In March 2023, Goldman Sachs economists Jan Hatzius and Joseph Briggs published an internal research note that circulated far beyond Wall Street. Their conclusion: generative AI could expose roughly 300 million full-time jobs globally to automation. Within weeks, the paper was cited in every major newspaper, dozens of congressional hearings, and more LinkedIn posts than anyone could count.
The number was arresting. It was also widely misread. The Goldman economists did not say 300 million jobs would disappear — they said those jobs contained tasks that AI could plausibly perform. The distinction matters enormously, and it's where the economic analysis actually starts.
Since the early 2000s, economists David Autor, Frank Levy, and Richard Murnane developed what is now called the ALM framework: the idea that technology displaces specific tasks, not entire occupations. Routine tasks — those that can be described by a finite set of rules — are the most vulnerable. Non-routine tasks, particularly those requiring physical dexterity in unpredictable environments or complex social judgment, are the most protected.
Generative AI has complicated this framework in one crucial way: it extended automation pressure from routine manual tasks (already hit by earlier robotics) into routine cognitive tasks — the kind performed by knowledge workers. Drafting standard contracts, producing first-pass financial summaries, answering tier-1 customer queries, generating code from specifications: these are now contested territory.
The clearest real-world cases come from sectors where AI was deployed at scale before 2024. Customer service was the first visible battleground. Klarna, the Swedish buy-now-pay-later firm, announced in February 2024 that its AI assistant — built on OpenAI technology — was handling the equivalent workload of 700 human customer service agents, resolving 2.3 million conversations in its first month with customer satisfaction scores matching human agents. The company simultaneously announced it had cut its global headcount from roughly 5,000 to 3,800 over the prior year.
In software development, GitHub Copilot reached 1.8 million paid users by early 2024. A controlled study by GitHub found developers completed tasks 55% faster with Copilot assistance. What's notable is what didn't happen: junior developer hiring at large tech firms dropped sharply in 2023–2024, while senior developer demand remained strong. The implication is that AI is compressing the value of entry-level coding tasks while amplifying senior engineers who can review, direct, and integrate AI output.
In legal services, firms including Allen & Overy deployed Harvey AI for contract review and legal research. The documented impact: due diligence reviews that took paralegals 10–12 hours were completed in under 2. The firm did not immediately lay off paralegals; instead it reduced paralegal hiring targets and began redeploying existing staff to higher-complexity work. This pattern — reduced hiring rather than mass layoffs — is now considered the typical first-wave impact.
In nearly every documented first-wave case, AI automation first reduces new hiring rather than triggering immediate layoffs. The workforce shrinks through attrition. This makes displacement slower, harder to see in quarterly unemployment figures, and more difficult for workers to anticipate.
MIT economist David Autor's 2024 working paper "Applying AI to Rebuild Middle Class Jobs" documented a counterintuitive finding: white-collar office work is more exposed to generative AI than blue-collar physical work. Occupations scoring highest on AI exposure include financial analysts, legal assistants, medical coders, copywriters, data entry clerks, and customer service representatives.
Occupations scoring lowest include electricians, plumbers, construction workers, childcare workers, and physical therapists. These jobs require real-time sensory judgment and physical manipulation in environments too variable for current robotic systems. The McKinsey framework calls this the "proximity paradox" — high-paying knowledge jobs are more immediately exposed than lower-paying physical jobs, inverting the usual pattern of technological displacement.
| Sector | AI Exposure Level | Documented Action (2023–24) |
|---|---|---|
| Customer Service | Very High | Klarna: 700 agent-equivalent AI; headcount –24% |
| Legal (paralegal tasks) | High | Allen & Overy Harvey AI; 80% time reduction on review |
| Software Dev (junior) | High | GitHub Copilot; 55% task completion speedup |
| Media / Copywriting | High | BuzzFeed, Sports Illustrated AI content incidents |
| Radiology (screening) | Moderate–High | FDA-approved AI screening tools at 700+ US hospitals |
| Construction Trades | Low | Robot bricklayers exist but not economically competitive |
| Childcare / Elder Care | Very Low | No significant deployment; regulatory and trust barriers |
Every previous general-purpose technology — electricity, the internet — eventually created more jobs than it displaced, though the transition took decades and the new jobs were often in entirely different sectors. Early evidence for AI is consistent with this pattern: LinkedIn reported that job postings mentioning "prompt engineering" grew 300% in 2023. Roles including AI trainer, AI auditor, model evaluator, and AI integration specialist were virtually nonexistent in 2020.
