Lesson 1

AI and Jobs

Automation, displacement, and the transformation of work.

Between 2000 and 2010, US manufacturing lost about 5.6 million jobs. Economists debated how much was automation vs. trade. The debate continues with AI: a 2023 Goldman Sachs report estimated generative AI could affect 300 million full-time jobs globally — not necessarily eliminating them, but transforming what those jobs involve. Some economists emphasize job creation from new industries; others emphasize that historical job transitions took decades and caused real suffering for the people in transition.

What the Research Shows

The economics of AI and employment is genuinely contested:

Task displacement, not job elimination: Most jobs contain some tasks AI can automate and others it can't. AI tends to automate specific tasks within jobs rather than eliminating jobs entirely — in the short term.
The transition problem: Even if total employment stays stable, displaced workers often lack skills for new jobs. Historical transitions (agricultural to manufacturing, manufacturing to services) took decades and generations.
Unequal exposure: High-education knowledge workers may see productivity gains; routine cognitive workers face higher displacement risk.

Policy Responses

Retraining and workforce development programs
Portable benefits (not tied to specific employers)
Automation taxes to fund transition support
Reduced work weeks to share productivity gains

The Honest Uncertainty

Economists disagree about AI's net employment effects. What's clear: the transition costs are real, unevenly distributed, and policy can either mitigate or amplify them.

Back Take Quiz

Quiz 1

AI and Jobs

4 questions — free, untracked, retake anytime.

does 'task displacement' mean in the context of AI and employment?

✓ Correct — ✅ ✓ Task displacement: AI automates specific tasks within jobs. Most jobs have some automatable tasks and some that aren't — so the first-order effect is job transformation, not elimination.

❌ ❌ Task displacement means AI automates specific tasks within jobs rather than eliminating jobs entirely — at least in the short term. Jobs transform before they disappear.

is the 'transition problem' significant even if total employment stays stable?

✓ Correct — ✅ ✓ Even if aggregate employment is stable, transitions are hard: displaced workers need new skills, and historical transitions have taken decades. The aggregate doesn't capture individual harm.

❌ ❌ The transition problem: even if total jobs stay stable, displaced workers need retraining, and transitions take decades. The aggregate stability doesn't prevent individual suffering.

workers face the highest displacement risk from AI?

✓ Correct — ✅ ✓ Routine cognitive work — predictable, pattern-following tasks — is most exposed to AI automation. Physical work and high-level judgment work face less near-term risk.

❌ ❌ Routine cognitive work (predictable, pattern-following tasks) is most exposed. Physical labor and high-level judgment work face less near-term displacement risk.

does an 'automation tax' proposal aim to do?

✓ Correct — ✅ ✓ Automation taxes: tax the productivity gains from automation to fund retraining and transition support for displaced workers — redistributing some of the gains.

❌ ❌ An automation tax would tax productivity gains from automation, using the revenue to fund retraining and transition support for displaced workers.

Back to Lesson Start Lab 1

Lab 1

Employment Policy Analysis

Develop a policy framework for AI-driven employment transition.

Lab 1 — Employment Policy Analysis

Analyze AI employment effects and develop a policy framework.

The AI opens: economists disagree on AI's net employment effects. What policy framework would you advocate for a government facing significant AI-driven task displacement in the next decade?
Evaluate the available policy tools against their limitations.
Address: how do you design policy that helps displaced workers without slowing beneficial AI adoption?

Consider: retraining timelines, the uneven distribution of displacement, and what 'transition support' actually requires in practice.

🔬 AI GuideLab 1

Back to Quiz Next: Lesson 2

Lesson 2

AI and Democracy

Disinformation, algorithmic amplification, and AI's effects on political systems.

The 2024 US election cycle saw the first widespread use of AI-generated political content: synthetic audio of candidates, AI-written targeted ads, and AI-generated voter suppression messages. A January 2024 robocall used a voice clone of President Biden discouraging New Hampshire Democrats from voting in the primary. The call was illegal — but it reached thousands of voters before being identified. Election security researchers noted that the marginal cost of influence operations had dropped to near zero.

