payoff to general-purpose technologies depends on accumulated learning and organisational capital that builds slowly over many adoption cycles.

Erik Brynjolfsson and colleagues at the MIT Productivity Lab have developed a framework they call the J-curve of AI adoption, in which firm-level productivity initially dips when AI is first adopted (because the firm is investing in training, reorganisation, and process redesign) before eventually rising above the pre-adoption baseline as complementary intangibles accumulate. The aggregate productivity statistics are noisy averages across firms at different points in the J-curve, and in the early phases of diffusion, the firms that have invested most in AI (and are currently in the trough of the J-curve) are also the largest and most productive, so their temporary productivity dip depresses the aggregate average.

6.3 AI and Within-Firm Productivity Heterogeneity

One of the most robust empirical findings in the micro-productivity literature is that AI adoption generates enormous within-sector heterogeneity (行业内异质性). Firms that successfully integrate AI into core business processes show large productivity gains; firms that adopt AI without accompanying organisational change show modest or negative returns. This heterogeneity mirrors findings from the ICT era — Nicholas Bloom and colleagues showed that the same computer hardware produced vastly different productivity outcomes depending on management quality — but appears to be even more pronounced for AI, perhaps because AI systems require more substantial complementary restructuring of data pipelines, workforce composition, and decision rights.

The policy implication is significant. If the productivity benefits of AI are concentrated in firms with particular organisational capabilities, then aggregate AI adoption rates may tell us relatively little about aggregate productivity outcomes. The distribution of organisational capability across firms matters as much as the distribution of AI access. This suggests that policies aimed purely at diffusing AI access — through subsidised cloud computing, open-source model provision, or SME digitalisation grants — may generate less productivity growth than policies that simultaneously address the complementary organisational barriers.

Part III: Platform Economics and Market Power

Chapter 7: Two-Sided Markets, Network Effects, and Winner-Takes-All

7.1 Platform Economics: The Basic Architecture

Platform economics (平台经济学) studies markets in which a firm creates value primarily by facilitating interactions between two or more distinct user groups. Parker and Van Alstyne, along with Rochet and Tirole in their parallel academic formulations, established the theoretical framework for understanding how these markets differ from standard pipeline businesses. A platform does not buy inputs and sell outputs in the conventional sense; instead, it provides infrastructure — physical, digital, or relational — that reduces transaction costs between groups whose interaction creates surplus. The key analytical novelty is that the two sides of the platform exhibit indirect network effects (间接网络效应): the value of the platform to users on side A depends on the number and quality of users on side B, and vice versa.

The canonical example is a credit card network. More cardholders make the card more attractive to merchants (more potential customers), and more merchants accepting the card make it more attractive to cardholders (more places to use it). The network effects are indirect because cardholders do not directly value the presence of other cardholders — they value the presence of merchants, and vice versa. This structure creates a fundamental pricing challenge: the platform must attract both sides simultaneously, and the optimal pricing structure depends on the relative elasticities of demand on each side and the intensity of the cross-side network effects.

The Rochet-Tirole pricing formula for a two-sided platform sets prices on each side according to:

where \( c_A, c_B \) are marginal costs of serving each side, \( \alpha, \beta \) are the cross-side network effect parameters, \( n_A, n_B \) are the number of users on each side, and \( \varepsilon_A, \varepsilon_B \) are demand elasticities. The platform subsidises the side with stronger cross-side effects and lower own-price elasticity, and charges the side that benefits most from the other side’s presence. This is why credit card companies historically charged merchants high fees while offering cardholders free services or cash-back rewards, and why search engines charge advertisers while providing search to users for free.

7.2 AI and the Intensification of Network Effects

AI intensifies platform network effects through at least two distinct channels. First, AI enables algorithmic personalisation (算法个性化) that makes the platform more valuable to each user as more users and their behavioural data are accumulated. A recommendation system that knows the viewing histories of a hundred million users can generate better predictions for any individual user than a system trained on a thousand users, even holding fixed the capabilities of the underlying model. This creates a direct link between platform scale and AI quality, meaning that larger platforms have better AI, which attracts more users, which generates more data, which improves the AI further. This data feedback loop (数据反馈循环) is a form of network effect that is qualitatively different from the simple presence-of-users effect in Rochet-Tirole: it is dynamic and self-reinforcing rather than static.

Second, AI enables platforms to monetise user data through targeted advertising (定向广告) with an efficiency that non-AI approaches cannot match. The value of an advertising impression is dramatically higher when the serving algorithm can predict with reasonable accuracy which users are in-market for the advertised good or service. Advertisers’ willingness to pay for targeted impressions is therefore a direct function of the platform’s AI quality, which is in turn a function of its data accumulation. Hal Varian, Google’s chief economist, has articulated this logic explicitly: data about user behaviour has value because it improves prediction, and better prediction makes advertising more valuable. This is an operationalisation of the Agrawal-Gans-Goldfarb prediction machine framework in a specific market context.

7.3 Winner-Takes-All Dynamics

The combination of indirect network effects, data feedback loops, and algorithmic scale economies creates strong tendencies toward winner-takes-all (赢者通吃) outcomes in platform markets. When platforms compete, each has an incentive to subsidise user acquisition aggressively in order to build the data and network scale needed to sustain competitive advantage. This produces dynamics that bear some resemblance to the standards wars of the technology industry — VHS vs. Betamax, Windows vs. Mac OS — but with the important difference that platform competition typically resolves into a single dominant firm rather than a standard shared across competitors.

The economist Jean Tirole has noted that winner-takes-all outcomes in platform markets can be welfare-improving in the short run (one large platform generates more efficient matching than two small ones) but welfare-reducing in the long run (the dominant platform uses its position to extract monopoly rents and suppress competitive entry). This tension is at the heart of the antitrust debates examined in Chapter 9.

Multi-homing (多归属) is the primary force that can sustain competitive platforms in the presence of winner-takes-all pressures. When users participate in multiple competing platforms simultaneously — as many users do with ride-hailing (Uber plus Lyft) or social media (Instagram plus TikTok plus Twitter/X) — the network effects advantage of the dominant platform is partially offset because the marginal cost of adding another platform is low. However, AI may be reducing multi-homing by creating switching costs (转换成本) rooted in data. A language model assistant that has been fine-tuned on a user’s preferences, communication style, and historical interactions becomes more valuable the longer it is used, creating a lock-in dynamic that discourages switching to competing assistants even if the competitors offer similar base capabilities.