The World Economic Forum's 2023 Future of Jobs report estimated that AI would displace 83 million jobs globally by 2027 while creating 69 million new ones — a net displacement of 14 million. Notably, the fastest-growing roles include green-energy technicians, data analysts, and AI/machine learning specialists: roles that require humans to work with AI, not in competition with it.
The economic evidence so far suggests AI is best understood not as a job-eliminator but as a task-redistributor: it compresses the value of routine cognitive tasks while amplifying the value of judgment, synthesis, and human-facing skills. The aggregate outcome — more jobs or fewer — depends heavily on whether productivity gains are reinvested in ways that generate new demand.
In this lab you'll work with an AI tutor to apply the task-level automation framework to specific real-world jobs. Think through which tasks within an occupation are routine vs. non-routine, and what "high AI exposure" actually means in practice.
In April 2024, MIT economists Shakked Noy and Whitney Zhang published the results of a randomized controlled trial — one of the most rigorous AI productivity studies conducted outside a lab. They recruited 444 college-educated professionals working in marketing, consulting, and data analysis and gave half of them access to ChatGPT. The AI-access group completed tasks 37% faster and produced output that blind evaluators rated 18% higher quality. Workers in the bottom half of the performance distribution gained the most — AI compressed the skill gap between mediocre and excellent performers.
The study's most important finding was what it didn't measure: what happened to wages, or headcount, or how the firm used the freed-up time. Productivity gains inside a controlled experiment don't automatically become wages in the real economy. That translation depends on market structure, labor bargaining power, and how competitive the industry is.
Economic historians have documented that the timing between productivity gains and wage gains from general-purpose technologies can span decades. The electrification of American factories between 1890 and 1920 raised manufacturing productivity substantially, but real manufacturing wages didn't fully reflect those gains until the 1940s — after labor organizing, New Deal legislation, and wartime demand reshaped bargaining dynamics.
The internet era provides a more recent cautionary tale. US labor productivity grew 2.6% annually from 1996–2004 during the dot-com boom. Real median wages grew 0.8% annually over the same period. The gap — sometimes called the "productivity-pay divergence" — reflected rising corporate profit shares, executive compensation, and returns to capital rather than broad wage increases. Economist Lawrence Mishel at the Economic Policy Institute documented that the productivity-pay gap has widened in every technology wave since 1979.
In the 2024 MIT RCT, workers with ChatGPT access completed tasks 37% faster and produced 18% higher-quality output. Critically, the performance boost was largest for lower-skilled workers — AI compressed the gap between median and top performers rather than amplifying it.
When AI raises productivity, the gains flow through three possible channels. First, lower prices: in competitive markets, productivity gains get passed to consumers as firms underprice each other. Legal services, financial advice, and software products may all get cheaper as AI lowers the cost of production. Second, higher wages: workers who are more productive may be able to bargain for more, especially in tight labor markets or when they have specialized AI skills. Third, higher profits: firms with market power can capture productivity gains as profit rather than sharing them with workers or customers.
The distribution across these three channels depends critically on market concentration. Industries already dominated by a few large players — cloud computing, search, enterprise software — are well-positioned to capture AI productivity gains as profit rather than distribute them through wages or prices. The AI investment race itself is highly concentrated: in 2023, three companies (Microsoft, Google, and Meta) accounted for a majority of AI infrastructure spending in the United States.
Economists David Autor, David Dorn, Lawrence Katz, Christina Patterson, and John Van Reenen documented what they called the "superstar firm" effect in a 2020 paper: as certain industries became winner-take-most markets, the surviving dominant firms captured increasingly large revenue shares while labor's share of income in those industries fell. AI may accelerate this dynamic.
