Lesson 1

AI and Jobs

Labor economics, automation history, and the political economy of technological displacement.

The Luddites of 1811-1816 were skilled textile workers who destroyed machinery that was displacing their livelihoods. History has remembered them as anti-technology; in fact, they were pro-skilled labor and anti-exploitation. Their uprising was suppressed by military force, and the displacement continued. Historian David Landes later calculated that the Industrial Revolution's productivity gains took roughly 60 years to produce broadly shared prosperity — and that the intervening decades were characterized by falling real wages for many workers despite rising aggregate output. The question for AI is whether technology policy can shorten that gap.

The Historical Pattern

Major technological transitions have followed a recurring pattern:

Productivity gains accrue first to capital owners
Labor displacement in affected sectors is immediate; new job creation is gradual
Aggregate employment eventually recovers, but the distribution changes — some groups benefit significantly, others are left behind permanently
Policy choices (labor law, education, social insurance) determine how the transition costs are distributed

What's Different About AI

AI displacement differs from previous automation in potentially important ways:

Breadth: Previous automation displaced specific physical tasks; AI affects cognitive work across many sectors simultaneously
Speed: Digital deployment is faster than physical capital investment — the displacement can happen before policy adapts
Skill reversal: Previous automation displaced low-skill work and created demand for high-skill work; AI may displace high-skill cognitive work while leaving physical low-skill work less affected

The Distribution Question

The economic gains from AI will accrue to someone. The distribution of those gains — between capital and labor, between high-skill and low-skill workers, between wealthy and developing countries — is determined by policy, not by technology.

Back Take Quiz

Quiz 1

AI and Jobs

5 questions — free, untracked, retake anytime.

does the Luddite example illustrate about technological displacement?

✓ Correct — ✅ ✓ The Luddite case: productivity gains came quickly; broadly shared prosperity took 60 years. The gap between aggregate gains and distributed benefits is real, prolonged, and policy-determined.

❌ ❌ The Luddite case illustrates that aggregate productivity gains don't automatically translate to broadly shared prosperity. The Industrial Revolution took ~60 years to produce shared gains — with severe displacement in between.

is the 'skill reversal' argument about AI displacement?

✓ Correct — ✅ ✓ Skill reversal: historical automation displaced low-skill physical work and increased demand for cognitive skills. AI may displace cognitive work — reversing the long-term trend where more education meant more protection.

❌ ❌ Skill reversal: historical automation displaced low-skill physical work, increasing demand for high-skill cognitive work. AI may displace cognitive work — potentially reversing the historical protective value of high education.

makes AI displacement potentially faster than previous technological transitions?

✓ Correct — ✅ ✓ Digital deployment: AI can be deployed at scale through software updates without the physical capital investment timelines of previous technological transitions. Displacement may outpace policy.

❌ ❌ AI deploys through software — without the physical capital investment timelines of previous transitions. Displacement can happen faster than policy frameworks adapt.

determines the distribution of AI's economic gains between capital and labor?

✓ Correct — ✅ ✓ Distribution of AI gains between capital and labor is determined by policy choices: labor law, taxation, education, and social insurance. Technology creates the gains; policy distributes them.

❌ ❌ The distribution of AI's economic gains is determined by policy choices — labor law, education, taxation, social insurance. Technology creates gains; policy determines who gets them.

might AI displacement be qualitatively different from the Industrial Revolution displacement in ways that make historical analogies unreliable?

✓ Correct — ✅ ✓ Breadth of simultaneous cognitive work displacement across many sectors is qualitatively different from the sector-specific physical displacement pattern of previous technological transitions.

❌ ❌ AI's breadth matters: it displaces cognitive work across many sectors simultaneously — unlike the sector-specific physical displacement of the Industrial Revolution. Historical analogies may not hold.

Back to Lesson Start Lab 1

Lab 1

Labor Economics and Policy

Analyze AI labor displacement and develop a distribution framework.

Lab 1 — Labor Economics and Policy

Analyze AI labor displacement from a political economy perspective.

The AI opens: historical transitions took decades to produce broadly shared prosperity. What policy framework would shorten that gap for AI-driven displacement — and what are the political economy obstacles to implementing it?
Address the skill reversal specifically: if AI displaces high-skill cognitive work, what does that mean for the education-as-protection model?
Develop your position on distribution: how should the economic gains from AI be distributed between capital and labor?

Consider: automation taxes, profit-sharing, shortened work weeks, education reform, and the political obstacles each faces.

🎯 AI GuideLab 1

Back to Quiz Next: Lesson 2

Lesson 2

AI and Democracy

Information ecosystems, epistemic foundations of democracy, and AI's structural threat to political self-governance.

