The steam engine eliminated the occupation of water carrier and created railroad engineer and factory foreman. The telephone eliminated telegraph operators and created phone-company executives. The personal computer eliminated typing pools and created software engineers. Every major technology has done this — and every time, the people whose work was displaced absorbed the majority of the cost while the gains concentrated elsewhere.
AI is the next round. The transition is compressed: tasks that were safe a year ago are automatable now, and the definition of a knowledge worker is shifting faster than institutions (employers, schools, governments) can adapt. The net number of jobs may be fine; the specific humans in specific jobs, facing specific transitions, are not.
This course is an honest assessment of work in the AI age. It covers what AI is actually automating, what it isn't, the historical pattern of automation booms, the economics of labor displacement, the specific occupations most affected, how organizations are (and aren't) handling the transition, and the questions parents, students, workers, and leaders should be asking about their own futures.
If you finish every module, here's who you become:
In the autumn of 2023, researchers at Harvard Business School embedded a controlled trial inside BCG, one of the world's premier strategy firms. They gave 758 consultants access to GPT-4 for specific tasks and withheld it from a control group. The results landed hard enough to reshape the internal debate about AI adoption at nearly every firm that read them.
The Fabrizio Dell'Acqua–led study, published in September 2023, found that consultants using GPT-4 completed 12.2% more tasks, did so 25.1% faster, and produced output that was rated 40% higher quality by blind evaluators — compared with the control group on the same assignments. These were not simple tasks; they included market sizing, creative ideation, and persuasive writing components drawn from real BCG client work.
Critically, the gains were largest at the lower end of the performance distribution. Consultants who had previously scored in the bottom third of the firm's internal benchmarks improved their output quality by over 43 percentage points on evaluated tasks. The study authors labelled this a "levelling up" effect — not a replacement of the weakest performers, but a compression of the performance gap between junior and senior staff.
A parallel 2023 study by Noy and Zhang at MIT assigned knowledge workers — this time, college-educated professionals writing reports, press releases, and analytical documents — either access to ChatGPT or no AI tool. Workers with AI completed assignments in 17 minutes on average versus 27 minutes in the control group. Quality ratings improved by 18 percentage points on a standardised rubric.
The finding echoed BCG's: the largest absolute gains accrued to workers who began with the weakest baseline writing skills. Workers already rated "excellent" showed moderate gains. Workers rated "poor" or "fair" at baseline showed dramatic convergence with the excellent group. The mechanism appeared to be task scaffolding — AI supplies the structural backbone of a document, freeing the worker to focus on judgment calls rather than formatting and phrasing.
Both studies measured specific, clearly defined cognitive tasks over short time horizons. Neither study measured what happens to overall employment, wage levels, or firm headcount over years — only immediate task-level productivity. Extrapolating from task gains to macroeconomic outcomes requires significant additional evidence.
In 2023, Erik Brynjolfsson and colleagues at Stanford studied a Fortune 500 company that deployed an AI assistant to 5,179 customer support agents. The tool provided real-time suggested responses drawn from a database of successful past conversations. After deployment, agents resolved 14% more issues per hour. Critically, new agents (fewer than two months on the job) improved by 35% — and attrition rates dropped, since the AI gave new hires the institutional knowledge that previously took months of trial and error to acquire.
The Stanford study introduced an economic wrinkle not present in the BCG or MIT work: the AI system was effectively transferring the tacit knowledge of the firm's best workers to its newest hires at near-zero marginal cost. This has implications for wages, training budgets, and the premium attached to experience — themes that recur throughout this module.
Before 2023, most claims about AI's productivity impact were either theoretical projections from economists or anecdotal reports from technology vendors with obvious conflicts of interest. These three studies — all peer-reviewed or working-paper quality, all using randomised or quasi-experimental designs — provided the first credible causal evidence from real workplaces.
Their limits are equally important. All three studied white-collar knowledge work. All three measured short-term task completion, not career trajectories. None measured whether firms retained the productivity gains or passed them to shareholders versus workers. And none controlled for the learning effects that might reverse over time as novelty wears off or as workers adapt their work style to accommodate AI. These are the open questions that the rest of this module pursues.
On narrow, well-defined cognitive tasks performed by knowledge workers, AI assistance in 2023–2024 produced real, statistically significant productivity gains of 14–40% depending on task type and worker baseline. These gains were not evenly distributed — lower-performing workers gained most. What these studies cannot tell us is whether those gains translate into job security, wage growth, or net employment increases. That question requires different evidence.
