## Introduction {.smaller} **Key Message:** This book challenges popular narratives about AI as an inevitable force or human-vs-machine conflict. Instead, it argues AI is fundamentally about: - Power - Control - Whose interests are served . . . **Opening Question:** Are you scared of AI? The tech industry wants you to be—but they want you scared of the wrong things. --- ## The Dominant Narrative {.smaller} ### The "Humans vs. Machines" Story **In Popular Culture:** - HAL 9000 (2001: A Space Odyssey, 1968) - The Terminator (1984) - The Matrix (1999) **Modern Tech Leaders:** - Elon Musk: "one of the biggest threats to humanity" - Sam Altman: AI poses extinction risk like nuclear war --- ## Two Main Fears {.smaller} ::: {.incremental} 1. **Superintelligent AI will turn against humanity** - Intelligence explosion - Self-preservation drives conflict - Existential risk 2. **Mass unemployment** - Goldman Sachs: 300 million jobs at risk - Inevitable technological displacement - Social unrest ::: . . . **Problem:** This narrative misses the real issues entirely. --- ## What the Old Story Misses {.smaller} ::: {.incremental} 1. **Technology is not fate** - Humans make choices about how AI is built and deployed 2. **AI is not that complicated** - Anyone can understand the basics 3. **The real conflict is between people**, not humans vs. machines 4. **Control matters** - Who decides AI's objectives? 5. **Democratic governance is the answer** - Not just better engineering ::: . . . **Central Argument:** The key issue is WHO controls AI objectives, not whether algorithms are perfectly optimized. --- ## Defining AI {.smaller} ### Core Definition > AI = "Automated decision-making using optimization to maximize some measurable reward" . . . **What This Means:** - Someone types into a computer: "This is the reward we care about" - Algorithm makes decisions to maximize that reward - Examples: maximize ad clicks, minimize costs, predict outcomes . . . **Key Insight:** Since AI requires picking an objective, the crucial question is: **Who gets to pick?** --- ## Real-World AI Applications {.smaller} ### Workplace - Amazon warehouses (robotics) - Uber driver management - Job candidate screening ### Online - Facebook/social media feeds (maximize engagement) - Google search results (maximize ad revenue) - Targeted advertising ($hundreds of billions annually) ### Consequential Domains - Predictive policing - Bail/incarceration decisions - **Warfare** (Gaza: Lavender & "Where's Daddy?" systems) --- ## The Means of Prediction {.smaller} **Definition:** The resources required to build AI . . . ### Four Components ::: {.incremental} 1. **Data** - Training examples, user behavior, surveillance 2. **Computational Infrastructure** - GPUs, data centers, cloud computing 3. **Technical Expertise** - Engineers, researchers, domain knowledge 4. **Energy** - Massive power requirements for training/running models ::: . . . **Who Controls These?** Capital owners, large tech companies, governments **Implication:** Control of these resources = control of AI objectives --- ## How Machine Learning Works {.smaller} ### The Shift from Expert Systems **Before (~1980s):** Expert systems - hand-coded human knowledge **Now:** Machine learning - statistical methods learn from data . . . ### Why Machine Learning Won ::: {.incremental} 1. Built on 200 years of statistical methodology 2. Scales almost arbitrarily with data 3. Leveraged internet data explosion 4. Benefited from computational advances ::: . . . **Key Success Factors:** Data + Computing Power + Statistical Methods --- ## Supervised Learning {.smaller} **Definition:** Predict outcome Y given features X ### How It Works 1. Train on many examples of (X, Y) pairs 2. Learn patterns in the data 3. Make predictions for new X values . . . ### Examples - **Online ads:** Predict click probability → show most profitable ads - **Hiring:** Predict job performance → decide who to interview - **Criminal justice:** Predict recidivism → decide bail/incarceration - **Large language models:** Predict next word → generate text - **Warfare:** Predict Hamas affiliation → target for bombing (10% error acceptable) --- ## The Identification Principle {.smaller} ### Thought Experiment Imagine unlimited data and computing power . . . **What Can Algorithms Learn?