Philosopher Jürgen Habermas argued that democracy requires a functioning public sphere: a shared information environment where citizens can form views through rational deliberation. Social media had already fragmented that sphere before generative AI. AI-powered synthetic media, personalized persuasion at scale, and the collapse of epistemic trust don't just make democracy harder — they may make the preconditions of democracy structurally unavailable. When citizens can't agree on basic facts, and when any video or audio can be fabricated, the deliberative foundation of self-governance erodes.

The Epistemic Foundation

Democratic self-governance assumes:

Citizens can access reasonably accurate information about the world
Citizens can distinguish credible from non-credible sources
There is a shared factual baseline — agreement on what is real — that political disagreement can be about values rather than facts
Manipulation at scale is costly enough to be limited

AI undermines each of these assumptions.

AI's Structural Democratic Threat

Personalized epistemic bubbles: Recommendation systems curate information environments individually — fragmenting any shared factual baseline
Synthetic evidence: AI-generated documents, audio, video, and images make "seeing is believing" unreliable
Attention capture: AI-optimized platforms maximize engagement through emotional activation — prioritizing outrage over deliberation
Manipulation at zero marginal cost: AI removes the human labor constraint on influence operations

The Habermas Concern

The public sphere isn't just an asset — it's a prerequisite. A democracy that loses the shared epistemic foundation required for deliberation may face challenges that no institutional reform can fix if the information environment is too degraded.

Back Take Quiz

Quiz 2

AI and Democracy

5 questions — free, untracked, retake anytime.

does Habermas's 'public sphere' concept mean for democracy?

✓ Correct — ✅ ✓ The public sphere is a prerequisite for democracy: a shared information environment enabling rational deliberation. Without it, democratic self-governance loses its epistemic foundation.

❌ ❌ Habermas's public sphere: a shared information environment where citizens deliberate rationally. It's a prerequisite for democracy — not just a useful feature. AI degrades this prerequisite.

is the 'shared factual baseline' assumption that AI threatens?

✓ Correct — ✅ ✓ The shared factual baseline: citizens agree enough on basic facts that debate is about values, not about whether events happened. AI-generated synthetic evidence and personalized information bubbles erode this.

❌ ❌ Shared factual baseline: enough agreement on basic facts that political disagreement is about values rather than reality itself. AI erodes this — turning factual questions into contested territory.

does 'personalized epistemic bubbles' represent a structural democratic threat?

✓ Correct — ✅ ✓ Personalized epistemic bubbles fragment the shared factual baseline. Citizens in different bubbles may disagree not just on values but on basic facts — making democratic deliberation structurally difficult.

❌ ❌ Personalized bubbles fragment the shared factual baseline. Citizens may inhabit different versions of reality — making deliberation structurally difficult rather than just contentious.

distinguishes AI's democratic threat from pre-existing concerns about media bias?

✓ Correct — ✅ ✓ AI is qualitatively different from media bias: near-zero marginal cost manipulation and synthetic evidence that makes any content potentially fabricated — not just slanted coverage of real events.

❌ ❌ AI's threat is qualitatively different from media bias: near-zero cost influence operations and synthetic evidence that makes any content potentially fabricated — not just biased coverage of real events.

might institutional reforms be insufficient if the information environment is too degraded?

✓ Correct — ✅ ✓ Democratic institutions are built on epistemic prerequisites. If citizens can't access reasonably accurate information or form views through deliberation, institutional reforms fix the wrong layer.

❌ ❌ Democratic institutions assume epistemic prerequisites — reasonably accurate information, shared facts, deliberative capacity. If those prerequisites are destroyed, institutional reforms operate on top of a broken foundation.

Back to Lesson Start Lab 2

Lab 2

Democracy and Epistemic Infrastructure

Analyze AI's structural democratic threat and governance responses.

Lab 2 — Democracy and Epistemic Infrastructure

Analyze AI's structural threat to democracy and develop a governance response.

The AI opens: Habermas argued the public sphere is a prerequisite for democracy. If AI degrades that prerequisite structurally, is the problem solvable through governance — or is there a level of information environment degradation that institutional democracy can't survive?
Develop your analysis of whether this is a solvable governance problem or a structural constitutional threat.
Address: what obligations do AI platform companies have for maintaining the epistemic foundations of democracy in the countries they operate in?

Consider: platform obligations, information environment regulation, provenance standards, media literacy, and international dimensions.

🎯 AI GuideLab 2

Back to Quiz Next: Lesson 3

Lesson 3

AI and Healthcare

The transformation of medicine — evidence, equity, and the political economy of health AI.

The FDA approved the first AI-based medical device in 1995 (a neural network for detecting heartbeat abnormalities). By 2023, over 500 AI/ML-based medical devices had been approved. The regulatory framework struggled to keep pace: traditional FDA approval is designed for static devices, but AI medical systems update continuously. In 2021, researchers published a study showing that an AI for diabetic retinopathy screening, FDA-approved with 90%+ accuracy on validation data, performed at 40% sensitivity when deployed on a different patient population — a real-world failure that validation hadn't caught.