You've read about three real studies measuring AI's impact on worker productivity (BCG, MIT, Stanford). Now interrogate that evidence more deeply. Ask the assistant about methodology, potential confounds, what the studies can and cannot prove, or how to apply these findings to a real workplace situation.
In March 2023, OpenAI researchers Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock published a paper titled "GPTs are GPTs" — a deliberate reference to economist Erik Brynjolfsson's concept of general purpose technologies. Their core claim: large language models expose workers to augmentation or automation of significant portions of their work, and this exposure is highest, paradoxically, among college-educated professionals rather than lower-wage manual workers.
Eloundou et al. used the U.S. Department of Labor's O*NET database — which catalogues thousands of occupations by their component tasks — and had human raters plus GPT-4 itself classify each task according to whether an LLM could reduce the time to complete it by at least 50%. An occupation was classified as "exposed" if more than 50% of its tasks met this criterion.
Results: approximately 80% of the U.S. workforce has at least 10% of their tasks exposed to LLMs. About 19% of workers are in occupations where more than 50% of tasks are exposed. Critically, the most exposed occupations include mathematicians, tax preparers, financial quantitative analysts, writers and authors, and web designers — not truck drivers, janitors, or agricultural workers, whose tasks remain largely physical and therefore outside current LLM capabilities.
The OpenAI paper is careful to note that "exposure" means AI could assist or automate a task — not that it will, or that the job will disappear. Decisions about whether to deploy AI, how to reorganise work, and whether gains accrue to workers or employers are made by humans and institutions, not by technical capability alone.
In a widely-read March 2023 report, Goldman Sachs economists Jan Hatzius and colleagues estimated that generative AI could automate tasks equivalent to 300 million full-time jobs globally — but immediately clarified that this figure described potential task automation, not net job loss. Their model distinguished between:
Substitution effect: Tasks currently done by humans that AI could perform instead, potentially reducing demand for labour in specific roles.
Complementarity effect: Productivity gains that expand output and demand for the goods and services workers produce, increasing total employment.
New job creation effect: Entirely new roles created by the technology itself and by firms managing and implementing AI systems.
Goldman's central scenario predicted that AI-driven productivity gains equivalent to a 7% boost to global GDP would occur over ten years, but that the net employment effect remained "highly uncertain" and dependent on policy, wage bargaining, and the pace of adoption.
Economic historian David Autor has spent two decades documenting what he calls "task-biased technological change." His central insight, refined from 2003 onward with collaborators Levy and Murnane: technology tends to automate routine cognitive tasks (data entry, basic analysis, rule-following) while complementing non-routine cognitive tasks (judgment, creativity, persuasion, novel problem-solving) and remaining largely unable to replicate physical dexterity in unstructured environments.
Applied to AI: a radiologist's job involves tasks ranging from reviewing scan images (increasingly automated) to communicating diagnoses to anxious patients, managing clinical relationships, and making judgment calls in ambiguous cases (not easily automated). The occupation persists even as specific tasks within it are transformed. Autor's framework predicts job transformation as the primary near-term effect — not wholesale job destruction.
When automated teller machines were introduced at scale in the United States from the late 1970s onward, economists predicted significant job losses in bank teller employment. The opposite occurred. By 2004, there were more bank tellers employed in the U.S. than before ATMs, according to economist James Bessen's research published in 2015. Why? ATMs reduced the cost of operating a bank branch, so banks opened more branches. Each branch needed fewer tellers, but the number of branches more than compensated. Meanwhile, the teller role itself shifted from cash handling toward customer service and financial advising — tasks ATMs could not perform.
This does not mean AI will follow the same pattern. ATMs automated one narrow mechanical task. LLMs can operate across dozens of cognitive task categories simultaneously. But the ATM case demonstrates that technological displacement of specific tasks routinely co-exists with job count stability or growth when the technology also reduces costs and expands markets.
Apply the task-exposure framework from Lesson 2 to a real occupation or industry you care about. Name a job or sector, and work with the assistant to identify which specific tasks within that role have high LLM exposure versus low exposure — and what that implies for how the role might transform.
When researchers at the International Monetary Fund analysed wage data around AI adoption across multiple countries in 2023, they encountered a consistent pattern: productivity gains from AI disproportionately accrued to capital owners and firms in the short run, with workers seeing modest wage gains only in occupations where AI complemented rather than substituted for their skills. The distributional question — who benefits — proved as important as the productivity question — how much more gets done.