Chapter 8: Data as a Competitive Moat: Algorithmic Pricing and Collusion

8.1 Data as a Non-Rival Asset

The economic properties of data differ from those of conventional inputs in ways that are consequential for competition analysis. Hal Varian has articulated the key distinction: data is non-rival (非竞争性) in the sense that its use by one firm does not diminish its availability to another, and it is often non-excludable (非排他性) in the absence of legal or technical restrictions. These properties suggest that data resembles a public good rather than a private input, which would imply that data should be shared freely to maximise social welfare. In practice, however, data is made excludable through trade secrecy, contractual restrictions, and technical means, and its non-rivalry is offset by the proprietary nature of the AI models trained on it.

The competitive significance of data moats depends on the specific market and the minimum efficient scale of data (数据最小有效规模) needed to train a competitive AI system. In some markets — image recognition, where millions of labelled images are needed to train a competitive model — data moats are formidable barriers to entry. In others — recommendation systems for niche content — relatively small datasets can produce satisfactory performance, and the barrier is lower. The antitrust question of whether data constitutes an essential facility (必要设施) that should be subject to mandatory sharing is actively debated by competition economists and regulators.

8.2 Algorithmic Pricing and the Collusion Problem

Algorithmic pricing (算法定价) refers to the use of automated systems to set prices dynamically in response to market conditions. The most benign version — adjusting prices in response to real-time supply and demand information — is economically beneficial: it improves market efficiency by aligning prices with marginal costs more accurately than humanly-administered price lists could. Airlines, hotels, and ride-hailing platforms use algorithmic pricing in essentially this form.

The concerning version involves the possibility that competing firms’ pricing algorithms, even without explicit communication or coordination, could converge to supra-competitive prices (超竞争价格) through a process that resembles collusion but involves no illegal agreement. If two firms in a market both use pricing algorithms that respond to each other’s prices, the algorithms may learn — through reinforcement learning or related techniques — that mutual price restraint is a Nash equilibrium strategy in the repeated game of price-setting. This is the algorithmic collusion problem identified by economists including Ariel Ezrachi and Maurice Stucke in their work Virtual Competition.

The antitrust law of collusion in most jurisdictions requires evidence of agreement or concerted action; purely parallel algorithmic behaviour, even if its effect is supra-competitive prices, may not meet the legal standard for liability. This creates a significant regulatory gap that competition authorities are beginning to address through updated guidelines and, in some jurisdictions, new legislative instruments.

8.3 Data Markets and Monopsony

A further dimension of data economics concerns the market for data itself. When firms acquire user data through free services — offering email, search, or social networking in exchange for the right to use behavioural data for advertising — they are operating as monopsonists (买方垄断者) in the data market: they are the only buyer, and users are the sellers who receive a non-monetary payment (the service) in exchange for their data. The welfare analysis of this monopsony position is complex. Users may or may not value their data at the price implicitly offered by the platform (the utility from the free service), and the platform may or may not be using the data in ways that users would consent to if fully informed.

Shoshana Zuboff’s concept of surveillance capitalism (监控资本主义) offers a sociological rather than economic analysis of this same phenomenon, arguing that the extraction of behavioural data is structurally coercive because users lack meaningful alternatives to the dominant platforms and lack the information needed to evaluate the implicit exchange. While Zuboff’s framework does not map neatly onto neoclassical welfare economics, it highlights the limitations of revealed-preference reasoning in markets characterised by information asymmetry and incomplete competition.

Chapter 9: Antitrust in the Age of AI: Google, Meta, and the Digital Markets Act

9.1 The Structural Economics of Big Tech Antitrust

The antitrust economics of AI platforms departs from both the post-Chicago School consumer welfare standard and its Chicago predecessors in important ways. The post-Chicago approach, developed through the work of economists including Jean Tirole and Jean-Charles Rochet, acknowledges that two-sided markets may require pricing structures that appear anticompetitive in a one-sided analysis — below-cost pricing on one side, bundling, tying — but that are efficient when the platform’s two-sidedness is properly accounted for. However, this approach was developed for markets where the network effects were relatively stable and the data advantage of the dominant platform was not self-reinforcing.

Modern AI platform markets combine several features that the existing antitrust framework handles imperfectly: the self-reinforcing data feedback loop; the capacity of dominant platforms to use data and AI to monitor competitive threats and preemptively neutralise them; the vertical integration of AI infrastructure (compute, data, foundational models) with downstream AI applications; and the pace of innovation that makes market definition unstable. The Department of Justice’s 2023 case against Google’s search advertising monopoly, and the FTC’s cases against Meta’s acquisition strategy, represent the first serious attempts to apply antitrust law to these features.

9.2 Google Search: Monopoly Through Quality or Foreclosure?

The DOJ’s case against Google centres on the claim that Google maintained its monopoly in general search through exclusive default agreements with browser developers (most importantly Apple) and device manufacturers (Android OEM agreements), paying billions of dollars annually to secure default search engine status. The anticompetitive theory is that these agreements foreclosed the distribution channels through which competing search engines could acquire the scale of user data needed to train competitive AI models, thereby erecting a data moat that potential competitors could not breach.

Google’s defence relied on the product quality argument: users choose Google search because it is genuinely better, and its superiority reflects not exclusionary conduct but superior investment in AI and engineering talent. The antitrust economic question is whether the quality advantage is endogenous to the scale advantages that the exclusive agreements helped create, or whether it would exist even absent the exclusionary conduct. This is an anticompetitive foreclosure (反竞争封锁) claim in the traditional sense, but its mechanism runs through data accumulation and AI training rather than through conventional input foreclosure.

As PHIL 451 (AI Ethics, Law, and Governance) explores, the legal question of what constitutes unlawful exclusionary conduct in AI platform markets is not yet settled, and the regulatory responses being developed in Europe may provide alternative frameworks to the consumer welfare standard of US antitrust law.

9.3 The EU Digital Markets Act

The European Union’s Digital Markets Act (数字市场法), which entered into force in 2022 and became applicable in 2023, represents a structural departure from the case-by-case antitrust approach. Rather than requiring proof of abuse after the fact, the DMA identifies gatekeepers (看门人) — large platform operators meeting quantitative thresholds of size, durability, and intermediation importance — and imposes ex ante obligations on their conduct. These include requirements to provide interoperability with competing services, to allow users to uninstall pre-installed applications, to refrain from self-preferencing (giving a platform’s own services preferential treatment in rankings), and to share data with competitors on fair and non-discriminatory terms.

The data-sharing obligation is particularly significant for AI economics. If dominant platforms are required to share the behavioural data that trains their AI systems with competitors and regulators, the data moat argument for their market power is substantially weakened. However, the design of data sharing obligations is technically challenging: the data must be shared in a form that is useful to competitors without revealing individual user privacy, and the terms must be specified precisely enough to be enforceable without being so prescriptive as to stifle the innovation that generates the data’s value in the first place. The DMA’s implementation details remain a major site of regulatory development as of the mid-2020s.