The Validation Gap

The gap between validation performance and real-world performance is the central challenge of healthcare AI deployment:

Distribution shift: Validation populations may not match deployment populations
Dataset bias: AI trained on data from academic medical centers may fail on community hospital populations
Temporal shift: Disease presentations change; AI trained on historical data may become less accurate over time
Feedback loops: AI recommendations influence clinical practice, which influences the data the AI is evaluated on

The Political Economy of Health AI

Healthcare AI investment flows toward high-margin specialties in wealthy markets:

More AI tools for radiology, dermatology, and ophthalmology than for primary care or psychiatry
More AI tools for wealthy-country diseases than neglected tropical diseases
More AI tools for large-volume procedures than for rare diseases

The Access Question

The healthcare AI tools most needed in low-income settings (primary care, diagnostic support for under-resourced facilities) attract the least investment because the commercial returns are lowest. This is a market failure, not a technical barrier.

Back Take Quiz

Quiz 3

AI and Healthcare

5 questions — free, untracked, retake anytime.

is the 'validation gap' in healthcare AI?

✓ Correct — ✅ ✓ Validation gap: AI performs well on validation data but fails in real-world deployment when the patient population differs from the validation population — a distribution shift problem.

❌ ❌ Validation gap: the difference between performance on validation data and real-world performance — often because the deployment population differs from the validation population.

is 'temporal shift' as a healthcare AI risk?

✓ Correct — ✅ ✓ Temporal shift: disease presentations and clinical practice evolve. AI trained on historical data may become progressively less accurate as the real world diverges from its training distribution.

❌ ❌ Temporal shift: disease presentations change over time. AI trained on historical data may become less accurate as the real world diverges from what it was trained on.

do healthcare AI investments concentrate in radiology, dermatology, and ophthalmology?

✓ Correct — ✅ ✓ Commercial investment flows toward high-margin, high-volume specialties in wealthy markets — not toward the greatest global health need. This is a market failure.

❌ ❌ Commercial investment concentrates in high-margin, high-volume wealthy-market specialties. This reflects commercial return on investment, not global health need.

regulatory challenge does continuously updating AI create for the FDA?

✓ Correct — ✅ ✓ FDA approval is designed for static devices. Continuously updating AI systems can change substantially post-approval — requiring ongoing evaluation frameworks the FDA has had to develop on the fly.

❌ ❌ Traditional FDA approval is for static devices. Continuously updating AI may change substantially post-approval — requiring regulatory frameworks that didn't exist when healthcare AI emerged.

diabetic retinopathy AI example illustrates:

✓ Correct — ✅ ✓ The retinopathy case: 90%+ validation accuracy → 40% real-world sensitivity on a different population. Validation performance didn't predict real-world performance — a validation gap with clinical consequences.

❌ ❌ The retinopathy example: 90%+ validation accuracy → 40% real-world sensitivity on a different population. Validation didn't catch the distribution shift — with clinical consequences.

Back to Lesson Start Lab 3

Lab 3

Healthcare AI Governance

Design a regulatory framework for healthcare AI deployment.

Lab 3 — Healthcare AI Governance

Analyze healthcare AI regulatory challenges and develop a governance framework.

The AI opens: traditional FDA approval doesn't handle continuously updating AI, and validation performance doesn't reliably predict real-world performance. Design a regulatory framework for healthcare AI that addresses both problems.
Address the market failure: commercial investment flows toward high-margin specialties, not greatest global health need. What policy would redirect investment toward higher-need applications?
Develop requirements for real-world performance monitoring of deployed healthcare AI.

Consider: pre-deployment requirements, post-deployment monitoring, demographic performance requirements, and access mandates.

🎯 AI GuideLab 3

Back to Quiz Next: Lesson 4

Lesson 4

AI and Education

Pedagogical theory, learning science, and the structural transformation of educational institutions.

Benjamin Bloom's 1984 "2 Sigma Problem" identified that students who received one-to-one tutoring performed two standard deviations better than those in conventional classrooms. Bloom called this a "problem" because one-to-one tutoring was economically unscalable — most students couldn't afford it. AI tutoring systems represent the first plausible technological answer to Bloom's 2 Sigma Problem: personalized, adaptive, available instruction at scale. But the research on what AI tutoring actually delivers — as opposed to what it promises — remains early and mixed.

What Learning Science Says About AI Tutoring

Spacing and interleaving: Optimal learning involves distributed practice across time and interleaved problem types. AI can enforce this more reliably than human schedules.
Immediate feedback: Learning is faster with immediate feedback on errors. AI provides this at scale without teacher bottlenecks.
Worked examples vs. generation: Research is mixed on whether AI should show worked examples or prompt students to generate answers — the right balance depends on skill level and task type.
Motivation and relationship: Human relationships are a significant driver of student motivation. AI tutors lack this — a limitation that may matter most for students who need it most.