AI's Democratic Risks

Synthetic disinformation: AI-generated audio, video, and text can fabricate candidate positions, fake scandals, or suppress turnout — at scale, at near-zero cost
Micro-targeted persuasion: AI enables psychological profiling and individually tailored political messaging designed to exploit specific vulnerabilities
Algorithmic amplification: Recommendation systems amplify emotionally activating (often divisive) content, distorting what citizens believe is the mainstream view
Epistemic collapse: When any content can be fabricated, citizens lose the ability to distinguish real from synthetic — producing paralysis or cynicism rather than informed participation

Democratic Safeguards

Synthetic content disclosure laws (several states enacted in 2024)
Platform takedown obligations for synthetic political content
Media literacy education
Content provenance standards (C2PA)

The Hard Problem

None of these safeguards are sufficient alone. The asymmetry between generating synthetic content (cheap, fast) and detecting and correcting it (expensive, slow) is fundamental — and gets worse as generation quality improves.

Back Take Quiz

Quiz 2

AI and Democracy

4 questions — free, untracked, retake anytime.

made the 2024 Biden voice clone robocall significant?

✓ Correct — ✅ ✓ The case showed the operational reality: synthetic disinformation can spread faster than detection and correction — and the cost of production has dropped to near zero.

❌ ❌ The significance: synthetic political disinformation reaches voters before detection, at near-zero production cost. The asymmetry between generation and correction is the problem.

is 'epistemic collapse' as a democratic risk from AI?

✓ Correct — ✅ ✓ Epistemic collapse: if citizens can't trust any video, audio, or text as authentic, the information environment that democracy requires breaks down — producing cynicism rather than participation.

❌ ❌ Epistemic collapse: when synthetic content is indistinguishable from real, citizens lose the epistemic foundation democracy requires — producing paralysis or cynicism rather than informed participation.

is the 'fundamental asymmetry' in synthetic political content?

✓ Correct — ✅ ✓ The fundamental asymmetry: generating synthetic content is cheap and fast; detecting it and correcting the public record is expensive and slow. This gap is structural.

❌ ❌ The fundamental asymmetry: generating synthetic content = cheap and fast. Detecting and correcting it = expensive and slow. This structural gap worsens as generation quality improves.

does algorithmic amplification distort democratic discourse?

✓ Correct — ✅ ✓ Algorithmic amplification: emotionally activating content gets more engagement, so recommendation systems surface more of it — distorting what citizens perceive as mainstream.

❌ ❌ Algorithmic amplification: emotionally activating (often divisive) content drives engagement, so recommendation systems surface it preferentially — distorting perceptions of mainstream views.

Back to Lesson Start Lab 2

Lab 2

Democratic Safeguards Analysis

Develop a governance framework for AI and democracy.

Lab 2 — Democratic Safeguards Analysis

Analyze AI's democratic risks and develop a governance framework.

The AI opens: given the fundamental asymmetry between synthetic content generation and detection/correction, what governance framework gives democracy the best chance of surviving AI-powered disinformation?
Evaluate the available safeguards against their limitations.
Address: what obligations do AI companies have when their tools are used for democratic manipulation?

Consider: platform obligations, legal frameworks, provenance standards, and the international dimension — disinformation crosses borders, governance doesn't.

🔬 AI GuideLab 2

Back to Quiz Next: Lesson 3

Lesson 3

AI and Healthcare

Diagnostic AI, drug discovery, and the transformation of medicine.

In 2020, DeepMind's AlphaFold solved protein structure prediction — a 50-year grand challenge in biology. Within two years, it had predicted structures for nearly all known proteins, accelerating drug discovery across diseases from malaria to cancer. Simultaneously, diagnostic AI systems for radiology, pathology, and ophthalmology began outperforming human specialists on specific tasks. The promise of AI in healthcare is significant and real. So are the risks: biased training data, over-reliance on AI diagnosis, and questions about who benefits from AI-driven healthcare advances.