Part IV: Innovation Economics

Chapter 10: AI as a General-Purpose Technology: R&D and Creative Destruction

10.1 The Schumpeterian Framework

Joseph Schumpeter’s vision of capitalism as a process of creative destruction (创造性破坏) — in which technological innovation generates new industries and firms while displacing incumbent ones — remains the most intellectually powerful framework for understanding the macroeconomics of technological change. Schumpeter distinguished between two modes of innovation: the incremental improvements by existing firms that he associated with competitive markets, and the radical innovations introduced by entrepreneurial entrants that he associated with the temporary monopoly rents of market-disrupting novelty. His insight that the prospect of monopoly rents is the primary incentive for radical innovation has important implications for AI: a world in which AI systems are freely shared and their outputs are not appropriable may generate less innovation investment than a world in which AI innovators can capture at least some of the social returns to their discoveries.

The Arrow replacement effect (阿罗替代效应) modifies this Schumpeterian picture. Arrow argued that an incumbent monopolist has less incentive to innovate than a competitive entrant, because the monopolist’s innovation replaces its own existing profit stream (the “replacement effect”) while the entrant’s innovation creates a new one. In the context of AI, this suggests that the concentrated structure of the AI industry — where a small number of very large firms control the frontier — may reduce the pace of radical innovation while maintaining rapid incremental improvement. Whether the current pace of AI progress reflects the optimal incentive structure for innovation is therefore a live question in the economics of AI.

10.2 AI as an Invention of a Method of Invention

The concept of a general-purpose technology (通用目的技术) maps onto AI with particular force in the domain of research and development. Agrawal and colleagues have argued that AI is not merely a technology that creates new products — it is a technology that accelerates the creation of other technologies. By automating the hypothesis-generation, literature-search, experimental-design, and data-analysis components of the research process, AI effectively reduces the cost of producing scientific knowledge, which is itself the input to all future technological innovation. This is what Bresnahan and Trajtenberg called the invention of a method of invention (发明方法的发明) when describing the role of general-purpose technologies in accelerating innovation.

The empirical evidence for AI as a research accelerator is still accumulating, but early results are striking. Jumper and colleagues at DeepMind demonstrated that AlphaFold could predict protein folding structures with accuracy comparable to experimental methods at a fraction of the cost and time, effectively solving a fifty-year-old grand challenge in structural biology. Studies of AI tools for materials science have shown measurable reductions in the time from hypothesis to experimental validation. In drug discovery, multiple pharmaceutical companies report that AI-assisted target identification has compressed timelines by months. The aggregate implication — that AI could substantially raise the productivity of research effort across the sciences — is potentially the most important economic effect of AI, though it operates on timescales that are difficult to capture in standard productivity measurement.

10.3 Concentration in AI Research and Its Consequences

One of the most concerning features of the current AI innovation landscape is the extraordinary concentration of frontier research in a small number of organisations. Training a state-of-the-art large language model as of 2024 required computational resources costing tens or hundreds of millions of dollars, accessible only to a handful of very large firms (OpenAI/Microsoft, Google DeepMind, Anthropic, Meta) and a small number of national laboratories. This compute concentration (算力集中) has structural implications for the character of AI innovation.

When frontier research is conducted primarily within large commercial organisations, the incentives shaping research are those of product development rather than scientific discovery. Topics that do not have near-term commercial applications are underfunded relative to their social value; negative results (showing that an approach does not work) are published less systematically than in academic science; and the data, model weights, and infrastructure used in frontier research are proprietary rather than shared. Acemoglu and colleagues have argued that these features mean that the direction of AI innovation (人工智能创新方向) is being shaped by the profit motives of a small number of firms, which may diverge significantly from the direction that would maximise social welfare.

10.4 Intellectual Property in AI

The intellectual property (知识产权) framework for AI is in a state of considerable legal and economic uncertainty. Three distinct IP questions bear on AI economics. First, who owns the output of AI systems — is a novel generated by a language model copyrightable, and if so, by whom? Current US copyright law requires human authorship, which creates ambiguity for AI-generated works. Second, do AI systems trained on copyrighted material infringe the copyrights of the training data? This question is currently being litigated in multiple jurisdictions and its resolution will significantly affect the cost structure of AI development. Third, are AI-assisted inventions patentable, and if so, are the patents held by the AI developer or the user who directed the AI?

The economic stakes of these questions are large. If AI-generated outputs are not copyrightable, the investment incentive to develop AI content-generation systems is weakened. If training on copyrighted material requires licensing fees, the cost of training frontier models rises significantly, further entrenching the advantage of firms with deep pockets. If AI-assisted inventions cannot be patented, the incentive to use AI in R&D may be reduced in industries that rely heavily on patent protection. The resolution of these questions will shape the direction and pace of AI innovation over the next decade.

Chapter 11: Concentration, Intellectual Property, and the Open-Source Question

11.1 The Industrial Organisation of AI

The industrial organisation (产业组织) of the AI sector exhibits features that distinguish it from most prior technology industries. At the infrastructure layer, the AI industry is dominated by three cloud providers — Amazon Web Services, Microsoft Azure, and Google Cloud — who collectively control the vast majority of the compute infrastructure on which AI training and inference runs. At the foundation model layer, a small number of vertically integrated firms (those three cloud providers plus Anthropic and a handful of others) maintain the frontier. At the application layer, a much larger and more competitive ecosystem of startups and incumbents deploys AI capabilities to specific use cases. This three-layer structure creates complex vertical relationships: the foundation model providers are also, in many cases, the cloud infrastructure providers who charge application developers for compute, and increasingly also competitors to those same application developers as they build their own AI-powered products.

The concentration at the infrastructure and foundation model layers is the primary source of competition policy concern. If a small number of firms control the infrastructure on which all AI applications depend, they have the potential to extract monopoly rents from application developers, to discriminate against applications that compete with their own products, and to shape the development trajectory of AI in ways that reflect their private interests rather than social welfare. The analogy to telecommunications infrastructure — where concerns about bottleneck control led to common carrier regulation — is frequently invoked in AI policy discussions.

11.2 Patent Races and First-Mover Advantage

In industries characterised by rapid technological change and strong network effects, the incentive to be first to market can generate patent races (专利竞赛) that result in socially excessive investment in R&D. The standard model of patent races, developed by Loury (1979) and Dasgupta and Stiglitz (1980), shows that when the winner of a patent race captures the entire market (winner-takes-all), competing firms will invest up to the expected value of the prize, potentially duplicating research efforts and racing to a deadline that arrives earlier than socially optimal. The social optimum would involve coordinated investment that avoids duplication, but such coordination may itself violate antitrust law.

