Study Guide

Format

• Covers weeks 1 to 9 with the focus on weeks 5 to 9, from Privacy Foundations to Making Silicon Valley 3.
• Two short answers at 10 points each, 30 minutes each. One long analytical essay at 20 points, 60 minutes. It is a hypothetical scenario and asks for the human contexts, the ethical stakes, and a specific recommendation.
• Graders look for four things in order: organize by what the question asks for, define every concept you use, explain every case with its context, then apply the concept to the case and argue. Name the author and title whenever you use a reading.
• Short-answer habit: paragraph 1 states the argument of the reading or lecture with its distinctions. Paragraph 2 explains the case or concept in its own context. Paragraph 3 applies and takes a position.
• Essay habit: read the scenario twice, write the tension in one sentence, pick two or three lenses from different units, explain each concept before applying it, end with one recommendation and what it costs.

Summary

1. Institutions make social order by making people legible. States, firms, and platforms simplify lives into data so they can govern from a distance. This is Scott's legibility and Foucault's discipline, and it runs from the census to the AFST to the platform data imperative. Tools: Power, Classification, Representation.
2. Privacy as individual control over information is a 1970s artifact that no longer fits. FIPs and PII assume data already exists and that consent at sign-up settles things. Malkin replaces this with contextual integrity, appropriate flows that match the norms of a context. Pressly goes further: privacy protects oblivion, the not-knowing that exists before information is created. Cases: Havasupai, Crisis Text Line, TSA scanners.
3. Surveillance disciplines by making people watch themselves, and it falls unevenly. Bentham's panopticon makes power visible but unverifiable, so the inmate becomes the principle of his own subjection. The NYPD's Domain Awareness System and the NSA show the digital version: bulk, cheap, repurposable, and with differential exposure.
4. Numbers persuade because of politics, and prediction is the new paradigm of governance. Porter shows quantification is a technology of distance adopted by weak elites under pressure from outsiders. Decision-making itself is a post-1945 product of cybernetics, game theory, and behaviorism. Automated decision systems extend this into a panoptic sort: COMPAS, AFST, Ofqual, credit scores, Fourcade and Healy's ordinal society.
5. Fairness and risk have a history, and that history is neoliberal. Risk moved from collective pools in the welfare state to actuarial fairness, paying for your own risk, through the 1970s insurance debates and welfare automation. Fairness as accuracy turns a social question into an engineering one. This is the lecture argument about co-production with neoliberalism, and it is where Winner's inherently political technology applies.
6. Futures are narratives that do work in the present. Narrative templates like 1984, Terminator, the Hero, and Andreessen's techno-capital machine decide what gets funded and feared. Le Guin's carrier bag, Haraway's cyborg, and Afrofuturism are alternatives. Hildebrandt's Diana shows technological regulation without law.
7. Automation and Silicon Valley are about capital, the military, and hidden labor, told as a story about genius. Noble shows numerical control won because the Air Force and management wanted control, not because it was better. Nicholas names three builders of the Valley: universities, military spending, and a culture and legal climate that tolerated failure and voided non-competes. Crawford, Sloane, Vinsel and Russell, and Fang supply the labor, finance, maintenance, and military layers the garage myth leaves out.

Week 5. Privacy (Lectures 11 and 12)

Tools: Sociotechnical systems, Narratives, Power, Co-production, Identity/Positionality, Vulnerability

Main claim. Privacy is not one natural thing. The current legal and technical approach was built in the 1970s U.S. out of fear of government databases and treats privacy as individual control over personal information. In the datafied world that approach fails, because data circulates, merges, and feeds inferences that no sign-up click anticipated.

Key terms

• Privacy Act of 1974, sectoral regulation by industry, self-regulation, and the FTC as enforcer. Regulation of privacy is thin.
• Fair Information Practices (FIPs): notice, choice or consent, access and correction, integrity and security, enforcement and redress.
• PII: information that distinguishes or traces identity, or is linked or linkable to a person. Protecting PII is the standard best practice.
• De-identification is not foolproof. Sweeney showed 87 percent of Americans are identified by zip code, birth date, and gender. The Netflix Prize dataset was re-identified and the second contest was cancelled.
• Differential privacy adds random noise so individuals cannot be re-identified while analysis stays useful. Used in the 2020 Census.
• Terms of service and privacy policies create a legal relationship, service in exchange for information. The work falls on the user, the choice is binary, and companies assume nobody reads.
• Opt-in and opt-out, individual consent.
• Limits of the inherited paradigm: it is about individuals, relies on self-management at one moment, assumes the problem is moving existing information, and equates ethical with legal.