IMF chief economist Pierre-Olivier Gourinchas released an analysis in January 2024 — timed to coincide with the Davos World Economic Forum — projecting that AI would affect 40% of jobs globally and up to 60% of jobs in advanced economies. The IMF's framing was explicitly distributional: in advanced economies, AI exposure is roughly split between augmentation (higher productivity, potentially higher wages) and displacement (tasks replaced, potentially lower wages or unemployment). In emerging markets, the share of jobs with high displacement risk is lower, but so is the capacity to manage the transition through social safety nets.
The report identified a specific risk: AI tends to raise the value of high-skill workers who can direct and verify AI output, while potentially depressing wages for mid-skill workers whose tasks are automated. This is different from the "hollowing out" of the middle that characterised the 1990s–2010s automation wave. AI's exposure is concentrated among cognitive rather than routine blue-collar tasks, affecting a different set of workers.
GitHub Copilot, launched in June 2022, provides one of the cleaner natural experiments in AI's wage effects. In a 2023 controlled study by Peng et al., programmers using Copilot completed coding tasks 55% faster on average. However, wage data for software developers did not deteriorate in 2022–2023 despite this productivity increase — according to U.S. Bureau of Labor Statistics data, median annual wages for software developers remained at approximately $120,000 and vacancy rates stayed elevated.
The likely explanation: software development faces persistent labour shortages, meaning employers cannot easily substitute AI for workers without the context and judgment needed to manage AI output. Productivity gains expanded output rather than reducing headcount. This is consistent with the complementarity effect — but it is a sector-specific outcome, not a universal law. Labour market conditions (tight vs. slack, unionised vs. non-unionised, high-skill vs. low-skill) profoundly shape whether productivity gains translate to wages.
When a few dominant employers control hiring in a sector, workers have limited ability to capture productivity gains through job switching or wage bargaining. Research by economists Azar, Marinescu, and Steinbaum (2020) documented significant employer market power (monopsony) in U.S. labour markets. In sectors with high monopsony concentration, AI productivity gains are more likely to accrue to employers rather than workers — irrespective of how much task efficiency improves.
MIT economists Daron Acemoglu and Pascual Restrepo have produced the most rigorous econometric work on automation and wages. Their 2019 paper "Robots and Jobs" found that each additional industrial robot per thousand workers in a commuting zone was associated with a wage decline of 0.42% and employment decline of 0.2 percentage points. Their 2022 extension argued that most automation since 1980 has been "so-so" — generating efficiency gains that flow to capital without creating enough new tasks to restore worker incomes.
Acemoglu has been consistently more pessimistic than many technology economists about AI's wage effects. In a widely-circulated 2023 MIT working paper, he argued that the productivity gains from current AI are likely to remain concentrated in specific task types, and that without policy intervention (portability of benefits, wage supplements, antitrust enforcement against labour market concentration), the gap between AI-exposed workers and everyone else will widen.
A 2024 NBER working paper by Humlum examined Danish firms that adopted AI software tools between 2018 and 2022. The study found that adoption raised firm-level revenue and profits, but workers in adopting firms saw only a 1–3% wage increase — substantially below the productivity gains implied by the firm-level revenue data. The "pass-through" rate from AI productivity gains to worker wages appeared low, consistent with a world where firms capture most of the surplus and return it to shareholders.
The pass-through was higher in firms with works councils (formal worker representation bodies) and in sectors with sectoral collective agreements — suggesting that institutional bargaining power shapes how AI gains are distributed, not just technical exposure levels.
AI productivity gains are real and documented. Worker wage gains from AI adoption are also real — but substantially smaller than the productivity gains, and highly variable depending on labour market tightness, unionisation, institutional bargaining structures, and the specific nature of the tasks automated. The evidence does not support either "AI will make everyone richer" or "AI will impoverish workers." It supports a more uncomfortable finding: outcomes depend on the distribution of power between workers and employers, which is shaped by policy and institutions, not technology alone.
Lesson 3 established that AI productivity gains don't automatically translate to worker wages — pass-through depends on institutions, market power, and labour conditions. Use this lab to explore specific scenarios where the wage question plays out differently, or to challenge the evidence you've read.
Between 1999 and 2011, the United States lost approximately 5.5 million manufacturing jobs — a substantial portion attributable to import competition from China following WTO accession. Economists David Autor, David Dorn, and Gordon Hanson documented in their landmark 2013 paper that workers in the most exposed local labour markets did not transition to new sectors as standard trade theory predicted. They remained in depressed regions, left the labour force, or found lower-wage service jobs. The transition costs — in lost income, health outcomes, and community cohesion — were severe and persistent. This "China Shock" episode established the evidentiary foundation for the modern understanding of transition risk.