The mechanism: AI allows the most capable firms to scale their productivity advantages more rapidly because AI capabilities scale with compute investment, and large firms have disproportionate access to capital. A regional law firm and a Big Law firm both get access to Harvey AI — but Big Law can layer AI onto client relationships, proprietary data, and existing brand trust in ways that amplify their competitive advantage rather than eroding it.
There is one documented channel through which AI productivity gains do reach workers relatively quickly: internal labor markets at firms that explicitly share gains. A 2023 study of call center workers at a large US software firm (the company has not been named but the study was conducted by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond) found that AI assistance raised the productivity of new and low-skilled workers by 35%. The firm's incentive structure paid workers partly on volume — so productivity gains translated directly to wages for those workers, without requiring any renegotiation of contracts.
This suggests the institutional structure around AI deployment matters as much as the technology itself. Piece-rate or productivity-linked compensation allows gains to flow automatically to workers. Fixed-wage structures do not — and fixed wages are the norm for most knowledge workers.
AI's productivity gains are real and documented. Whether those gains translate to broad wage increases, lower consumer prices, or concentrated corporate profits depends on market structure, labor bargaining power, and institutional design — not on the technology itself. History suggests we should not assume automatic broad distribution.
Use this lab to work through the distribution question for specific industries. Given what you know about AI productivity gains and market structure, analyze which channel — wages, prices, or profits — is most likely to capture gains in industries you care about.
In 2023, the Brookings Institution's Mark Muro and Sifan Liu published an analysis mapping AI exposure across US metropolitan areas. Their finding was stark: the cities best positioned to benefit from AI are overwhelmingly the same cities that have already benefited most from the digital economy. San Jose, San Francisco, Seattle, Austin, and Boston — already wealthy knowledge-economy hubs — scored highest on both AI investment concentration and workforce AI readiness. Toledo, Youngstown, Scranton, and Fresno scored lowest on both.
The mechanism is self-reinforcing: AI talent concentrates where AI investment is, AI investment concentrates where AI talent is, and the regulatory and educational infrastructure that produces AI-ready workers takes decades to build. The geography of AI benefit is not a technical question — it's a political economy question.
The most consistent finding across labor economics research on AI is an education gradient in both risk and opportunity. Workers without four-year college degrees are concentrated in physically demanding service jobs — retail, food service, transportation — that face lower near-term AI exposure but also fewer AI-augmentation opportunities. Workers with graduate degrees are in the highest-exposure but also highest-augmentation potential roles.
The nuance matters: exposure is not the same as harm. A radiologist with high AI exposure may find their productivity augmented 5x. A data entry clerk with high AI exposure may find their role eliminated. The difference is whether the occupation has sufficient non-routine cognitive components that AI augments rather than replaces.
MIT economist David Autor's 2024 analysis documented that AI could potentially increase the earning power of workers without four-year degrees — if AI tools are deployed to augment skilled trades, health support roles, and technician positions that currently require expensive credentialing. But he emphasized this outcome requires deliberate policy and training investment; it will not happen automatically.
The international dimension of AI inequality is potentially the most severe. Countries in Sub-Saharan Africa, South Asia, and Latin America built economic development strategies around manufacturing export growth and business process outsourcing (BPO) — the same sectors most exposed to AI automation.
The Philippines built a $30 billion BPO sector employing over 1.3 million workers in customer service, data processing, and content moderation — the exact roles now being automated by AI. McKinsey's 2023 analysis estimated that AI could displace 30–50% of BPO employment globally over the next decade. For countries where BPO represents 5–8% of GDP and provides the primary source of middle-class employment, this is not a sectoral adjustment story — it's an economic crisis risk.
Meanwhile, the countries that develop AI — the US, China, UK, Canada — capture the intellectual property value and platform rents. The IMF's January 2024 report "AI Will Transform the Global Economy" found that advanced economies are roughly 60% more AI-ready than emerging economies, and that without intervention, AI is likely to widen global income inequality between nations.