Cases

• Havasupai blood samples. Samples given to ASU for diabetes research were used for other studies and returned in 2010. Identity is collective and relational, consent is necessary but not enough, and sovereignty sits inside colonial history.
• Crisis Text Line, 2022. Shared anonymized conversations with its for-profit spin-off Loris.ai. Data was scrubbed of PII and users had consented, so FIPs thinking says fine. The context shifted from trauma support to customer-service training, consent came in a moment of crisis, and the people who supplied the data got nothing.
• TSA body scanners. A privacy impact assessment under FIPs disabled storage and blurred faces, then generic avatars arrived in 2022. The harm was never PII. It was the body, dignity, male and female as the norm, and who gets pulled aside.

Reading: Malkin, "Contextual Integrity, Explained"

• Privacy is appropriate information flow. A flow is appropriate when it matches the norms of the context it happens in.
• Five parameters of a flow: data type, data subject, sender, recipient, transmission principle. Change one and the flow can become a violation. Actors are roles, and purpose may deserve its own parameter.
• Rejects binaries like public versus private and sensitive versus not. Public store visits become invasive when aggregated into a profile.
• Not procedural. Consent, encryption, and a FIPs checklist do not make a flow appropriate. Norms come from real people's expectations.
• Lessons: Alexa recordings reviewed by contractors broke expectations even though data never left Amazon. Do Not Track became a fingerprinting signal. Like buttons are fine on news sites and wrong on health sites. Judge remote proctoring by whether it serves learning.
• Limits: it predicts feelings rather than describing thought, it favors existing norms like salary secrecy, and inferred data such as the predicted pregnancy must be judged on its own terms.

Reading: Pressly, The Right to Oblivion

• Privacy is valuable because it prevents information from being created, not because it lets us control information that exists.
• The ideology of information assumes every part of life already is or can become data. It naturalizes information, reduces privacy to access control, and marks the boundaries of extraction for states and platforms. By the time you reach privacy settings it is already too late.
• Even defenders like contextual integrity accept the information framing.
• Neoliberal critique via Harcourt: privacy became a form of property with costs and benefits, traded for services.
• Privacy is not secrecy. A secret is known by someone and durable. Privacy is destroyed when breached. The nothing-to-hide slogan misses the point, because hiding is for those with something to hide and privacy is for something else. A lie keeps a secret and destroys privacy, because it creates false information.
• Oblivion is a not-knowing with no fact of the matter, only ambiguity and potential. Respecting privacy means deliberate obliviousness, as if it never happened. Examples: neighbors heard through a wall, the pushy acquaintance who treats you as a repository of information.
• Modern privacy began in the 19th century against snapshot photography and newspapers. The novelty today is acceptance of the ideology of information, not any device.

Week 6. Making Social Order: Surveillance, Numbers, Decisions (Lectures 13 to 15)

Tools: Power, Sociotechnical systems, Positionality, Representation, Performativity, Co-production, Agency

Main claim. Institutions make social order three ways. Surveillance watches people so deviations can be acted on. Numbers make institutions trusted by outsiders. Framing people as decision-makers lets institutions steer them with nudges. Each is a historical product co-produced with politics.

Lecture 13. Surveillance and Security

• Three narrative templates: 1984 means control, the panopticon means discipline, global analytics means security, finding and preempting bad actors.
• History: wartime internal enemies, then 9/11, the PATRIOT Act, warrantless NSA collection, then Snowden in 2013 revealing Verizon metadata on 120 million subscribers and PRISM.
• Normalized surveillance is ubiquitous, suspicionless, bulk, cheap, predictive, and repurposable. The new thing is targeted precision.
• Differential exposure: everyone is watched, some are targeted. State and corporate boundaries are porous: Flock, Palantir, Anduril, ICE vendors.
• Responses: individual agency is not enough, so contestation and regulation such as GDPR data minimization.

Lecture 14. Quantification and Objectivity

• Data is a rhetorical practice, a strategy of communication that persuades.
• Trust in numbers is historical: Galileo's primary properties, political arithmetic, double-entry bookkeeping as a technology of trust, Condorcet's numbers against secrecy, the judge as automaton.
• Varieties of objectivity: truth to nature, disciplinary, mechanical (rules and algorithms), intersubjective.
• Numbers persuade because they are transparent, condensed, commensurable, portable, rule-bound, and look fair.
• Performativity: the metric changes what it measures. Case: U.S. News rankings.
• Objectivity is modern and co-produced with quantification, democracy, and the rule of law. It means public uniform standards instead of personal trust.