Credential Inflation and Institutional Risk

If AI can produce most of the work that generates academic credentials, institutions face a fundamental legitimacy challenge: their credentials may no longer signal what employers and society assume they signal. This is not just an academic integrity problem — it's an institutional sustainability problem.

The Reformulation Question

What human capabilities does education exist to build? Once that question is answered, assessment can be redesigned around AI-resistant evidence of those capabilities. Without answering that question first, institutions will keep adapting tactically without solving the structural problem.

Back Take Quiz

Quiz 4

AI and Education

5 questions — free, untracked, retake anytime.

was Bloom's '2 Sigma Problem'?

✓ Correct — ✅ ✓ Bloom's 2 Sigma: one-to-one tutoring = +2 standard deviations in performance vs. conventional classrooms. The 'problem' was that tutoring was economically unscalable. AI tutoring is the first plausible solution.

❌ ❌ Bloom's 2 Sigma: students with one-to-one tutoring performed 2 standard deviations better than conventional classroom students — but tutoring was economically unscalable. AI represents the first plausible answer.

does learning science say about 'spacing and interleaving' that AI tutoring can optimize?

✓ Correct — ✅ ✓ Spacing and interleaving: distributed practice across time and mixed problem types produce better retention. AI can enforce these schedules more consistently than human-managed classroom environments.

❌ ❌ Spacing and interleaving: distributed practice across time and mixed problem types produce better learning. AI can enforce these schedules reliably — something classroom logistics often can't.

might AI tutoring fail for students who need it most?

✓ Correct — ✅ ✓ Human relationship is a motivation driver that AI tutors can't replicate. This limitation may matter most for students with low intrinsic motivation or limited support structures — those who most need the human element.

❌ ❌ Human relationship drives student motivation in ways AI can't replicate. This limitation may matter most for students who most need support — those with low intrinsic motivation or poor home environments.

is the 'institutional sustainability' problem posed by AI and academic credentials?

✓ Correct — ✅ ✓ Institutional sustainability: if credentials no longer signal actual capability (because AI generates the work), the value that universities provide — credentialing — faces a legitimacy crisis.

❌ ❌ Institutional sustainability: if AI generates the work that produces credentials, credentials no longer signal what employers and society assume they signal. This is an institutional legitimacy crisis.

must the 'reformulation question' precede assessment redesign?

✓ Correct — ✅ ✓ Without answering 'what capabilities does education exist to build?', institutions adapt tactically — making assignments AI-harder — without addressing the structural legitimacy question.

❌ ❌ Without the reformulation question — what capabilities does education exist to build? — assessment redesign is tactical adaptation without structural solution.

Back to Lesson Start Lab 4

Lab 4

Education Transformation Analysis

Analyze AI's structural challenge to educational institutions.

Lab 4 — Education Transformation Analysis

Analyze AI's structural challenge to educational institutions.

The AI opens with Bloom's 2 Sigma Problem: AI tutoring may represent the first scalable answer. But the 2 Sigma gain came from human tutors, not AI. What evidence should we require before deploying AI tutoring at scale?
Address the institutional sustainability problem: if credentials lose legitimacy, what replaces them as signals of human capability?
Develop your answer to the reformulation question: what human capabilities should education build in an AI world?

Consider: learning science evidence requirements, alternative credentialing models, and the equity implications of transitioning from traditional credentials.

🎯 AI GuideLab 4

Back to Quiz Next: Lesson 5

Lesson 5

AI and the Environment

Energy systems, climate modeling, and the geopolitics of AI's environmental impact.

In 2023, Microsoft reported a 34% increase in water consumption from 2021-2022, largely attributable to AI operations — AI data centers require significant water cooling. In the same period, the company announced major AI-climate partnerships, using AI for energy grid optimization and carbon accounting. The same company was simultaneously a significant environmental burden and a significant environmental tool. This duality isn't hypocrisy — it reflects the genuine complexity of AI's environmental role. What matters is the net balance and whether policy structures that balance correctly.

The Full Environmental Footprint

Energy: Training and inference at scale; data center power consumption projected to grow significantly with AI adoption
Water: Data center cooling consumes significant water, increasingly in water-stressed regions
Hardware: GPU production requires rare earth materials; hardware cycles are short, creating e-waste and mining pressure
Land: Data center construction and associated infrastructure

AI as Climate Tool — The Evidence

Grid optimization: DeepMind's AI reduced Google data center cooling energy by 40% — the same technology applied to national grids could reduce significant waste
Climate modeling: AI-accelerated models can run at higher resolution with less compute — improving prediction and planning
Materials discovery: AI accelerating the discovery of better battery materials, solar cell components, and carbon capture materials
Agriculture: AI precision agriculture reducing fertilizer and water use, cutting agricultural emissions

The Policy Design Problem

The environmental costs of AI (energy, water, hardware) fall on current populations. The climate benefits of AI tools (better grid management, materials discovery) accrue over decades. Pricing the present cost against the future benefit requires policy — the market won't do it automatically.