The IMF's "AI Will Transform the Global Economy" report found that nearly 40% of global employment is exposed to AI. In advanced economies, about 60% of that exposed employment may be augmented. In emerging economies — which have less complementary infrastructure — the proportion facing displacement rather than augmentation is significantly higher.
Within the United States, AI exposure maps onto pre-existing demographic inequalities in complex ways. The National Bureau of Economic Research's 2023 working paper by Zanele Munyikwa, Emma Mishel, and colleagues found that Black and Hispanic workers are overrepresented in high-exposure low-augmentation roles — particularly in administrative support, customer service, and data processing. White and Asian workers are more often in high-exposure high-augmentation roles — software development, financial analysis, management consulting.
The gender dimension shows a distinct pattern: women are disproportionately employed in clerical, administrative, and customer service roles — historically the most routine-cognitive work — giving them higher AI displacement exposure than men on average. However, women are also concentrated in care-sector roles (nursing, education, social work) that are among the least AI-automatable, creating a bimodal distribution of risk and relative security.
Several documented policy experiments address geographic and demographic AI inequality. Finland's universal basic income experiment (2017–2018), while pre-dating the generative AI wave, demonstrated that unconditional income support improved wellbeing and did not reduce work effort, providing evidence for UBI as a distributional buffer. Canada's 2023 AI and Society Fund allocated C$200 million specifically to AI adoption support for small and medium enterprises in non-hub cities. The EU's AI Act mandates impact assessments for high-risk AI deployments, with explicit attention to demographic disparate impact.
In the United States, the Biden administration's 2023 Executive Order on AI included workforce provisions directing federal agencies to develop AI training programs targeting displaced workers, but allocated no dedicated new federal funding. The scale of policy response has not matched the scale of the transformation documented by economists.
AI will raise aggregate output. The question is not whether there are gains to distribute — there are — but who has the political and market power to claim them. Geography, education, race, gender, and national income level all predict position in the distribution queue. Aggregate gains do not resolve distributional questions; they only raise the stakes of those questions.
Use this lab to analyze a specific region, country, or demographic group and work through their AI economic exposure — both risks and opportunities. Apply the frameworks from the lesson to generate grounded analysis.
In January 2024, the International Monetary Fund released its landmark "AI Will Transform the Global Economy. Let's Make Sure It Benefits Humanity." The document was notable not for its economic projections — those tracked closely with earlier McKinsey and Goldman estimates — but for its explicit call for proactive redistribution policy. "In most scenarios," IMF Managing Director Kristalina Georgieva wrote, "AI will likely worsen overall inequality... It is therefore critical that countries put in place strong social safety nets and retraining programs."
The policy debate that followed mapped onto a familiar ideological landscape: market optimists argued that productivity gains would create enough new wealth to fund redistributive programs voluntarily; labor advocates argued that historical evidence shows this never happens without legal mandate; and a third camp — perhaps most credibly grounded in economic evidence — argued that the specific form redistribution takes matters as much as whether it happens at all.
Retraining is the most politically popular policy response and has the weakest evidence base for large-scale effectiveness. The Trade Adjustment Assistance (TAA) program, which provided retraining funds to US workers displaced by trade, was studied extensively. A 2012 analysis by Kara Reynolds and a 2019 study by economists at the Federal Reserve found that TAA participants were no more likely to be employed four years after displacement than non-participants, and had lower wages on average — partly because the program provided extended income support that allowed workers to take longer training but then re-entered slower job markets.
More successful are sectoral retraining programs that partner with specific employers to train workers for known job openings. The Washington State Workforce Training and Education Coordinating Board documented that sector-based programs had significantly better employment and wage outcomes than general skills training. Germany's dual apprenticeship system — heavily employer-integrated — is the most-cited international success case, maintaining low youth unemployment even through periods of industrial disruption.
General retraining programs have poor evidence of effectiveness. Sectoral programs with direct employer partnerships have significantly better outcomes. The scale of AI displacement is projected to exceed the capacity of any retraining system currently funded in any country.