In the AI context, the patent race dynamic is complicated by the fact that the most valuable AI “inventions” are not discrete, patentable innovations but rather large models trained on massive datasets — assets whose value lies in their accumulated parameters rather than in any specific novel algorithm. The competitive dynamic is therefore less like a patent race (where the winner gets a legal monopoly for a fixed term) and more like a data race (数据竞赛): firms compete to accumulate the data and compute needed to train the best model, with the winner gaining a de facto competitive advantage that may be durable even without formal intellectual property protection.

Part V: Development Economics and the Global South

Chapter 12: Data Colonialism and the Digital Divide

12.1 The Global Geography of AI Production

The geography of AI production is strikingly concentrated. The majority of frontier AI research is conducted in the United States, with secondary hubs in the United Kingdom, Canada, and China. The semiconductor supply chain, which provides the physical infrastructure for AI computation, is concentrated in Taiwan (TSMC), South Korea (Samsung), and the Netherlands (ASML for lithography equipment). The training data for large language models is predominantly in English, with secondary coverage of other high-resource languages (Chinese, Spanish, French, German) and minimal coverage of the hundreds of languages spoken in sub-Saharan Africa, South and Southeast Asia, and Latin America. This geographic concentration of AI production is not accidental; it reflects historical patterns of capital accumulation, educational investment, and institutional development that track closely with earlier patterns of industrial concentration.

The digital divide (数字鸿沟) in AI has both access and participation dimensions. The access dimension concerns whether users in lower-income countries can use AI systems at all — a question constrained by internet connectivity, electricity infrastructure, and the affordability of AI-enabled devices. The participation dimension concerns whether researchers, developers, and entrepreneurs from lower-income countries participate in the creation and governance of AI — a question constrained by the concentration of talent in a small number of elite institutions and the incentives created by immigration policies in AI-producing countries.

12.2 Couldry and Mejias: Data Colonialism

Nick Couldry and Ulises Mejias’s concept of data colonialism (数据殖民主义) provides the most analytically developed critique of the global political economy of AI data production. Drawing on the historical parallel to colonial resource extraction, they argue that the systematic appropriation of human behavioural data by large technology companies — primarily based in the United States and China — constitutes a new form of resource extraction that enriches the extracting powers at the expense of the populations whose data is extracted. The extracted resource is not mineral wealth or agricultural output, as in historical colonialism, but social quantification (社会量化): the transformation of human behaviour, relationships, and experience into data that can be processed by AI systems and monetised through advertising, financial products, and platform services.

The colonial analogy has limits — unlike colonial subjects, most AI data contributors nominally consent to data collection through terms of service agreements, and they receive services in return — but Couldry and Mejias argue that the consent is illusory and the exchange is structurally unequal. The populations whose data is collected have little ability to negotiate the terms of collection, little transparency about how their data is used, and no share in the profits generated from AI systems trained on their data. Meanwhile, the firms collecting the data are headquartered in jurisdictions with strong legal systems that protect intellectual property and trade secrets, making it difficult for data-originating communities to assert ownership claims.

12.3 Leapfrogging vs. Lock-In

The historical literature on development and technology has identified leapfrogging (跨越式发展) — the ability of late-developing economies to skip intermediate technological stages by adopting the latest available technology — as one of the potential advantages of late industrialisation. The mobile telephone revolution in sub-Saharan Africa is the canonical example: countries that never built extensive fixed-line telephone infrastructure went directly to mobile networks, enabling communication and financial services that would have required decades of fixed infrastructure investment in an earlier era.

Whether AI enables a similar leapfrogging dynamic for the Global South, or instead creates new forms of lock-in (锁定) that entrench the technological advantage of early developers, is one of the central questions of AI development economics. The optimistic case for leapfrogging points to the declining marginal cost of accessing frontier AI capabilities through API-based cloud services: a Kenyan entrepreneur can access GPT-4 level language models through the same API as a Silicon Valley startup, at the same price per token. If the binding constraint on AI adoption in developing economies is access to frontier AI rather than the ability to adapt it to local contexts, then cloud APIs could be a powerful leapfrogging mechanism.

The pessimistic case notes several countervailing forces. First, the data on which foundation models are trained is dominated by high-income country content, meaning that model performance degrades systematically for low-resource languages and cultural contexts — a form of algorithmic bias (算法偏见) that is not merely a fairness concern but a direct constraint on the utility of AI for non-English-speaking populations. Second, the APIs through which developing countries access frontier AI are priced in hard currency and subject to payment infrastructure requirements that many users in lower-income countries cannot meet. Third, the regulatory frameworks governing AI — including data protection law, liability rules, and sector-specific AI regulations — are being designed primarily in the US and EU, without meaningful participation from the Global South, potentially creating regulatory templates that are poorly adapted to lower-income country contexts. UNCTAD’s 2021 Digital Economy Report documents these asymmetries in considerable detail, finding that the US and China together accounted for over 70 percent of global AI patent filings and over 80 percent of top AI researchers, while the entire continent of Africa produced less than one percent of each.

Chapter 13: Leapfrogging, Mobile AI, and Development Pathways

13.1 Mobile AI and Financial Inclusion

The most compelling concrete examples of AI-enabled leapfrogging involve the integration of AI with mobile phone infrastructure in developing economies. The M-PESA mobile money system in Kenya is the foundational case: launched by Safaricom in 2007, it provided a mobile payment and banking system to millions of Kenyans who had never had access to formal banking, using the simple technological base of SMS messaging. By the time smartphone penetration and machine learning had matured, M-PESA’s successor systems were using AI-based credit scoring to extend microloans to millions of small businesses using behavioural signals from mobile phone usage patterns — purchase history, call patterns, geographic mobility — rather than the credit bureau records that do not exist for the majority of people in developing countries.

The economic significance of AI-based alternative credit scoring (替代信用评分) in developing economies is substantial. Traditional credit scoring requires a documented history of formal financial transactions — bank accounts, credit cards, formal employment — that the majority of people in low-income countries lack. Mobile-based AI credit scoring sidesteps this requirement by using behavioural signals that are available for anyone who owns a smartphone. Multiple peer-reviewed studies have found that mobile-based AI credit scores are predictive of default risk with accuracy comparable to traditional scores, suggesting that the new data sources are genuine substitutes for the conventional information, not merely noisier proxies.

13.2 AI and Agricultural Productivity

Agriculture employs the majority of the labour force in many lower-income countries, and AI applications in agricultural productivity represent one of the clearest potential development payoffs. Satellite imagery combined with machine learning models can provide crop health monitoring, yield prediction, and pest identification at spatial and temporal resolutions that were prohibitively expensive with ground-based methods. Weather forecasting AI is improving the accuracy and lead time of agricultural weather alerts that help smallholder farmers time planting and harvest decisions. AI-based soil analysis from mobile phone photographs is being piloted in India and sub-Saharan Africa as a cheap alternative to laboratory soil testing.