Back Take Quiz

Quiz 5

AI and the Environment

5 questions — free, untracked, retake anytime.

does the Microsoft water consumption case illustrate 'genuine complexity' rather than hypocrisy?

✓ Correct — ✅ ✓ Genuine complexity: the same AI infrastructure that consumes water and energy also enables climate solutions. It's not hypocrisy — it's a real net-balance question that policy must address.

❌ ❌ The complexity is genuine: the same company's AI operations that consume water and energy also enable climate tools. The question is the net balance and whether policy structures it correctly.

is the 'policy design problem' in AI and climate?

✓ Correct — ✅ ✓ The temporal mismatch: current environmental costs vs. long-horizon climate benefits. Markets underprice future benefits relative to current costs — policy must correct this.

❌ ❌ Policy design problem: AI's environmental costs are immediate; climate benefits accrue over decades. Markets discount future benefits, making the net balance look negative without policy correction.

did DeepMind's AI cooling optimization demonstrate?

✓ Correct — ✅ ✓ DeepMind's 40% cooling energy reduction demonstrated that AI-optimized systems management could produce significant efficiency gains — with implications for national grid management.

❌ ❌ DeepMind's 40% cooling reduction: proof that AI-optimized systems management produces significant efficiency gains. The same approach applied to national grids could reduce waste substantially.

is hardware production an underappreciated part of AI's environmental footprint?

✓ Correct — ✅ ✓ Hardware production involves rare earth mining, short product cycles (e-waste), and manufacturing footprint — environmental costs not captured in energy or water consumption metrics.

❌ ❌ Hardware production: rare earth mining, short GPU cycles creating e-waste, manufacturing footprint — environmental impacts not captured in energy/water metrics but significant.

does AI materials discovery offer for climate?

✓ Correct — ✅ ✓ AI materials discovery: accelerating identification of better battery materials, solar cell components, and carbon capture materials — potentially compressing the timeline for clean energy adoption.

❌ ❌ AI materials discovery: accelerating identification of better battery materials, solar cell components, and carbon capture materials — potentially compressing clean energy transition timelines.

Back to Lesson Start Lab 5

Lab 5

AI Climate Policy Design

Design policy for AI's environmental tradeoffs.

Lab 5 — AI Climate Policy Design

Analyze the environmental tradeoffs and design a policy framework.

The AI opens: AI's environmental costs are immediate (energy, water, hardware); its climate benefits accrue over decades. How do you design policy that captures this temporal mismatch?
Develop a framework for mandatory environmental disclosure, carbon pricing, and investment prioritization for AI climate applications.
Address the geopolitical dimension: if one country imposes strict AI environmental standards and others don't, does that just move the environmental burden while undermining competitiveness?

Consider: temporal discounting, carbon pricing, mandatory disclosure, renewable requirements, and international coordination challenges.

🎯 AI GuideLab 5

Back to Quiz Next: Lesson 6

Lesson 6

AI and Inequality

Structural inequality, capability distribution, and the political economy of AI's distributional effects.

Economists Daron Acemoglu and Simon Johnson, in their 2023 book "Power and Progress," argued that technological progress does not automatically benefit everyone — history shows it benefits whoever controls the technology and shapes its deployment. The gains from agricultural, industrial, and information technology revolutions were not automatically broadly shared; they depended on specific institutional arrangements, labor movements, and policy choices. Their argument for AI: the distributional outcome will be determined not by the technology but by who captures the gains and who sets the policy.

The Political Economy of Distribution

Three mechanisms determine who benefits from AI:

Ownership: Those who own AI systems (compute, models, data) capture a larger share of AI-generated productivity gains
Labor bargaining power: AI that increases employer leverage over workers (through surveillance, productivity measurement, task automation) reduces labor's share of gains
Policy: Taxation, antitrust, access requirements, and labor law determine how AI gains are redistributed

Global Inequality Dimension

AI development is concentrated in a few countries (US, China, EU). AI's economic benefits concentrate similarly. For most of the world's countries:

No significant AI development capacity
Consumer of AI tools built elsewhere, often with data from their populations
Potential benefit from AI healthcare/education tools — if access is structured correctly
Exposure to AI economic disruption (automation of outsourced service work) without the gains of AI productivity

Acemoglu's Warning

If AI is left to market forces without deliberate policy intervention, the gains will concentrate with AI developers and owners — replicating and amplifying the pattern of previous technological transitions where the powerful captured the gains and the costs were distributed more broadly.

Back Take Quiz

Quiz 6

AI and Inequality

5 questions — free, untracked, retake anytime.

is the core argument of Acemoglu and Johnson's 'Power and Progress'?