** - The probability distribution of Y for any observed value of X - How likely different outcomes are given features . . . **What They CANNOT Learn:** 1. Distribution of Y for values of X never observed in data 2. Causal effects (correlation ≠ causation) . . . ### Crucial Implication for Democracy - Technical details of algorithms don't matter as much - What matters: **What decision problem is being solved?** - Public can debate objectives without needing technical expertise --- ## Overfitting vs. Underfitting {.smaller} ### The Scylla and Charybdis of Machine Learning **Overfitting (Developing Superstitions)** - Extrapolates random patterns from past data - Works great on training data, fails on new data - Example: Seeing a few cars run red lights → concluding it's legal . . . **Underfitting (Being Stubborn)** - Refuses to learn from evidence - Sticks to prior beliefs despite contradictions - Example: Looking left in UK traffic because cars "should" drive on right . . . **Solution:** Balance through model complexity, cross-validation, regularization --- ## Deep Learning & Neural Networks {.smaller} ### What Are Neural Networks? - Artificial neurons loosely inspired by biological brains - Layers of mathematical functions - Millions or billions of parameters (weights) . . . ### Why They Succeeded (2000s onward) ::: {.incremental} - Massive datasets (internet) - Powerful GPUs for computation - Transformer architecture (2017) - revolutionary for language ::: . . . ### Applications - Image recognition - Large language models (ChatGPT, Claude) - Generative AI (text, images, video) **Energy Cost:** Training large models = enormous power consumption --- ## Exploration vs. Exploitation {.smaller} **The Dilemma:** Should you do what seems best (exploit) or try something new (explore)? . . . ### Multi-Armed Bandit Problem - Like choosing slot machines in a casino - Must balance: trying different options vs. using best known option - **Optimism in face of uncertainty:** Try things that might be good . . . ### Real Applications - Medical trials (which treatment to test?) - Online advertising (which ads to show?) - Reinforcement learning (chess, Go, self-driving cars) **AlphaGo/AlphaZero:** Combined deep learning + tree search + self-play --- ## Social Welfare Framework {.smaller} ### Core Question What makes a society good? . . . **Answer:** A society is good if it's good for the people in it. . . . ### Three Sub-Questions ::: {.incremental} 1. **Who counts?** All people affected by AI 2. **How to measure individual welfare?** - Opportunities vs. outcomes? - Objective well-being vs. subjective utility/preferences? 3. **How to aggregate?** Trade-offs between different people's welfare ::: . . . **Welfare Weights:** How much we care about improving welfare of different people (typically: care more about those worst off) --- ## Individual Welfare Concepts {.smaller} ### Different Approaches **Revealed Preferences (Economics)** - What people choose = what they want - Problem: Choices may be constrained, manipulated, or uninformed . . . **Capabilities (Amartya Sen)** - Freedom to achieve valuable functionings - Focus on opportunities, not just outcomes . . . **Primary Goods (John Rawls)** - Basic resources everyone needs - Rights, liberties, income, opportunities . . . **For AI:** Must specify what outcomes we're trying to optimize for --- ## Agents of Change {.smaller} ### Who Can Align AI with Social Welfare? **Requirements:** 1. Leverage over those who control means of prediction 2. Values/interests aligned with social welfare 3. Capacity to effect change . . . ### Key Actors ::: {.incremental} - **Workers** (unions, strikes, organizing) - **Consumers** (boycotts, switching costs, network effects) - **Media/Public Opinion** (bad press, transparency) - **The State** (regulation, antitrust, standards, labor law) - **Courts** (litigation, enforcement) ::: . . . **Problem:** Ideological obfuscation prevents effective action --- ## Ideological Obfuscation {.smaller} ### Three Forms of Misdirection **1. Optimization Errors Framing** - "The algorithm made a mistake" - Hides that objectives themselves are the problem - Example: Algorithm maximizing engagement → promotes outrage (not a bug!) . . . **2. Humans vs. Machines Framing** - Obscures conflicts between different people - Focuses on fictional AI apocalypse - Ignores real harms happening now . . . **3. Technical Expertise Required Framing** - "Too complicated for public debate" - "Leave it to the engineers" - Denies