The World Economic Forum's "Future of Jobs Report 2023" surveyed approximately 800 companies covering 11.3 million workers across 27 industries. The report projected that by 2027, AI and automation would displace 83 million jobs globally while creating 69 million new jobs — a net loss of 14 million positions, or 2% of current employment. The new jobs projected include AI and machine learning specialists, data analysts, digital transformation specialists, cybersecurity experts, and roles in green energy transition.
The critical problem is not the aggregate number — it is the match quality. A displaced data entry clerk in a mid-sized city does not straightforwardly become an AI prompt engineer. The skills required, the geographic location of new opportunities, and the age and education profile of displaced workers all create frictions that aggregate job-count comparisons obscure.
Economic historians Daron Acemoglu and Pascual Restrepo documented in a 2018 paper that during the first Industrial Revolution, real wages for English workers stagnated for approximately 60–70 years after mechanisation began — despite dramatic increases in industrial output. This period, which economic historians call the "Engels Pause" (after Friedrich Engels' 1845 documentation of Manchester factory conditions), is the historical archetype of a transition that worked out eventually but imposed catastrophic costs on the living generation during the transition itself.
More recent evidence: the U.S. Bureau of Labor Statistics Trade Adjustment Assistance data shows that workers displaced from manufacturing jobs after trade shocks took an average of three to four years to reach stable re-employment — and many never returned to their prior wage level. A 2019 paper by Autor, Maestas, Mullen, and Strand found that workers displaced in mid-career (ages 45–54) experienced permanent earnings declines averaging 15–20% relative to comparable non-displaced workers, and significantly elevated disability application rates.
Public retraining programs for displaced workers have a mixed evidence record. A landmark 2021 meta-analysis of U.S. federal retraining programs by Heckman, LaLonde, and Smith found that while classroom training showed modest positive effects for adult women, effects for adult men were near zero or negative. The implication: creating new jobs is not the same as enabling displaced workers to access those jobs. Training frictions and credential barriers are large and poorly addressed by most existing policy tools.
LinkedIn's 2023 "Future of Work" report, drawing on job posting data from over 950 million users, identified the fastest-growing job categories associated with AI: AI prompt engineers, machine learning operations (MLOps) engineers, AI ethics officers, AI-augmented healthcare coordinators, and AI content editors. These roles are concentrated in three metropolitan areas — San Francisco/Silicon Valley, New York, and Seattle — and require programming, data science, or graduate education backgrounds for the technical roles.
The geographic and educational concentration matters enormously. A call-centre worker in a mid-sized Midwestern city whose job is automated by an LLM is not in a realistic position to relocate to San Francisco to become a prompt engineer. Geographic immobility — driven by homeownership, family ties, and community anchors — is one of the most robust findings in labour economics, and it is a central reason why aggregate job-count optimism does not translate into individual worker security.
ChatGPT reached 100 million users in 60 days after its November 2022 launch — the fastest product adoption in recorded history. For context, it took the telephone 75 years to reach 100 million users; the internet, 4 years; Instagram, 2.5 years. The pace of LLM diffusion into enterprises accelerated even faster: McKinsey's 2023 survey of 1,700 organisations found that 79% of respondents had had at least some personal exposure to generative AI, and 22% were using it regularly in their work. One year earlier, these figures were effectively zero.
This speed creates a specific challenge: labour markets, educational systems, and policy institutions that adapted to previous technological transitions over decades are now being asked to adapt over months. The question is not whether transition eventually occurs — it will — but whether the institutional infrastructure for managing that transition exists at the required speed.
What the evidence shows, taken together: AI is producing real productivity gains on specific cognitive tasks; those gains are currently concentrated in knowledge work; jobs will transform rather than uniformly disappear; wage pass-through from productivity to workers is low and institutionally contingent; and the speed of transition is unprecedented, which is the central challenge. None of this is reason for either complacency or panic — but it is a precise set of claims, grounded in documented data, that should anchor every subsequent discussion of what to do.
Lesson 4 identified the speed problem, the retraining gap, and geographic immobility as the central friction points in the AI-labour transition. In this lab, think through what responses — from individuals, firms, or governments — the evidence actually supports. Be specific; vague "more training" answers don't engage seriously with the retraining effectiveness data.