✓ Correct — ✅ ✓ Acemoglu and Johnson: technology progress doesn't automatically benefit everyone. Historical evidence shows the distribution depends on ownership, institutions, and policy — not on the technology itself.

❌ ❌ Acemoglu and Johnson: technological progress doesn't automatically produce broad benefit. The distribution depends on who controls the technology, labor bargaining power, and policy — not on the technology itself.

does AI potentially reduce labor's share of productivity gains?

✓ Correct — ✅ ✓ AI can increase employer leverage — through surveillance, productivity measurement, and task automation — reducing workers' ability to bargain for higher wages as productivity rises.

❌ ❌ AI can reduce labor bargaining power through surveillance, productivity measurement, and task automation — increasing employer leverage and reducing labor's share of productivity gains.

is the 'data from their populations' concern for developing countries?

✓ Correct — ✅ ✓ Data extraction: developing country populations contribute data to AI systems built and owned elsewhere — receiving the costs (data use) without capturing the development gains (ownership, productivity).

❌ ❌ Data contribution without ownership: developing country populations contribute data to AI systems built and owned elsewhere. They provide the input; others capture the gains.

three mechanisms does Acemoglu identify as determining AI's distributional outcomes?

✓ Correct — ✅ ✓ Acemoglu's three mechanisms: ownership (who holds the assets), labor bargaining power (who can claim gains), and policy (who sets the redistribution rules).

❌ ❌ Acemoglu's distribution mechanisms: ownership of AI systems, labor bargaining power, and policy choices (taxation, antitrust, access requirements, labor law).

does the concentration of AI development in the US, China, and EU mean for most of the world?

✓ Correct — ✅ ✓ AI concentration: most countries have no significant development capacity. They're exposed to AI disruption (automation of outsourced service work) without the gains of AI ownership and development.

❌ ❌ AI concentration: most countries have no AI development capacity. They're exposed to AI disruption (service work automation) without the gains — a global inequality amplifier.

Back to Lesson Start Lab 6

Lab 6

Political Economy of AI Distribution

Analyze the political economy of AI's distributional effects.

Lab 6 — Political Economy of AI Distribution

Analyze the distributional mechanisms and develop a redistribution framework.

The AI opens with Acemoglu's argument: the distribution of AI gains depends on ownership, labor bargaining power, and policy. Which of these three levers do you think is most important to get right — and why?
Develop a framework for international AI equity — how should developing countries that contribute data and experience disruption without development gains be addressed?
Address: what institutional arrangements would prevent AI gains from concentrating with AI developers and owners?

Consider: antitrust, data dividends for developing countries, international AI benefit funds, labor law reform, and progressive AI taxation.

🎯 AI GuideLab 6

Back to Quiz Next: Lesson 7

Lesson 7

AI and Power

Concentration of AI capability, geopolitical competition, and the governance of transformative technology.

As of 2024, the five largest AI companies (Microsoft/OpenAI, Google/DeepMind, Meta, Amazon, Anthropic) collectively control the majority of frontier AI development capacity. The compute required to train frontier models has grown exponentially; only organizations with billions in capital can participate. AI has become a geopolitical competition, with the US and China both making it a national strategic priority. This concentration of a potentially transformative technology in a handful of organizations and governments raises questions about accountability that the world's governance frameworks are not yet equipped to answer.

The Concentration Problem

Compute concentration: Frontier AI training requires compute accessible only to well-capitalized organizations — a structural barrier to entry
Data concentration: Large internet platforms hold data advantages that compound over time
Talent concentration: A small number of researchers with critical AI expertise concentrate in a few organizations
Network effects: AI systems that gain adoption improve through more data and feedback — widening advantages over time

Geopolitical Dimensions

AI has become a geopolitical competition along several dimensions:

Military AI: Autonomous weapons, intelligence analysis, cyber operations
Economic AI: Industrial productivity, financial systems, logistics
Surveillance AI: Mass surveillance capabilities now accessible to most governments
Influence AI: State-sponsored information operations using AI-generated content

The Governance Gap

The organizations developing the most capable AI are subject to the laws of the countries they're in — which don't have adequate governance frameworks for technology with global effects. International governance of AI is nascent. The gap between AI capability and governance capacity is widening.

Back Take Quiz

Quiz 7

AI and Power

5 questions — free, untracked, retake anytime.

structural factor creates the concentration of frontier AI development in a few organizations?

✓ Correct — ✅ ✓ Compute requirements for frontier training have grown exponentially — creating a structural capital barrier that concentrates frontier AI development in the most well-capitalized organizations.

❌ ❌ Exponentially growing compute requirements create a capital barrier: only organizations with billions can train frontier models. This is a structural concentration mechanism.

do network effects widen AI competitive advantages over time?