Universal Basic Income (UBI) has been piloted in at least eight countries since 2016. The most rigorous evidence comes from Finland's 2017–2018 experiment, which provided €560 per month to 2,000 unemployed workers unconditionally. Published results: recipients reported significantly higher wellbeing and mental health, marginally higher employment rates than the control group (counter to critics' predictions), and greater willingness to take entrepreneurial risk. The Finnish government ended the experiment on schedule and did not expand it, citing cost.
Kenya's GiveDirectly UBI program — the largest randomized controlled trial of cash transfers in history, covering 20,000 recipients — found positive effects on consumption, assets, and local economic activity, with no evidence of reduced labor supply. Stockton, California's SEED program (2019–2021) provided $500/month to 125 low-income residents; recipients showed higher full-time employment rates than the control group at 12 months.
The economics case for UBI as an AI response centers on transition support: providing income security during labor market transitions reduces the pressure to take any available job and allows workers to pursue better-matched retraining. Critics note the cost: a US UBI of $1,000/month for all adults would cost approximately $3 trillion annually, exceeding the current federal discretionary budget.
Bill Gates publicly endorsed a robot tax in a 2017 Quartz interview, arguing that if a robot displaces a worker, it should pay payroll taxes equivalent to what the worker would have paid. The argument is intuitive: the current tax system is optimized for labor income, so as capital displaces labor, the tax base shrinks even as the need for social support grows.
The European Parliament voted in 2017 against a proposed robot tax resolution, with critics arguing it would slow AI adoption and investment. South Korea implemented a partial version in 2017: it eliminated a tax incentive for companies investing in automation, effectively making automation slightly more expensive — the only enacted policy globally that approximates a robot tax.
Economic analysis by MIT economists Daron Acemoglu and Pascual Restrepo argues that the US tax code already implicitly subsidizes automation over labor by taxing labor income (payroll taxes) but not capital investment. Their 2020 paper found that this tax asymmetry has meaningfully contributed to labor displacement by making automation economically attractive even in cases where it provides modest productivity gains.
The EU AI Act (fully in effect 2025) takes a risk-tiered regulatory approach: banning certain AI applications (social scoring, real-time biometric mass surveillance in public spaces), requiring human oversight for high-risk applications (credit decisions, employment screening, medical devices), and imposing transparency requirements on general-purpose AI systems. The Act's labor provisions are limited but include protections against purely algorithmic employment decisions.
At the collective bargaining level, several US unions have successfully negotiated AI protections into contracts. The Writers Guild of America's 2023 contract after its 148-day strike established rights over AI use in writing — AI cannot be used to generate script content that replaces union writers, and existing material cannot be used to train AI without consent. The Screen Actors Guild negotiated residual payment rights when AI likenesses are used commercially. These contracts represent the first systematic application of collective bargaining to AI labor questions.
Denmark's "flexicurity" model — combining easy hiring and firing with robust unemployment support and active labor market programs — is frequently cited as the best existing institutional structure for navigating technology displacement. Danish workers displaced by automation receive up to 90% of prior wages for up to two years while in active labor market programs, funded by a 2% GDP expenditure on labor market policy (compared to 0.1% in the United States).
Synthesizing the evidence across policy approaches, a few conclusions stand out. First, no single intervention is sufficient — the scale and speed of AI-driven labor market transformation exceeds what any individual policy can address. Second, institutional structures matter more than individual programs — Denmark's flexicurity works because it's a coherent system, not a collection of individual programs. Third, employer involvement is critical — retraining programs without employer partnerships consistently underperform. Fourth, redistribution must be funded — the mechanisms that generate AI productivity gains (capital, IP, platform rents) are the natural tax base for redistribution, but taxing them requires political will that has been consistently absent.
Sectoral retraining with employer partnerships: strongest employment outcome evidence. UBI-style income support: strong wellbeing and transition evidence, prohibitive cost at scale. Robot/capital taxes: theoretically sound, politically and practically difficult. Collective bargaining AI provisions: emerging but limited to organized sectors. Denmark-style flexicurity: best-evidenced systemic approach, requires 20x US current labor market investment.
Use this lab to design or evaluate policy responses to specific AI economic disruption scenarios. Apply the evidence from the lesson — what works, what doesn't, and why — to build a credible policy argument for a specific context.