Lecture 15. Making Decisions

• Decision-making is a 1940s to 1960s invention built for WWII logistics and Cold War strategy from cybernetics (Wiener), behaviorism (Skinner), game theory, Simon's bounded rationality, and behavioral economics.
• Choice architecture and nudges, libertarian paternalism: you are free to choose but the designer arranges the options. Power is second-order decision-making.
• Case: Facebook 2010 voting messages raised turnout. Zuboff: change behavior at scale. Yeung's hypernudge: decision, data, assessment against goals, reconfigured architecture, nudge again.
• Framing agency as decision-making leaves out relationships, creativity, and obligation.

Reading: Foucault, "Panopticism"

• The leper is excluded, a binary of in and out. The plague town is disciplined: sealed, every street watched, everyone fixed in place and written down. The 1800s merge the two: asylums and schools both brand and individually train.
• The Panopticon: a ring of cells around a tower. The inmate is seen but does not see, "the object of information, never a subject in communication." "Visibility is a trap."
• Power is visible but unverifiable, so you act as if always watched and become "the principle of your own subjection." It is depersonalized, lives in the architecture, is cheap and preventative, works as a laboratory, and generalizes to schools, hospitals, and factories.

Reading: Vasquez, "The N.Y.P.D. Is Teaching America How to Track Everyone Every Day Forever"

• Over 3 billion dollars on cameras, license plate readers, social media scraping, drones, DNA, and iris scans feeding the Domain Awareness System, which keeps data for years and predicts movement.
• No warrant for about 90 percent of the tools. The gang database holds 13,000 people, 99 percent of color. A Marcy Houses teen is added for where he lives or who he posts with. Over 24,000 phones seized in 2024.
• Chilling effect, Planned Parenthood visitors as a future risk, DOGE and Palantir breaking federal data walls. Privacy laws carve out police, so the fix is to ban collection.

Reading: Porter, "Objectivity and Authority"

• Objectivity means impersonality and fairness. Numbers give the appearance of both.
• Who uses numbers? People under pressure from powerful outsiders. Secure elites trust their own judgment. Quantification is a weapon of the weak and a technology of distance.
• Case: the Corps des Ponts et Chaussées, École Polytechnique engineers who defined public utility with numbers, a language credible to the public but not comprehensible to it. Standardizing methods standardized values.
• Technocracy is not quantification. French engineers trusted instinct and used numbers only where they could not control decisions. Postwar America was far more mechanical.

Week 7. Automated Decision-Making: Fairness, Prediction, Risk (Lectures 16 to 18)

Tools: Power, Narratives, Classification, Identity/Positionality, Sociotechnical systems, Agency, Performativity, Representation, Co-production

Main claim. Courts, welfare agencies, insurers, and platforms use algorithms to sort people into scores that decide life chances. The standard story says algorithms can be biased, so make them fair, and fair means accurate. That story misses proxies, who is in the data, feedback loops, due process, stigma, agency, and institutional goals. Seen whole, ADM is a panoptic sort, a form of governance, and its notion of fairness was co-produced with insurance and neoliberalism.

Lecture 16. Fairness. Case: COMPAS

• Risk scores for bail, parole, and sentencing. ProPublica: unfair because error rates differ by race, 44.9 percent of Black defendants versus 23.5 percent of white labeled high risk without reoffending, 28.0 versus 47.7 percent labeled low risk who did. Northpointe: fair because scores are calibrated, a 7 reoffends at about 60 percent for both groups. Both metrics cannot hold when base rates differ, and both assume fair equals accurate.
• Beyond garbage in, garbage out: what does the data represent, what proxy stands for the ground truth (what is recidivism, what is a crime). Health costs as proxy for health needs made Black patients at equal scores sicker. Facial classifiers fail worst on darker-skinned women, which is intersectionality. Who bears the errors depends on positionality, as with Cook County property taxes.
• Anti-discrimination law framing asks whether error rates vary by protected class, yet decision algorithms are built to discriminate. Fairness as accuracy turns a social problem into a technical one and can obscure the systemic issue.