The empirical evidence on AI’s agricultural productivity effects in developing economies is still sparse — most published studies involve pilots with small samples — but the mechanistic case is strong. The information failures that suppress agricultural productivity in low-income countries — lack of access to timely weather information, poor knowledge of optimal planting practices for specific soil and microclimate conditions, inability to identify crop diseases before they spread — are precisely the kind of problems that AI prediction systems are designed to address. The scaling question is whether the satellite and mobile infrastructure needed to deliver these AI services can reach the smallholder farmers who stand to benefit most.

13.3 Brain Drain and Talent Flows

One of the most economically significant interactions between AI and development is the brain drain (人才流失) of AI talent from developing economies to the small number of high-income countries and major tech firms that dominate the AI industry. AI research is among the most internationally mobile forms of human capital: a machine learning PhD trained in India or Brazil can be hired by a US firm remotely or through immigration, receiving salaries that are an order of magnitude above what local institutions can offer. The result is a systematic outflow of the most AI-capable researchers from the countries that most need them.

The magnitude of this effect is difficult to quantify precisely, but surveys of AI researcher populations at major US universities and firms consistently show that a disproportionate fraction of AI talent was born or educated in developing countries — predominantly India and China. Whether this represents a genuine welfare loss for the origin countries (the researchers’ human capital was developed partly at public expense) or a form of gains from trade (through remittances, diaspora networks, and knowledge transfer) is contested, but the dominant economics view is that brain drain in high-value knowledge industries produces net negative externalities for origin countries when the public investment in higher education is not recovered through taxation or knowledge spillovers.

The concept of digital sovereignty (数字主权) — the aspiration of developing countries to maintain independent control over their AI infrastructure, data, and regulatory frameworks — is partly a response to these talent and technology dependency concerns. Countries including India, Brazil, and several African nations have developed or proposed policies that require local data storage, that condition market access on technology transfer, or that invest public resources in domestic AI research capacity. These policies face the tension between the efficiency gains from participating in the global AI ecosystem and the strategic autonomy gains from maintaining domestic AI capacity.

Part VI: Financial Systems and Macroeconomic Risk

Chapter 14: Algorithmic Trading, High-Frequency Markets, and Flash Crashes

14.1 The Microstructure of Algorithmic Financial Markets

The transformation of financial markets by algorithmic and high-frequency trading represents one of the earliest and most complete examples of AI displacing human labour in a high-skill, high-wage profession. By the mid-2010s, high-frequency trading (高频交易) firms accounted for the majority of equity market volume in the United States and a substantial share in European markets. The business model of HFT involves using computational speed advantages — co-located servers physically adjacent to exchange matching engines, fibre optic and microwave networks that shave microseconds off data transmission — to front-run slower market participants and capture the bid-ask spread.

The economic welfare analysis of HFT is contested. Proponents argue that HFT increases market liquidity — the ability of investors to transact at the quoted price without moving the market — by narrowing bid-ask spreads and increasing quote depth. Empirical studies do generally find that HFT activity is associated with reduced bid-ask spreads in normal market conditions, which benefits retail investors who transact at posted prices. Critics argue that the liquidity provided by HFT is phantom liquidity (幽灵流动性): it is available in normal conditions but evaporates precisely when it is most needed, during periods of market stress, because HFT algorithms are designed to withdraw when the signal-to-noise ratio of price discovery deteriorates.

14.2 Flash Crashes and Correlated Algorithm Failures

The most dramatic evidence for the systemic risk of algorithmic trading is the series of flash crashes (闪崩) that have occurred since the widespread adoption of algorithmic market-making. The May 6, 2010 flash crash, in which the Dow Jones Industrial Average fell nearly one thousand points in minutes before partially recovering, is the canonical event. The SEC-CFTC joint report attributed the crash to a large sell order (a Kansas-based mutual fund selling S&P 500 futures) being executed by an algorithm that did not account for the liquidity conditions in the market, triggering a cascade of algorithmic responses that temporarily withdrew liquidity and amplified the price decline.

The economic mechanism behind flash crashes is a form of correlated strategy failure (相关策略失效). When many market participants use similar algorithmic strategies, their responses to a given market signal are correlated in a way that amplifies shocks rather than absorbing them. The standard market microstructure model assumes that market participants have heterogeneous beliefs and strategies that cause their trades to offset each other in aggregate; correlated algorithmic strategies violate this heterogeneity assumption and produce market dynamics that can be highly unstable. The Bank for International Settlements has documented multiple flash crashes in foreign exchange markets, bond markets, and commodity markets since 2010, suggesting that the phenomenon is not isolated to equity markets but is a general feature of algorithmically intermediated financial markets.

14.3 AI and the New Generation of Trading Systems

Beyond HFT, which operates on millisecond timescales, the more recent wave of AI adoption in financial markets involves machine learning systems operating on human timescales — minutes, hours, and days — but making investment decisions of much greater complexity. These systems use natural language processing to analyse news and earnings call transcripts, computer vision to analyse satellite images of retail parking lots or shipping container movements, and reinforcement learning to adaptively construct trading strategies in response to changing market conditions.

The economic analysis of these systems raises different concerns from those associated with HFT. Because they operate on longer timescales and make more complex judgements, the correlated strategy failure risk is different in character — it is less a matter of millisecond liquidity withdrawal and more a matter of macro-level herding (宏观层面羊群效应) if many large funds adopt AI systems trained on similar data with similar architectures. The concern articulated by several central bank research departments is that if a large fraction of institutional investment AUM is managed by AI systems that share similar underlying logic, market prices could deviate persistently from fundamental values and correct suddenly and sharply when the shared logic is falsified by an unexpected event.

Chapter 15: AI in Credit, Robo-Advice, and Consumer Finance

15.1 AI in Credit Scoring: Efficiency vs. Equity

The application of AI to consumer credit scoring offers a clear illustration of the tensions between economic efficiency and distributional equity that run through many AI applications. Traditional credit scoring, exemplified by the FICO score, uses a relatively small number of variables — payment history, amounts owed, length of credit history, types of credit used, new credit applications — because those are the variables that the three major credit bureaus have historically collected. AI-based credit scoring can incorporate a much larger array of variables, including non-traditional signals such as mobile phone usage patterns, social network connections, browsing history, and even patterns in written communication.