✓ Correct — ✅ ✓ Network effects: more adoption → more data and feedback → better AI → more adoption. This compounding dynamic means early advantages widen over time.

❌ ❌ Network effects: more adoption → more data and feedback → better AI → more adoption. AI advantages compound over time, making early leads structural rather than temporary.

is the 'governance gap' in AI?

✓ Correct — ✅ ✓ Governance gap: AI organizations are governed by national laws designed for different contexts, with no adequate international framework for technology that has global effects.

❌ ❌ Governance gap: AI is governed by national laws not designed for global-effect technology. International AI governance frameworks are nascent. The gap between AI capability and governance capacity is widening.

distinguishes AI surveillance capability from previous surveillance technology?

✓ Correct — ✅ ✓ AI democratizes surveillance capability: what previously required sophisticated intelligence agencies can now be implemented by most governments, dramatically expanding the population of states with mass surveillance capacity.

❌ ❌ AI democratizes mass surveillance: capabilities previously accessible only to sophisticated state actors are now accessible to most governments — dramatically lowering the cost of population monitoring.

makes AI geopolitical competition different from previous technology competitions?

✓ Correct — ✅ ✓ AI's geopolitical significance spans all major dimensions of national power simultaneously: economic productivity, military capability, surveillance, and information operations — unlike previous technologies that affected fewer dimensions.

❌ ❌ AI affects economic productivity, military capability, surveillance capacity, and information operations simultaneously — giving a single technology strategic implications across all major dimensions of national power.

Back to Lesson Start Lab 7

Lab 7

AI Power and Governance

Analyze AI concentration and design international governance frameworks.

Lab 7 — AI Power and Governance

Analyze the concentration problem and design governance responses.

The AI opens: a handful of companies and governments control the most capable AI. This concentration is structurally driven by compute requirements and network effects — not just market outcomes. What governance framework would address this structural concentration?
Address the international governance gap: AI has global effects but is governed by national laws. What international governance architecture would be adequate?
Develop your position on military AI: should autonomous lethal AI weapons systems be governed differently from civilian AI?

Consider: compute governance, antitrust for AI, international treaties, civilian/military distinctions, and the verification challenges of AI arms control.

🎯 AI GuideLab 7

Back to Quiz Next: Lesson 8

Lesson 8

The Future We Choose

Agency, values, and the possibility of shaping AI's societal trajectory.

Historian Langdon Winner's 1980 essay "Do Artifacts Have Politics?" argued that technologies are not politically neutral — they embed the values of those who design them, and some technologies have structural implications that constrain political choices for generations. The highway systems designed for cars make car-free cities structurally difficult. Nuclear power plants require centralized control and create proliferation risks. Winner's question for AI: what political and social values are being embedded in AI systems now, and what future choices are being foreclosed by the systems we're building?

Artifacts Have Politics

AI systems embed political choices that their designers may not consciously recognize:

What is optimized: engagement, profit, equity, safety?
Whose values are encoded in training data and RLHF?
Who controls the systems: corporations, governments, open-source communities?
What futures do the systems make more or less likely?

The Future Is Not Predetermined

AI's societal trajectory is not predetermined. Historical examples of successfully governed transformative technologies include:

Nuclear power: international safeguards, non-proliferation treaty
Commercial aviation: international safety standards, air traffic coordination
Pharmaceuticals: evidence requirements, regulatory approval, post-market surveillance
Internet: DNS governance, content moderation frameworks (imperfect but functional)

All of these involved significant political contestation. None were automatic. All required sustained engagement from people who understood the technology and cared about its social effects.

Why You're Here

AI literacy isn't just about knowing how AI works. It's about being equipped to participate in the decisions about what AI should do, who it should serve, and what future it should help build. Those decisions are being made now — and they need informed participants.

Back Take Quiz

Quiz 8

The Future We Choose

5 questions — free, untracked, retake anytime.

is Langdon Winner's argument that 'artifacts have politics'?

✓ Correct — ✅ ✓ Winner's argument: technologies aren't politically neutral — they embed designers' values and create structural constraints on future choices. Some technology choices foreclose political options for generations.

❌ ❌ Winner: technologies embed the values of their designers and have structural political implications. Some technology choices constrain political options for generations — they have politics built in.

political choices are embedded in AI systems that their designers may not consciously recognize?

✓ Correct — ✅ ✓ Embedded choices: optimization objectives (engagement, profit, equity), whose values RLHF encodes, who controls the systems. These are political choices made as technical ones.

❌ ❌ Embedded political choices: what is optimized, whose values are encoded in training data and RLHF, and who controls the systems. These technical choices are political choices.

do the historical examples of nuclear power, aviation, and pharmaceuticals demonstrate about governing transformative technology?

✓ Correct — ✅ ✓ Historical examples: effective governance of transformative technology is possible. Nuclear safeguards, aviation safety standards, pharmaceutical approval — all required contestation and deliberate design.