Where AI Is Transforming Healthcare

Drug discovery: AlphaFold and similar tools have dramatically accelerated the identification of drug targets — potentially compressing timelines from years to months
Diagnostic imaging: AI systems for reading X-rays, MRIs, pathology slides, and retinal scans are approaching or exceeding specialist accuracy on specific conditions
Clinical decision support: AI tools that surface relevant information and flag potential errors during clinical care
Administrative burden: AI-assisted documentation to reduce physician burnout

Healthcare AI Risks

Diagnostic bias: Systems trained on non-representative populations perform worse for underrepresented groups
Automation bias in clinical settings: Physicians may defer to AI recommendations without sufficient scrutiny
Access inequality: AI-enhanced healthcare may benefit wealthy countries and patients first, widening health disparities

The Distribution Question

AlphaFold's protein predictions are publicly available — anyone can use them. This is the exception, not the rule. Most healthcare AI is proprietary. Who benefits from AI healthcare advances depends heavily on how IP and access are structured.

Back Take Quiz

Quiz 3

AI and Healthcare

4 questions — free, untracked, retake anytime.

was AlphaFold's protein structure prediction a landmark achievement?

✓ Correct — ✅ ✓ AlphaFold solved protein structure prediction — a 50-year unsolved problem. Its public release made drug target identification dramatically faster across many diseases.

❌ ❌ AlphaFold solved protein structure prediction — a 50-year grand challenge — and made those predictions publicly available, accelerating drug discovery across diseases.

is the 'access inequality' risk in healthcare AI?

✓ Correct — ✅ ✓ Most healthcare AI is proprietary and expensive. Without active policy to address access, AI healthcare advances may benefit wealthy populations first — widening disparities.

❌ ❌ Access inequality: most healthcare AI is proprietary and expensive. Without active access policies, AI health advances benefit wealthy populations first and widen global health disparities.

is automation bias particularly dangerous in clinical settings?

✓ Correct — ✅ ✓ Automation bias in clinical settings: if physicians defer to AI without scrutiny, a tool designed to support clinical judgment becomes a substitute for it — with potentially deadly consequences.

❌ ❌ Automation bias: physicians deferring to AI without scrutiny convert a decision support tool into an autonomous decision-maker. This is dangerous when the AI is wrong.

makes diagnostic AI both promising and risky?

✓ Correct — ✅ ✓ Diagnostic AI is genuinely capable — approaching specialist accuracy — but trained on non-representative data. It may work well for the populations in its training data and poorly for others.

❌ ❌ Diagnostic AI approaches specialist accuracy but often trains on non-representative populations — meaning it may work well for some groups and poorly for others.

Back to Lesson Start Lab 3

Lab 3

Healthcare AI Policy

Develop a framework for equitable healthcare AI deployment.

Lab 3 — Healthcare AI Policy

Analyze the distribution of benefits and risks from AI in healthcare.

The AI opens: AlphaFold made protein predictions publicly available; most healthcare AI is proprietary. What IP and access policies would best ensure that AI healthcare advances benefit the widest population?
Develop a framework for healthcare AI deployment that addresses bias, automation risk, and access.
Address: what regulatory requirements should govern diagnostic AI before clinical deployment?

Consider: FDA-style approval processes, real-world performance monitoring, demographic performance requirements, and open-access models.

🔬 AI GuideLab 3

Back to Quiz Next: Lesson 4

Lesson 4

AI and Education

Personalized learning, academic integrity, and the transformation of education.

Within weeks of ChatGPT's release in late 2022, teachers reported a wave of AI-generated essays. Schools responded with everything from bans to AI detectors (which proved unreliable). But the more significant disruption may be structural: if AI can complete most standard academic assignments, what is education for? Some educators argued for redesigning assessment around tasks AI can't do; others argued that AI is a tool students should learn to use, as calculators were a tool math students learned to use.