The efficiency argument for AI credit scoring is straightforward: if non-traditional variables are predictive of credit risk, using them will reduce the rate of default by enabling lenders to distinguish good from bad credit risks more accurately. This should allow lenders to extend credit to some borrowers who would have been denied under traditional scoring (expanding financial inclusion) while pricing credit more accurately for those who receive it. Empirical studies in both developed and developing countries find that machine learning models using non-traditional data do predict default more accurately than FICO alone, and the improvement is largest for “thin file” borrowers — those with limited formal credit history.

The equity concern arises because many of the non-traditional variables that are predictive of credit risk are also correlated with protected characteristics (受保护特征) — race, gender, national origin — in ways that can produce disparate impact (差异性影响) on protected groups even without any intent to discriminate. If living in a predominantly Black neighbourhood is predictive of default risk (because neighbourhood is correlated with income, employment stability, and access to health insurance), an AI model that uses geolocation will discriminate against Black borrowers in ways that violate the Fair Housing Act and Equal Credit Opportunity Act, even if race is never explicitly included in the model. This is the challenge of proxy discrimination (代理歧视): the model does not use race, but it uses variables that are so highly correlated with race that the effect is nearly identical.

15.2 Robo-Advisors and the Democratisation of Wealth Management

Robo-advisors (机器人理财顾问) — automated investment management platforms that use AI to construct and rebalance portfolios based on client goals and risk preferences — represent one of the genuinely democratising applications of AI in financial services. Traditional human financial advisors are costly: their fees, typically 1–1.5 percent of assets under management annually, make professional investment advice economically inaccessible for individuals with less than roughly $100,000 in investable assets. Robo-advisors reduce this threshold dramatically, offering diversified portfolio management for fees of 0.25–0.5 percent of AUM with minimum investment requirements as low as zero.

The empirical evidence on whether robo-advisors generate better investment outcomes for their clients than traditional human advisors is mixed. Studies that compare robo-advised portfolios to human-advised portfolios find broadly similar risk-adjusted returns, which is consistent with both groups constructing well-diversified low-cost index portfolios — the investment strategy that decades of financial economics research has shown to be optimal for most retail investors. The real advantage of robo-advisors may not be in portfolio performance per se but in behavioural intervention (行为干预): robo-advisors are better at preventing clients from making the panic-driven buy-high-sell-low decisions that reduce the actual returns of human-advised retail investors.

15.3 Regulatory Frameworks for AI in Finance

The regulatory landscape for AI in financial services is evolving rapidly across jurisdictions. The core challenge is that existing financial regulation — the Securities Exchange Act, the Dodd-Frank Act, Basel III, the EU’s MiFID II — was designed around human decision-makers and human-scale decision processes. When investment decisions are made in microseconds by algorithms, when credit is extended based on features generated by machine learning models that no human fully understands, and when entire financial products are structured and priced by AI systems, the compliance frameworks designed for human decision-making become difficult to apply.

The Securities and Exchange Commission has begun developing guidance on the use of AI in investment advice, focusing on the disclosure obligations that AI-based advisors face and the standards for algorithmic fairness in credit and insurance. The Financial Conduct Authority in the United Kingdom has taken a principles-based approach, requiring firms using AI in regulated activities to demonstrate that the AI meets the FCA’s existing conduct standards — including suitability, fairness, and explainability — without specifying particular technical approaches. The Basel Committee on Banking Supervision has addressed AI in the context of model risk management, extending existing model validation frameworks to ML models while acknowledging that the opacity of deep learning models poses challenges for standard model validation practices.

Chapter 16: Systemic Risk, Central Banks, and the Governance of AI in Finance

16.1 Macroprudential Perspectives on AI

The macroprudential framework — the approach to financial regulation that focuses on risks to the financial system as a whole, rather than to individual institutions — is the appropriate lens for analysing the systemic risks of AI in finance. Individual financial institutions adopting AI make decisions that are rational from their own perspective, but the collective result may be a financial system that is more fragile than any individual institution’s risk management would suggest. This is the canonical systemic risk externality (系统性风险外部性): each institution imposes risks on the broader system that it does not fully internalise.

AI in finance creates at least three distinct channels for systemic risk externalities. First, the correlated model channel (相关模型渠道): if a large fraction of financial institutions use AI models trained on similar data with similar architectures, their risk assessments and trading decisions will be correlated, increasing the likelihood that many institutions simultaneously take losses in a scenario where the shared model is wrong. Second, the opacity channel (不透明渠道): AI models are harder to audit and validate than rule-based systems, increasing the probability that model errors are not detected until they cause significant losses. Third, the speed channel (速度渠道): AI enables faster execution of financial decisions, which can accelerate the transmission of shocks through the financial system and reduce the time available for regulatory intervention.

The BIS’s 2019 Annual Economic Report on big tech in finance identifies a related concern: the entry of large technology platforms into financial services creates a new category of systemically important institution that may not be captured by existing prudential frameworks. A firm like Alibaba’s Ant Financial or Apple Card has access to customer relationships, data, and distribution networks that give it structural advantages over traditional banks, but it may be regulated by a different framework (payments regulation, data protection law) that does not include the capital adequacy and liquidity requirements that apply to systemically important banks.

16.2 Central Bank AI and Macroeconomic Surveillance

Central banks are themselves major adopters of AI, using machine learning systems for macroeconomic surveillance (宏观经济监测) and financial stability monitoring. Natural language processing systems monitor news flows, social media, and financial market commentary to detect early signals of financial stress or economic sentiment shifts. Machine learning models are used to construct more accurate near-term forecasts of inflation, output, and employment than traditional econometric models, particularly during periods of structural change when historical relationships may not be stable. Network analysis algorithms map the interconnections among financial institutions to identify systemically important linkages that are not visible in standard bilateral exposure data.

The Federal Reserve, the European Central Bank, and the Bank of England have all invested significantly in AI for surveillance, though the public documentation of these systems is limited for reasons of financial market sensitivity. The IMF has developed its own AI surveillance tools and provides technical assistance to member countries’ central banks in developing similar capacities. The governance of AI in central banking raises distinct questions: because central bank actions can move financial markets, the opacity of AI-assisted decision processes is particularly consequential, and there are democratic accountability concerns about the degree to which monetary policy decisions may be delegated to algorithmic systems.

16.3 Regulatory Responses and the Governance Gap

The governance gap between the pace of AI adoption in financial services and the pace of regulatory development is widely acknowledged by both regulators and industry participants. The technical complexity of AI systems makes them difficult for non-specialist regulators to evaluate; the pace of AI development means that regulations designed for today’s systems may be obsolete before they take effect; and the global character of financial markets means that nationally-based regulation can be circumvented by regulatory arbitrage.