❌ ❌ Nuclear safeguards, aviation standards, pharmaceutical approval: transformative technologies have been effectively governed. Not automatically — through sustained political contestation and deliberate institutional design.

is AI literacy fundamentally a civic capacity?

✓ Correct — ✅ ✓ AI literacy is civic: decisions about AI's values, optimization targets, and governance are being made now. Informed participation in those decisions requires enough understanding to evaluate claims and alternatives.

❌ ❌ AI literacy is civic because AI's societal trajectory is being shaped by current decisions — about values, optimization, governance. Those decisions need informed participants, not just technical experts.

does it mean that AI's societal trajectory is 'not predetermined'?

✓ Correct — ✅ ✓ Not predetermined: AI's societal outcomes depend on contested choices about governance, ownership, values, and policy — not on inevitable technological trajectories. History shows this can go differently based on human choices.

❌ ❌ Not predetermined: AI's societal outcomes — who benefits, what's optimized, how it's governed — are contested and shapeable. Historical examples of effective technology governance show this.

Back to Lesson Start Lab 8

Lab 8

Synthesis: The Future We Choose

Synthesize AI's societal implications and your role in shaping them.

Lab 8 — Synthesis: The Future We Choose

Synthesize the module and develop your position on AI's societal trajectory.

The AI opens with Winner's argument: technologies embed political choices. What political choices are being embedded in AI systems right now — and which are most important to contest?
Drawing on the full module — employment, democracy, healthcare, education, environment, inequality, and power — what is the most important AI governance priority of the next decade?
Address your role: as an informed person in an AI world, what specific contribution do you intend to make to how this technology develops?

This is a synthesis lab. Draw on any and all of the module — be specific about which governance choices matter most and why.

🎯 AI GuideLab 8

Back to Quiz Module Test

Module 4 Test

8 questions covering all lessons. Free, untracked, retake anytime.

is the 'skill reversal' argument about AI and employment?

✓ Correct — ✅ ✓ Skill reversal: historical automation displaced low-skill physical work. AI may displace high-skill cognitive work — reversing the education-as-protection model.

❌ ❌ Skill reversal: historical automation displaced low-skill work. AI may displace high-skill cognitive work — reversing the long-term trend where higher education meant more protection.

'public sphere' is important for democracy because:

✓ Correct — ✅ ✓ The public sphere is a democratic prerequisite: a shared information environment where citizens can form views through deliberation. AI degrades this prerequisite.

❌ ❌ Habermas: the public sphere is a prerequisite for democracy — a shared information environment enabling rational deliberation. Without it, democratic self-governance loses its foundation.

diabetic retinopathy AI example shows:

✓ Correct — ✅ ✓ Retinopathy case: 90%+ validation accuracy → 40% real-world sensitivity on a different population. Validation didn't catch the distribution shift.

❌ ❌ The retinopathy case: validation performance doesn't predict real-world performance when deployment population differs from validation population — the validation gap.

2 Sigma Problem identified:

✓ Correct — ✅ ✓ Bloom's 2 Sigma: one-to-one tutoring = +2 standard deviations. Economically unscalable for most students. AI tutoring is the first plausible answer.

❌ ❌ Bloom's 2 Sigma: one-to-one tutoring = +2 standard deviations over conventional classrooms. The 'problem' was that it was unscalable. AI tutoring is the first plausible solution.

'environmental double role' creates a policy design problem because:

✓ Correct — ✅ ✓ Temporal mismatch: immediate environmental costs vs. long-horizon climate benefits. Markets discount future benefits — policy must bridge this gap.

❌ ❌ Policy design problem: immediate AI environmental costs vs. long-horizon climate benefits. Markets undervalue future benefits — policy must correct the temporal mismatch.

and Johnson's core argument about AI and inequality is:

✓ Correct — ✅ ✓ Acemoglu and Johnson: distribution of AI gains depends on ownership, labor bargaining power, and policy choices — not on the technology.

❌ ❌ Acemoglu and Johnson: AI's distributional outcomes are determined by ownership, labor bargaining power, and policy — not by the technology itself.

structural factor drives concentration of frontier AI development?

✓ Correct — ✅ ✓ Compute requirements for frontier training grow exponentially — creating a structural capital barrier that concentrates frontier AI development in the most well-capitalized organizations.

❌ ❌ Exponentially growing compute requirements create a capital barrier: only organizations with billions can train frontier models. This is a structural concentration mechanism.

Winner's 'artifacts have politics' argument means:

✓ Correct — ✅ ✓ Winner: technologies embed designers' values and constrain future choices. Some technology choices foreclose political options for generations — they have politics built in.

❌ ❌ Winner's argument: technologies embed their designers' values and create structural constraints on future political choices — some foreclosing options for generations.

Back Academy Home