AI in Education — The Opportunity

Personalized learning: AI tutors can adapt to individual pace, identify gaps, and provide instant feedback — potentially democratizing access to high-quality instruction
Teacher support: AI can reduce administrative burden (grading, lesson planning) and free teachers for higher-value human interaction
Accessibility: AI translation, text-to-speech, and adaptive interfaces can support students with different learning needs

AI in Education — The Challenge

Academic integrity: AI-generated work is hard to detect and getting harder — requiring redesigned assessment
Learning vs. credential: Using AI to complete assignments may produce credentials without learning — devaluing education
Dependency risks: Students who offload writing, problem-solving, and analysis to AI may atrophy skills they need
Equity: AI tutoring tools vary widely in quality; access varies by income

The Core Question

If AI can do most of what we ask students to do, we need to ask: what skills does being human in an AI world require — and design education to build those.

Back Take Quiz

Quiz 4

AI and Education

4 questions — free, untracked, retake anytime.

is the most significant structural disruption AI poses to education?

✓ Correct — ✅ ✓ The structural disruption: if AI can do the assignments, the assignments no longer test learning. This forces educators to rethink what assessment is actually for.

❌ ❌ The structural disruption: standard assignments no longer reliably measure learning if AI can complete them. Education needs to redesign assessment around what AI can't replace.

calculator analogy for AI in education suggests:

✓ Correct — ✅ ✓ The calculator analogy: calculators were controversial in math education and are now standard tools. AI may follow the same arc — a tool students learn to use, not avoid.

❌ ❌ The calculator analogy: calculators were once controversial in math education and are now standard. AI may similarly become a tool students are taught to use effectively.

is the 'learning vs. credential' risk in educational AI use?

✓ Correct — ✅ ✓ Learning vs. credential: if students use AI to produce the work without doing the learning, the credential (grade, degree) becomes decoupled from actual capability.

❌ ❌ Learning vs. credential: AI can produce the credential-generating output (essay, problem solution) without the student doing the learning — decoupling credentials from capability.

is the equity concern about AI tutoring tools?

✓ Correct — ✅ ✓ Equity: AI tutoring tools vary in quality and access varies by income. Without deliberate equity focus, AI education tools may widen existing disparities.

❌ ❌ Equity concern: AI tutoring quality varies widely and access varies by income. Without deliberate focus on equity, AI education tools may widen rather than narrow disparities.

Back to Lesson Start Lab 4

Lab 4

Educational AI Framework

Design a framework for AI in education.

Lab 4 — Educational AI Framework

Design a framework for AI in education that maximizes benefits and minimizes harms.

The AI opens: if AI can complete most standard assignments, what should education assess instead? Propose a redesigned assessment framework.
Address where AI should and shouldn't be permitted in educational settings.
Develop a position on the 'calculator analogy' — is learning to use AI well a core educational goal?

Consider: what skills humans need in an AI world, what AI can't replace, and how to design for equity across income levels.

🔬 AI GuideLab 4

Back to Quiz Next: Lesson 5

Lesson 5

AI and the Environment

Energy consumption, climate modeling, and AI's environmental double role.

Training GPT-4 was estimated to consume roughly the equivalent of a small country's electricity for several weeks. Inference — running queries on deployed models — consumes energy continuously at scale. A Goldman Sachs analysis estimated that a ChatGPT query uses roughly 10x the energy of a Google search. At the same time, AI is being used for climate modeling, energy grid optimization, materials discovery for clean energy, and reducing logistics inefficiencies. AI is simultaneously a significant contributor to energy demand and a potential tool for addressing climate change.

AI's Environmental Costs

Training costs: Large model training requires significant compute — and therefore energy. Carbon footprint varies by energy source of the data center.
Inference costs: Continuous deployment at scale adds up — especially as AI becomes integrated into everyday queries
Hardware: GPU manufacturing requires rare materials; hardware cycles create e-waste
Water cooling: Data centers require significant water for cooling

AI as a Climate Tool

Climate modeling at higher resolution and lower cost than previous approaches
Electrical grid optimization — reducing waste in power distribution
Materials discovery for batteries, solar cells, and other clean energy components
Agricultural optimization to reduce emissions and improve efficiency

The Net Question

Whether AI is net positive or negative for climate depends on whether its applications to climate solutions outpace its energy footprint — a question that depends heavily on policy, energy sourcing, and which AI applications are prioritized.