Several principles are emerging from the regulatory discussions of the major jurisdictions. Explainability (可解释性) — the requirement that AI-based financial decisions can be explained to affected parties in terms they can understand — is consistently identified as a priority, though the technical definition of explainability and the appropriate level of explanation for different use cases remain contested. Auditability (可审计性) — the ability for regulators and internal compliance functions to examine how AI systems make decisions — is seen as a prerequisite for meaningful oversight. Model governance (模型治理) frameworks, adapted from existing model risk management practices in banking, are being extended to cover the full lifecycle of AI systems in financial applications, from design and training through deployment and ongoing monitoring. PHIL 451 (AI Ethics, Law, and Governance) examines the regulatory frameworks for AI more broadly; in the financial context, the distinctive feature is the direct systemic risk externality that makes governance a macro-prudential as well as a consumer protection concern.

Part VII: Distributional Consequences and Policy Responses

Chapter 17: Inequality, Rents, and the Political Economy of AI

17.1 The Mechanics of AI-Driven Inequality

The distributional consequences of AI operate through multiple channels that must be distinguished carefully to assess their relative magnitudes and policy implications. Korinek and Stiglitz, in their 2017 NBER working paper, provide a taxonomy of the mechanisms through which AI could increase inequality. The first is the labour displacement channel (劳动力替代渠道): AI displaces workers from tasks, reducing their wages or employment. The distributional impact depends on which workers are displaced — if primarily low-skill workers, inequality rises; if primarily high-skill workers, it may fall — and whether the displaced workers find alternative employment at comparable wages.

The second is the capital income channel (资本收益渠道): AI is embodied in capital, and when AI raises the return to capital by replacing labour in production, the income share of capital rises relative to the income share of labour. Since capital is far more unequally distributed than labour income in all advanced economies — the top quintile of the wealth distribution in the US owns approximately 85 percent of all financial assets — a rising capital share mechanically increases inequality in the distribution of income.

The third is the superstar firm channel (超级明星企业渠道): AI creates and sustains winner-takes-all markets in which a small number of highly productive firms capture a disproportionate share of industry revenue. Since the incomes of workers at superstar firms are also elevated by the firm’s rents — through efficiency wages, profit-sharing, and sorting of high-skill workers — the concentration of activity in superstar firms translates into concentration of wage income as well as capital income. Autor, Dorn, Katz, Patterson, and Van Reenen document the rise of superstar firms in the US economy and find that it accounts for a significant fraction of the observed increase in the labour income share accruing to top earners.

17.2 The Political Economy of AI Adoption

The political economy of AI is shaped by the fact that the gains from AI are concentrated among capital owners and high-skill workers while the losses are concentrated among middle-skill workers in specific occupations and geographic regions. This distributional asymmetry creates political tensions that are already manifesting in debates about automation taxation, worker retraining programmes, platform regulation, and trade policy.

The political economy of adjustment (调整政治经济学) — how societies manage the transition costs of technological change — has been studied extensively in the trade literature, where it is known that the adjustment costs of trade-displaced workers are large, persistent, and concentrated in specific communities, while the gains from trade are diffuse and economically invisible to most voters. Autor, Dorn, and Hanson’s “China Shock” research finds that communities exposed to import competition from China experienced not only economic losses but also political polarisation — higher vote shares for extreme candidates of both parties — suggesting that the failure to compensate trade-displaced workers has political as well as economic consequences. The structural similarity between trade displacement and automation displacement implies that similar political dynamics could follow from AI-driven job losses in robotics-exposed communities.

17.3 Rents and the Concentration of AI Income

A distinctive feature of AI-driven income concentration is its dependence on economic rents (经济租金) — incomes that exceed the minimum necessary to attract the relevant resources into their current use. The profits of AI platform firms substantially exceed the competitive return on their invested capital; the salaries of AI researchers at frontier firms substantially exceed what the same individuals could earn in alternative employments; and the returns to the data assets of dominant platforms exceed the cost of acquiring them. These rents are the product of market power rooted in network effects, data moats, and the winner-takes-all dynamics analysed in Part III.

The rent character of AI income concentration matters for policy because rents can, in principle, be taxed without distorting economic activity — this is the fundamental theorem of rent taxation, traceable to Henry George’s land value tax argument. A tax on the rents of AI platform companies, if properly designed, would not reduce the incentive to invest in AI development (because the investment would still earn the competitive return after tax) but would redistribute the above-normal returns from capital owners to society broadly. The practical challenges of rent taxation in the AI context are considerable — identifying and measuring rents separately from competitive returns is difficult, and firms have substantial ability to minimise taxable income through transfer pricing and jurisdictional arbitrage — but the theoretical case for targeting rents rather than normal profits is well-established in public economics.

Chapter 18: Policy Toolbox: UBI, Retraining, Compute Taxation, and Beyond

18.1 Universal Basic Income and the Automation Dividend

Universal Basic Income (全民基本收入), or UBI, has emerged as one of the most discussed policy responses to the distributional risks of AI-driven automation. In its most common formulation, UBI provides an unconditional cash transfer to all adults regardless of employment status, replacing or supplementing the patchwork of means-tested welfare programmes that characterise most advanced economy safety nets. The economic case for UBI in the AI context rests on two arguments: first, that the rate of displacement may be too rapid for existing retraining and adjustment programmes to handle, creating a need for income support that is not conditioned on job-seeking behaviour; and second, that the rents generated by AI and automation constitute a social dividend that should be distributed broadly rather than captured solely by capital owners.

The economics of UBI are complex and the empirical evidence from pilot programmes — including Finland’s 2017–2018 experiment and various US municipal pilots — is useful but limited in scope. The Finnish experiment, which provided €560 per month to two thousand unemployed individuals for two years, found modest positive effects on wellbeing and labour market attachment, but the sample was too small and the benefit level too modest to provide definitive evidence on the macro-level effects of a universal programme. The critical economic parameters — the labour supply elasticity of UBI recipients, the fiscal cost of the programme at scale, and the macroeconomic effects of the transfer — remain highly uncertain, and the estimated effects are sensitive to these parameters.

The funding mechanism for UBI is also economically consequential. If UBI is funded by consolidating existing welfare programmes (as in Milton Friedman’s negative income tax proposal), it may actually reduce the protection available to the most vulnerable workers if the UBI amount falls below the effective benefit level they currently receive. If it is funded by new taxes — on capital income, on platform firm revenues, or on automation equipment — the funding mechanism itself has distributional consequences that must be factored into the welfare analysis.

18.2 Worker Retraining and the Limits of Human Capital Policy

Worker retraining (工人再培训) programmes are the most politically palatable policy response to automation displacement, and most advanced economy governments have made them a central plank of their AI policy response. The appeal is intuitive: if AI eliminates certain jobs, policy should help workers acquire the skills to do the jobs that AI cannot. This logic underlies a wide range of programmes from community college subsidies and lifelong learning accounts to sector-specific apprenticeship schemes and on-the-job training tax credits.