possibility of democratic governance . . . **Effect:** Preserves status quo, prevents collective action --- ## The Control Problem Reframed {.smaller} ### Conventional View (AI Safety/Ethics) - How to align AI with "human values"? - How to make algorithms fair, unbiased, safe? - Focus on better engineering, better optimization . . . ### This Book's View - Who controls the means of prediction? - Whose objectives do algorithms serve? - How to democratize control over AI? . . . ### Example - Hiring Algorithm - **"Fairness" approach:** Does it predict productivity without bias? - **Welfare approach:** Whose job prospects improve? Who benefits? . . . **The Real Issue:** Objectives chosen by capital owners are good for profits, not society --- ## Value Alignment Problem {.smaller} ### The Paper Clip Maximizer Thought Experiment - AI told to maximize paper clip production - Converts entire world into paper clips - Humans try to stop it → AI fights back - **Problem:** Incomplete objective specification . . . ### Real-World Analogs - **Social media:** Maximize engagement → outrage, addiction, mental health harms - **Hiring:** Maximize predicted productivity → discriminates against caregivers - **Health insurance:** Maximize profits → screens out people who need care - **Gaza bombing:** Maximize suspected Hamas kills → 10% error rate acceptable, 15-20 civilian deaths per target --- ## Value Alignment Solutions {.smaller} ### Solutions Attempted - Inverse reinforcement learning (infer objectives from behavior) - Principal-agent models (alignment through incentives) . . . **Problem:** These assume benevolent principals. What if principals' interests conflict with social welfare? . . . ### Real Solution Democratic control over who sets objectives --- ## Privacy and Data Governance {.smaller} ### The Privacy Problem - Machine learning requires vast data - Data collection creates negative externalities (surveillance, manipulation) - Individual consent is inadequate (systemic effects) . . . ### Technical Solutions **Randomized Response:** Add noise to individual data **Differential Privacy:** Guarantee limited information leakage - Privacy budget: Total allowed information disclosure - Trade-off: More privacy = less accuracy . . . ### Governance Solutions - Regulate data collection (GDPR, etc.) - Data trusts (collective governance) - Pigou taxes on data collection - **Democratic control of surveillance** --- ## Automation and the Workplace {.smaller} ### The Luddite Question Will AI cause mass unemployment? . . . ### Economic Theory - Technology affects relative productivity of workers - Can complement some workers (increase wages) - Can substitute for others (decrease wages) - Distribution of gains depends on labor market structure . . . ### With AI - Some augmentation (making workers more productive) - Some replacement (automating tasks) - **Outcome depends on choices, not technological determinism** . . . **Critical Question:** Who decides what gets automated and how? --- ## Democratic Workplace Technology {.smaller} ### Scandinavian Model (1970s-80s) UTOPIA, DEMOS projects - Workers participate in technology design - Unions involved in decisions - Focus on augmentation over replacement - Collective resources approach . . . ### Principles ::: {.incremental} 1. Workers should have input on technology affecting their work 2. Technology should augment, not just replace 3. Codetermination: workers on company boards ::: . . . ### Examples - **Nursing:** AI for diagnosing conditions vs. robot replacements - **Warehousing:** Assistive tools vs. full automation + surveillance --- ## Fairness vs. Welfare Paradigms {.smaller} ### Fairness Paradigm (Gary Becker) - Discrimination = departing from profit maximization - Non-monetary considerations in decisions - Fair = perfectly optimized for profits - No discrimination when predictions are accurate . . . ### Welfare Paradigm (Martin Luther King Jr.) - Focus on outcomes for disadvantaged groups - Democracy, integration, brotherhood - Who wins and loses from algorithm? - Baseline inequality and heterogeneous treatment effects matter . . . ### Example - Affirmative Action - **Fairness view:** Bad (hurts profits, introduces bias) - **Welfare view:** Can be good (reduces inequality, helps disadvantaged) --- ## Algorithmic Fairness Definitions {.smaller} ### Common Definitions (all related) - Balance for positive/negative class - Equality of false positives/negatives - Equality of opportunity - Conditional statistical parity - Individual fairness . . . ### Unifying Theme - All converge when predictions are perfect - All say "no bias" when profits are maximized - Focus on treatment equality, not welfare equality . . . ### Problem with Equal Treatment Like giving everyone same eyeglasses - Doesn't account for baseline inequality - Doesn't account for heterogeneous effects - Can worsen outcomes for disadvantaged --- ## Explainability - Three Types {.smaller} ### 1. Explaining Decision Functions (for engineers) - How does algorithm map inputs → outputs? - Methods: Simple models, surrogate models, feature importance - Goal: Debug, ensure robustness, transferability . . . ### 2. Explaining Decisions (for those affected) - Why was this particular decision made? - Methods: Counterfactuals ("what if"), local explanations - Goal: Individual recourse, contestability . . . ### 3. Explaining Decision Problems (for democratic governance) - What is the system trying to achieve? - What data/features does it use? - **Most Important for Democratic Governance** . . . **Kafka's The Trial:** Protagonist never learns crime, verdict, or reasons --- ## The Limits of Technical Solutions {.smaller} ### Why "Just Make Engineers Nice" Doesn't Work - Engineers work for profit-maximizing corporations - Individual ethics constrained by structural requirements - Can be fired for refusing profitable-but-harmful projects - Scope of individual agency is limited . . . ### Corporate Social Responsibility (CSR) - Often just public relations - Doesn't change fundamental profit motive - Can be abandoned when costly . . . ### Examples of Failure - Facebook: Frances Haugen whistleblower - Amazon: Warehouse conditions despite employee concerns - Google: Timnit Gebru fired for research on AI harms . . . **Conclusion:** Need structural changes, not just individual ethics --- ## Forms of Democratic Governance {.smaller} ### Beyond Representative Democracy **1. Sortition (Random Selection)** - Used in ancient Athens - Modern: Citizens' assemblies, juries - Advantages: No professional politicians, allows deliberation - Stratified sampling for representation . . . **2. Liquid Democracy** - Everyone can vote on all issues - Can delegate vote to anyone (revocable, issue-specific) - Developed by Lewis Carroll (19th century) - Tools: LiquidFeedback software . . . **3. Workplace Democracy** - Workers on corporate boards (codetermination) - Democratic decision-making in firms - Collective bargaining over technology --- ## Democratic Control - Why It Matters {.smaller} ### Current State - Objectives set by capital owners - Maximize profits, not social welfare - Those affected by AI have no say . . . ### Democratic Alternative - Those impacted by AI decisions get input - Public deliberation about objectives - Redistribute power, not just wealth . . . ### Requirements for Real Democracy ::: {.incremental} 1. **Actual power** - Not just "consulting" or "participatory design" 2. **Deliberation** - Not just voting 3. **Multiple levels** - Workplace, local, national, global 4. **Public understanding** - Simplified through identification principle ::: --- ## Key Mechanisms for Control {.smaller} ### Regulatory Interventions - Antitrust enforcement (break up monopolies) - Data protection laws (GDPR) - Labor law (strengthen unions) - Standards and interoperability requirements . . . ### Economic Interventions - Pigou taxes on externalities (data collection, energy use) - Public ownership of computational infrastructure - Socialized data centers - Data trusts/commons . . . ### Social Movements - Worker organizing - Consumer advocacy - Public pressure campaigns - Litigation --- ## Ancient Questions, Modern Context {.smaller} ### AI Confronts Fundamental Questions **Epistemology (How do we know?)** - Learning from experience - Overfitting vs. underfitting - Exploration vs. exploitation - Causation vs. correlation . . . **Ethics (What should we do?)** - Individual welfare measurement - Social welfare aggregation - Fairness vs. equality . . . **Political Philosophy (How should we govern?)** - Democracy and deliberation - Power and control - Agency and accountability . . . **These aren't new questions** - they're ancient philosophical problems given new urgency by AI --- ## The Path Forward {.smaller} ### Core Recommendations ::: {.incremental} 1. **Reject Inevitability Narratives** - Technology is choice, not fate 2. **Focus on Control of Objectives** - Not optimization errors, but who decides? 