Back Take Quiz

Quiz 5

AI and the Environment

4 questions — free, untracked, retake anytime.

is a ChatGPT query more energy-intensive than a Google search?

✓ Correct — ✅ ✓ LLM text generation is computationally much more intensive than indexed search retrieval — requiring significantly more energy per query.

❌ ❌ Generating text requires significantly more computation than retrieving pre-indexed search results. LLM inference is computationally expensive in a way search is not.

does AI's 'environmental double role' mean?

✓ Correct — ✅ ✓ AI's double role: significant energy consumer and carbon emitter on one hand; powerful tool for climate modeling, grid optimization, and clean energy materials discovery on the other.

❌ ❌ AI's double role: it contributes to energy demand and emissions while also being a potential tool for climate solutions. Whether it's net positive depends on policy and prioritization.

does the carbon footprint of training a model vary by location?

✓ Correct — ✅ ✓ Carbon footprint depends on the energy source. Training in a region powered by renewables produces far less CO2 than training in a region powered by coal.

❌ ❌ The carbon footprint of training depends on the energy source of the data center. Renewable-powered data centers produce far less CO2 than fossil-fuel-powered ones.

would determine whether AI is net positive or negative for climate?

✓ Correct — ✅ ✓ The net question: do AI's climate applications (grid optimization, materials discovery, modeling) deliver more carbon reduction than AI's direct energy consumption adds? Depends heavily on policy and prioritization.

❌ ❌ Net climate impact: whether AI's climate solution applications outpace its direct energy footprint. This depends on policy, energy sourcing, and which AI applications are prioritized.

Back to Lesson Start Lab 5

Lab 5

AI and Climate Policy

Develop a climate policy framework for AI.

Lab 5 — AI and Climate Policy

Analyze AI's environmental impact and develop a climate policy framework for AI.

The AI opens: if AI has significant energy costs but also potential climate benefits, how do you design policy to maximize the ratio of climate benefit to climate cost?
Develop a framework for environmental AI governance — what requirements, incentives, and restrictions would you implement?
Address: should AI companies be required to disclose energy consumption and carbon footprint per product?

Consider: mandatory disclosure, renewable energy requirements, prioritization of high-climate-value AI applications, and hardware lifecycle standards.

🔬 AI GuideLab 5

Back to Quiz Next: Lesson 6

Lesson 6

AI and Inequality

How AI's benefits and harms are distributed — and what drives the distribution.

A 2021 study by researchers at MIT found that facial recognition systems had substantially higher error rates for darker-skinned women than lighter-skinned men — up to 34 percentage points difference in some systems. These systems were being used in hiring platforms, law enforcement, and financial services. The people most harmed by the AI errors were also among the most vulnerable. AI systems deployed without equity consideration tend to amplify existing inequalities — because existing inequalities are embedded in the data and systems AI learns from.

How AI Amplifies Inequality

Biased training data: Historical data reflects past discrimination — AI learns and replicates these patterns
Unequal access: AI productivity tools benefit those with access; those without fall further behind
High-stakes deployments: AI in hiring, lending, criminal justice, and healthcare affects life outcomes — and errors harm vulnerable populations disproportionately
Geographic concentration: AI development and its economic benefits concentrate in a few cities and countries, widening geographic inequality

How AI Could Reduce Inequality

AI tutoring democratizing access to high-quality education
AI diagnostic tools extending specialist-quality healthcare to underserved areas
AI translation reducing language barriers
AI-assisted legal tools providing access to legal guidance for those who can't afford lawyers

The Pattern

Whether AI increases or decreases inequality depends almost entirely on who it's designed for, who has access to it, and who bears the cost of its errors. These are policy choices, not technical inevitabilities.

Back Take Quiz

Quiz 6

AI and Inequality

4 questions — free, untracked, retake anytime.

do AI systems deployed without equity consideration tend to amplify existing inequalities?

✓ Correct — ✅ ✓ Historical data embeds historical discrimination. AI trained on that data learns and replicates those patterns — amplifying existing inequalities unless actively designed to address them.