Lecture 17. Prediction. Case: Ofqual

• UK 2020, an algorithm downgraded A-level grades when exams were cancelled. Students protested that the algorithm did not know them. Even a perfectly accurate prediction can be unfair.
• Fairness beyond accuracy: due process and procedural rights, predictive policing against the Fourth Amendment; stigma from the risky label and its circulation and durability; performativity, predictions as self-fulfilling prophecy through feedback loops; agency, predictions foreclose what people can do; institutions, a human in the loop does not fix it, responsibility shifts, and risk assessment becomes a category of good governance.
• Porter's line applies: objectivity is an adaptation to the suspicions of powerful outsiders, the judge as automaton. Rational decision-making is post-1940s and co-produced with managerial order.
• Cybernetic choice architecture: decision, data, assessment against institutional goals, reconfigured architecture, repeat. Google Maps rerouting is a digital speed bump. Personalization at the level of populations.

Lecture 18. Risk. Case: unisex insurance debates

• Timeline of risk: 17th century marine insurance and actuarial methods; industrial revolution collectivizes risk with workmen's compensation and Quetelet's average man; the welfare state, New Deal and Social Security in the 1930s, a shared pool; individualizing risk through private life insurance and actuarial fairness, paying for your own risk, which is panoptic discipline plus biopower (Bouk); the 1970s unisex debates, feminists called sex-based pricing discrimination and insurers called it actuarially fair (Horan); neoliberalism eroded the welfare state and automated public welfare; today granular data gives individualized risk, telematics, credit, and the prediction paradigm.
• Fairness has a history. It was co-produced with justice as fairness in political philosophy, anti-discrimination law, and the spread of insurance risk techniques into other governance. Actuarial fairness truncates more radical claims to social equality. Risk assessment, actuarial fairness, and neoliberalism co-produced each other.

Reading: Eubanks, "The Allegheny Algorithm" (Automating Inequality)

• The AFST is the best case of child-welfare prediction, open and participatory, and it still profiles the poor, confusing parenting while poor with poor parenting.
• Opening: screener Pat Gordon, Stephen scores 5 and Krzysztof 14, because the family's history with public services drove the score, not the allegation.
• Three human choices. Outcome variables are proxies, community re-referral and child placement within two years, so it predicts decisions, not maltreatment. Predictive variables: 131 of 287 kept by stepwise probit regression, about a quarter measure poverty (TANF, SNAP, SSI, Medicaid) and a quarter contact with CYF or juvenile probation. Validation: 76,964 referrals, AUC 76 percent, about a mammogram, about 3,633 wrong predictions in 2016.
• Arguments: data covers only public-service users, so the rich are invisible ("we definitely oversample the poor"). Bias enters at referral, not screening, and referral is the proxy, so the bias is baked in; Black children were 48 percent of foster care and 18 percent of county kids. Nuisance calls contaminate the data (Harriette). Mandated reporters make help into surveillance, so privacy and due process are the price of services. Records are kept forever. Feedback loop: a high score brings scrutiny, which brings re-referrals, which confirms the score. The human in the loop is trained by the algorithm; managers say a mismatch means you are missing something.
• Automated discretion is the discretion of the few, human discretion the discretion of the many, flawed but fixable. "Models are opinions embedded in mathematics." ADM reframes shared social decisions as systems engineering problems.

Reading: Fourcade and Healy, "Classification Situations" (The Ordinal Society)

• Data double: the you rebuilt from traces. Panoptic sort (Gandy): a difference machine that measures and allocates opportunity, always expanding.
• Classification situation: your position in a generated system of categories that decides life chances, often postdemographic and behavioral rather than race or gender.
• Nominal names kinds and hides hierarchy, marked versus unmarked. Ordinal ranks by score. The novelty is fusing naming and ranking with massive data, an ordinal society.
• Classifiers fail because training data comes from an unequal world: misgendered Black women, Amazon's hiring tool, overrepresented policed groups, shortcut learning on which hospital an X-ray came from, opacity that blocks audit and contestation.
• Eigencapital: accumulated data position that pays off. Embodied as ease, skipping the line. Objectified as work to produce good data. Institutionalized as one circulating score, credit scores, social credit, Airbnb trustworthiness.
• Visibility and invisibility are both traps: no credit history is a bad score. Those outside are the lumpenscoretariat. Waiting time tracks power.

Week 8. Futures, AI Narratives, Automation (Lectures 19 to 21)

Tools: Narratives, Performativity, Identity/Positionality, Power, Labor, Co-production

Main claim. Talking about the future is never neutral. Futuring shapes what gets built, funded, feared, and regulated, and it works through reusable narrative templates. AI is a future framed by science fiction and transhumanism. Automation is a social relationship between workers, owners, and machines, and its design is a choice, which is Winner's point from week 1.