3. **Democratize the Means of Prediction** - Data, computation, expertise, energy 4. **Create Institutions for Deliberation** - Multiple forms of democracy 5. **Build Broad Coalitions** - Workers, consumers, citizens across domains ::: --- ## Concrete Examples of Change {.smaller} ### Workplace - Union veto over surveillance technology - Worker input on automation decisions - Codetermination boards for AI deployment ### Consumer Protection - Bans on manipulative algorithms - Interoperability requirements - Data portability rights ### Criminal Justice - Democratic oversight of predictive policing - Transparency requirements - Community control over surveillance ### Online Platforms - User control over algorithms - Public-interest alternatives - Break up monopolies --- ## Addressing Objections {.smaller} **"AI is too complex for democratic governance!"** → Identification principle: Focus on objectives, not technical details . . . **"The market will solve these problems!"** → Market incentives = profit, not social welfare . . . **"We need to move fast or China will win!"** → Race to the bottom harms everyone . . . **"Engineers can be ethical!"** → Individual ethics insufficient against structural profit motive . . . **"This is unrealistic/utopian!"** → Current system is failing; alternatives exist --- ## Summary - Main Arguments {.smaller} ### 1. The Real Problem Not human vs. machine, but who controls AI objectives Capital owners maximize profits ≠ social welfare . . . ### 2. The Means of Prediction Data, computation, expertise, energy Control of these = control of AI Currently concentrated in few hands . . . ### 3. The Solution Democratic governance of AI - multiple forms, multiple levels Real power for those affected --- ## Summary (continued) {.smaller} ### 4. Why This Matters - AI amplifies existing power structures - Without intervention, increases inequality - With democratic control, can serve everyone . . . ### 5. How to Get There - Worker organizing - Regulation and public ownership - New democratic institutions - Reject ideological obfuscation --- ## Discussion Questions {.smaller} ::: {.incremental} 1. What AI systems affect your life? Who controls their objectives? 2. In your domain (work, study, community), who should have input on AI decisions? 3. What forms of democratic governance seem most promising? Most feasible? 4. How can we build coalitions for democratic AI control? 5. What are the biggest obstacles? How to overcome them? 6. How do we balance expertise and democracy in technical decisions? 7. What can individuals do? What requires collective action? ::: --- ## Key Takeaways {.smaller} ::: {.incremental} - **AI is not fate** - it embodies human choices about power and priorities - **The conflict is between people** - not humans vs. machines - **Control of objectives is key** - not optimization or technical details - **The means of prediction matter** - data, computation, expertise, energy - **Democracy is the answer** - but needs institutional innovation - **Change is possible** - through workers, consumers, regulation, organizing - **Ancient questions, modern stakes** - epistemology, ethics, political philosophy ::: . . . **Your voice matters** - in deliberation and collective action --- ## Final Message {.center} > The future of AI is not predetermined. > > It will be shaped by struggles over control, power, and democracy. > > The question is: **Will you participate?** --- ## Further Reading {.smaller} ### The Book Kasy, Maximilian. *The Means of Prediction: How AI Really Works (and Who Benefits)*. University of Chicago Press, 2025. . . . ### Related Works - Kate Crawford, *The Atlas of AI* (2021) - Daron Acemoglu & Simon Johnson, *Power and Progress* (2023) - Shoshana Zuboff, *The Age of Surveillance Capitalism* (2019) - Hélène Landemore, *Open Democracy* (2020) - Erik Olin Wright, *How to Be an Anti-Capitalist* (2019) . . . ### Key Topics Differential privacy • Value alignment • Democratic workplace design • Fairness in ML • Externalities and public goods --- ## Thank You {.center} Questions?
**Contact & Resources:** - Book: University of Chicago Press - Author: Maximilian Kasy (University of Oxford) - Topics: AI Ethics, Democratic Governance, Political Economy of Technology