❌ ❌ AI amplifies existing inequalities because training data reflects historical discrimination. AI learns those patterns and replicates them in new contexts.

makes facial recognition errors in hiring or criminal justice more serious than average AI errors?

✓ Correct — ✅ ✓ High-stakes decisions + highest error rates for most vulnerable populations = compounded harm. Errors in hiring, lending, and criminal justice determine life outcomes, not just user experience.

❌ ❌ High stakes + highest error rates for the most vulnerable: errors in hiring, criminal justice, and lending determine life outcomes — and the groups with highest error rates are often already disadvantaged.

could AI tutoring reduce inequality if deployed equitably?

✓ Correct — ✅ ✓ AI tutoring could democratize access to high-quality personalized instruction — something previously only available to those who could afford private tutors.

❌ ❌ AI tutoring could democratize personalized, high-quality instruction — giving students in underserved areas access to what was previously only available through expensive private tutoring.

determines whether AI increases or decreases inequality?

✓ Correct — ✅ ✓ Distribution of AI's benefits and harms is determined by design choices and policy decisions — not by technology itself. These are choices, not inevitabilities.

❌ ❌ Whether AI increases or decreases inequality depends on policy choices: who it's designed for, who has access, and who bears the cost of errors. These aren't technical inevitabilities.

Back to Lesson Start Lab 6

Lab 6

AI Equity Framework

Develop a comprehensive framework for AI equity.

Lab 6 — AI Equity Framework

Develop a comprehensive framework for AI equity.

The AI opens: given that AI tends to amplify existing inequalities without deliberate intervention, what policy and design requirements would you mandate for high-stakes AI deployments?
Develop your equity framework across design, deployment, and access dimensions.
Address: what democratic mechanisms should govern who bears the costs and who receives the benefits of AI?

Consider: demographic performance requirements, mandatory access provisions, community input in high-stakes deployments, and benefit-sharing mechanisms.

🔬 AI GuideLab 6

Back to Quiz Module Test

Module 4 Test

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

tends to automate specific tasks within jobs rather than eliminating jobs entirely. This is called:

✓ Correct — ✅ ✓ Task displacement: AI automates specific tasks within jobs, transforming what those jobs involve before (and if) it eliminates them.

❌ ❌ Task displacement: AI automates tasks within jobs rather than eliminating jobs entirely — at least in the short term.

fundamental challenge of AI-powered political disinformation is:

✓ Correct — ✅ ✓ The fundamental asymmetry: generating synthetic political content is cheap and fast; detecting it and correcting the record is expensive and slow.

❌ ❌ The fundamental asymmetry: generation is cheap and fast; detection and correction are expensive and slow. This gap is structural and worsens as generation improves.

was significant for healthcare AI because:

✓ Correct — ✅ ✓ AlphaFold solved protein structure prediction, a 50-year grand challenge, and made those predictions publicly available — dramatically accelerating drug discovery.

❌ ❌ AlphaFold solved a 50-year grand challenge (protein structure prediction) and made results public — potentially accelerating drug discovery across many diseases.

'learning vs. credential' risk in educational AI use means:

✓ Correct — ✅ ✓ Learning vs. credential: AI can generate the work that produces the credential without the student developing the underlying capability.

❌ ❌ AI can produce the work that generates credentials (grades, degrees) without the student actually learning. This decouples credentials from capability.

'environmental double role' refers to:

✓ Correct — ✅ ✓ AI's double role: significant energy consumer and carbon contributor, while also a potential tool for grid optimization, climate modeling, and clean energy materials discovery.

❌ ❌ AI's environmental double role: it contributes to energy demand and emissions while also being a potential tool for climate solutions.

AI increases or decreases inequality depends primarily on:

✓ Correct — ✅ ✓ The distribution of AI's benefits and harms is determined by design and policy choices — not by the technology itself.

❌ ❌ Whether AI increases or decreases inequality is determined by policy choices: who it's designed for, who has access, and who bears the cost of errors.

Back Academy Home