Lecture 19. Data Futures

• Narratives are selective, meaningful, descriptive and directive. An 1860 ad said the sewing machine would "banish ragged humanity."
• Data technology does more than predict: it recycles the past, steers action, and preempts futures, as in Minority Report.
• Two purposes of futuring: impose order and control, or open imagination and build common ground.
• Narrative templates are cookie cutters pressed onto reality. 1984 gives Orwellian, Big Brother, thought police. Afrofuturism (Octavia Butler, Black Panther) is an alternative.
• Cases: Songdo smart city as everyday utopia, Diana's onlife world as speculative scenario, Black Mirror as dystopia, The Matrix as invention rather than likeness to life.

Lecture 20. AI Futures

• Templates: AI controls or replaces us (Terminator), we fuse with it (2001, singularity), it becomes a companion (Her). All are stories about what it means to be human.
• Existential risk and longtermism: Yudkowsky, Bostrom, MacAskill. Doomers versus boomers, Yudkowsky versus Andreessen and e/acc.
• Transhumanism is old: Huxley coined it, I. J. Good's 1966 ultraintelligent machine and intelligence explosion, Kurzweil's singularity. Sci-fi becomes blueprint: Neuromancer to Neuralink, the Torment Nexus meme. Musk calls humans a biological bootloader.
• Longevity: Bryan Johnson, cryonics, Amodei. Class reading: Roy Batty in Blade Runner, robots as the working class.
• AI as mirror: ELIZA, Pygmalion, Data on trial, Frankenstein as Prometheus.
• Alternatives: Le Guin's carrier bag, Haraway's cyborg, the unfaithful offspring of militarism and capitalism.

Lecture 21. Automation and Labor

• Two narratives: automation brings abundance (Brynjolfsson and McAfee) or unemployment and inequality (Wiener). The answer is it depends on tasks, wages, labor share, and power.
• Technology as capital (Marx, Noble, Braverman): machines substitute capital for labor and displace control over the work process. Prony's logarithm project divided labor into a hierarchy that Babbage's Difference Engine embodied.
• Industrial revolution pattern: longer hours, deskilling, surveillance, Marx's machinery as the most powerful weapon against strikes. Luddites fought over control and conditions. Taylorism and CNC continue it.
• Today: Brynjolfsson's "Canaries in the Coal Mine" on early-career workers, Autor and Thompson on jobs as bundles of tasks, AI intensifying work like More Work for Mother, Solow's paradox, Baumol's cost disease, Kenyan workers under 2 dollars an hour detoxifying ChatGPT.
• Firms' AI bets assume huge labor productivity gains. Factors of AI value: data, compute, expertise, labor, energy.

Reading: Hildebrandt, Smart Technologies and the End(s) of Law (introduction)

• Diana's onlife day: her PDA Toma tunes her alarm, routes her transit with insurance in mind, and withholds a message. Daughter Lindsay is biometrically matched to risk profiles at school. Grandfather Jacob's care robot Henry balances insurer and hospital interests and does not report his stroke. He dies.
• The story is a scenario study meant to sensitize, not a Luddite fantasy.
• Law versus technological regulation: law is made by a legislator, can be violated, and can be contested in court. Design regulates as a side effect, cannot be violated, and is invisible. Cookies were built for shopping carts and became tracking without a vote.
• Solove: Kafka's The Trial fits better than Big Brother. Pre-emptive computing acts on your inferred future. Remedy: legal protection by design.

Reading: Le Guin, "The Carrier Bag Theory of Fiction"

• Early humans mostly gathered, yet we tell the mammoth hunter story because it has action and a Hero. Following Elizabeth Fisher, the first tool was a bag or sling, not a spear.
• The killer story is spear-shaped and makes conflict the point of narrative. A novel is a bag that holds people in relation, and a Hero in a bag looks like a potato.
• Sci-fi as the myth of technology is Promethean and tragic. Redefine technology as a cultural carrier bag and the stories become less apocalyptic.

Reading: Andreessen, "The Techno-Optimist Manifesto"

• Technology plus markets is the techno-capital machine, an endless engine of growth. Everything good is downstream of growth. Intelligence and energy are the key resources, AI is a universal problem solver, and delaying it costs lives.
• Against UBI, ESG, the precautionary principle, tech ethics, and Nietzsche's Last Man. For nuclear power and 50 billion people.
• Analyze it as a narrative template from a VC: conquest language, assumes markets distribute fairly, dismisses labor and regulation. The opposite story from Noble, Marx, and Le Guin.