The empirical evidence on the effectiveness of retraining programmes is, however, sobering. A large body of evaluation literature — much of it focused on the US Trade Adjustment Assistance programme, which provides retraining to workers displaced by international trade — finds that the average programme participant earns less in the years following training than a comparable non-participant, net of the forgone earnings during training. The reasons are several: retraining programmes frequently train workers for occupations that are already in decline by the time the curriculum is designed; the duration of programmes is often insufficient to build genuine competence in new technical domains; and older workers face steeper human capital depreciation that reduces the expected return to retraining investment.

These findings do not imply that retraining is never effective — there are programme designs and target populations for which the evidence is more positive — but they do suggest that retraining cannot be the principal policy instrument for addressing the distributional consequences of AI automation. The scale of the adjustment challenge implied by serious AI displacement scenarios — potentially tens of millions of workers requiring occupational transition over a decade — exceeds the demonstrated capacity of existing retraining infrastructure by a wide margin.

18.3 Automation Taxation and Compute Levies

Automation taxation (自动化税) — levying taxes on the automation of tasks previously performed by human workers — has been proposed by economists including Daron Acemoglu and by policy advocates including Bill Gates in a widely discussed 2017 interview. The economic rationale is that current tax systems are biased in favour of automation because they tax labour (through payroll taxes and social insurance contributions) but not capital equipment at the same effective rate. This tax bias distorts the automation decision margin: firms may find it privately profitable to automate tasks even when the social benefit of automation (including the value of the jobs displaced) does not exceed the social cost.

Formally, the distortion can be characterised using the standard automation decision condition. A firm automates a task when the cost of the automation technology falls below the cost of labour for that task:

where \( p_{K_{AI}} \) is the user cost of AI capital, \( w \) is the wage, and \( \tau_L \) is the effective labour tax wedge. An automation tax \( \tau_{AI} \) on AI capital would modify this to:

effectively raising the threshold below which automation is chosen and preserving employment in tasks that are only marginally cheaper to automate than to perform with labour. The revenue from the automation tax could in principle fund retraining programmes or UBI, creating an automatic stabiliser that grows with the pace of automation.

The principal objection to automation taxation is the concern that it will slow beneficial innovation — that taxing automation will reduce the incentive to invest in AI development and delay the productivity gains from which broader social improvements should ultimately flow. Acemoglu acknowledges this concern but argues that if the current direction of AI investment is already biased toward labour-replacing automation and away from task-complementing innovation, an automation tax may actually improve the direction of innovation by reducing the private profitability of displacement-heavy AI relative to augmentation-focused AI.

Compute taxation (算力税) is a more radical variant that targets the underlying resource that enables AI development. Large AI training runs consume enormous amounts of electricity and specialised semiconductor resources; a tax on these inputs could raise revenue that could fund public AI research or safety work while also — through the price signal — nudging AI development toward more computationally efficient approaches. The design challenges are significant: compute is an intermediate input with many legitimate non-AI uses, and a broad compute tax would impose costs on activities far removed from AI automation. Targeted levies on large training runs above a compute threshold (measured in petaflop-days, for example) would be more precise but also more easily gamed by distributing training across nominally separate runs.

18.4 Regulatory Responses: Antitrust, Data Governance, and the Public Option

Beyond fiscal instruments, the policy toolbox for managing AI’s distributional consequences includes structural regulatory approaches. Antitrust enforcement against AI platforms — limiting acquisitions, requiring interoperability, mandating data sharing — can reduce the rent capture of dominant firms and lower barriers to competitive entry, which should over time increase the competitive pressure to share AI productivity gains with workers and consumers. The DMA in Europe and renewed antitrust activity in the United States represent the current frontier of this approach.

Data governance (数据治理) frameworks — including data protection law, data portability requirements, and rights to algorithmic explanation — can shift bargaining power between data subjects and the firms that use their data, potentially improving the terms of the implicit data exchange. The GDPR in Europe and various US state-level privacy laws represent the current state of play, though most economists who study data markets find that existing governance frameworks are poorly designed to address the competition and distributional concerns associated with AI data accumulation.

A more radical proposal, advanced by economists including Daron Acemoglu and by policy think tanks including the Roosevelt Institute, is the public option in AI (人工智能公共选项): the provision of foundational AI capabilities as a public good by government or quasi-public institutions, analogous to the role played by public universities, public research laboratories, and public utilities in earlier phases of economic development. A public option in AI — whether in the form of publicly funded foundational models, publicly owned compute infrastructure, or public datasets — would reduce the dependence of downstream AI applications on privately owned monopoly infrastructure and could be designed to prioritise social welfare objectives over profit maximisation. The political economy of such a proposal is challenging in the US context, but analogues exist in other countries: France’s Mistral AI has partial public support; Canada’s Canadian Institute for Advanced Research (CIFAR) hosts national AI institutes; and multiple European governments have funded national AI research infrastructure.

18.5 A Political Economy of Reform

The fundamental challenge for AI policy is that the actors who stand to gain most from AI-driven economic change — large technology firms, capital owners, high-skill workers — also have disproportionate political influence. The actors who stand to lose most — middle-skill workers in automatable occupations, communities dependent on manufacturing and routine cognitive industries, populations in developing countries whose data is extracted without compensation — are politically weaker and face collective action problems in organising to demand redistribution.

This political economy suggests that effective AI policy will not emerge spontaneously from the play of market forces. It will require deliberate political choices to internalise the externalities of AI displacement, to tax the rents of AI-enabled market power, to invest publicly in the foundational AI research that private markets underprovide, and to govern the global AI ecosystem in ways that give all populations a stake in its benefits. The history of previous general-purpose technology transitions — industrial capitalism, electrification, the ICT revolution — suggests that these political choices can be made, and that when they are made well, the result is a more broadly shared prosperity than the unconstrained market would produce. The study of the economics of AI is ultimately the study of what those choices are and what considerations should guide them.

HIST 415 (A History of Artificial Intelligence) traces the institutional and political choices that shaped earlier phases of computing development, from the wartime science mobilisation that produced the first computers to the Cold War research funding that sustained AI through its long winters. PHIL 451 (AI Ethics, Law, and Governance) develops the normative frameworks — from utilitarian welfare analysis to Rawlsian distributive justice and capabilities approaches — that can guide the political choices that the economics alone cannot determine. Together, the three courses provide a foundation for thinking seriously about one of the central challenges of the coming decades: how to govern a technology whose economic consequences are as profound as those of any in human history.