Reading: Noble, "Social Choice in Machine Design"

• Two ways to automate machine tools. Record-playback records a machinist and replays, so skill stays on the floor. Numerical control programs parts from blueprints, so skill moves to the office. N/C won although record-playback was as reliable and suited small shops.
• Why: horizontal relations, the Air Force spent over 62 million dollars, created the market, and pushed the complex APT language only large firms could afford. Vertical relations, N/C removes the machinist as the source of intelligence, ends soldiering, lets managers set rates by tape time. GE called it a management system. Ideology: engineers' distrust of human error mirrors capital's distrust of labor. Drucker: automation extends Taylorism.
• Reality: deskilling mostly failed because machines were unreliable, and the intelligence of production stayed with the workforce. Argues against technological determinism.

Week 9. Making Silicon Valley (Lectures 22 to 24)

Tools: Narratives, Identity/Positionality, Power, Performativity, Sociotechnical systems, Labor

Main claim. Silicon Valley is a place, a system for producing wealth and social order, and a story that makes the system look natural and inevitable. The standard story is about materials, technologies, and entrepreneurs. Underneath are networks, the military, and capital, and the winners of each wave ride the next: chips, PCs, internet, platforms, AI.

Lecture 22. Making Silicon Valley 1

• Before: the Valley of Heart's Delight, orchards on unceded Ohlone land.
• Four layers of the standard story: electronics (San Jose radio, vacuum tubes, ship-to-shore wireless for the Navy), Stanford (Terman, HP, WWII radar tubes, SLAC, land deals with a liberal-arts sheen), silicon (Bell Labs transistor, Shockley moves west, the traitorous eight found Fairchild, the integrated circuit, Intel), computing (mainframes, PCs, internet, smartphones).
• Eisenhower's 1961 farewell warned of the military-industrial complex.
• Keys: regional network effect, free movement of labor, the military-industrial complex, a chart up and to the right, venture capital. Who disappears: the Ohlone, the Navy, Stanford's land deals.

Lecture 23. Making Silicon Valley 2

• Californian Ideology: counterculture fused with free-market libertarianism through faith in technology.
• The garage myth hides Stanford, networks, and capital. Disruptive innovation and the lean startup became an operating manual. Move fast and break things slides into fake it till you make it: SBF, Elizabeth Holmes.
• The dot-com bust did not end the story. Survivors rode smartphones and social media.
• Platforms: multi-sided markets, enclosed spaces for data collection, network effects toward monopoly. The data imperative: everything can and should be rendered into data and accumulated.

Lecture 24. Making Silicon Valley 3

• Platforms run a cybernetic nudge loop: decision becomes data, choice architecture is adjusted, next decision is steered.
• Gig economy and microlabor feed AI, from Mechanical Turk to annotation shops. Cloud computing and data centers are the hidden material layer.
• Scaling laws made scale the strategy and capital the moat, so Nvidia and compute are strategic resources. Factors: data, compute, expertise, labor, energy. Strategies: frontier models, hyperscalers, conglomerates, open source.
• Defense tech is back: Palantir, Anduril, OpenAI Pentagon work, Starlink in Ukraine as a private sovereign.
• Politics: Andreessen's firm backed Trump, CEOs at the inauguration, Thiel saying freedom and democracy are incompatible.

Reading: Nicholas, "Silicon Valley and the Emergence of Investment Styles"

• Three things built the Valley: universities, military spending, and a distinctive culture and legal climate. The cluster created demand for venture capital and shaped how VCs invest.
• Three investment styles: Arthur Rock bet on people, Tom Perkins on technology, Don Valentine on market size.
• Stanford under Terman owned patents, built the Stanford Industrial Park, and let professors start companies. At MIT that could get you fired.
• The military as early adopter bought over half of Bay Area integrated circuits until 1967 and paid for the learning curve, so chip prices fell from 31.60 to 2.33 dollars in five years. Cost-plus contracts built the manufacturing base. When defense spending fell, the Valley switched product lines and Boston's Route 128 did not.
• Culture rewarded risk, accepted failure, and drew immigrants. California Section 16600 voids non-competes, which is why Fairchild could exist.

Reading: Vinsel and Russell, "Hail the Maintainers"

• Innovation is an overvalued ideology. It replaced "progress" after the late 1960s as a smaller word with no moral content. 1980s innovation policy made Silicon Valley the model every region copied; Schumpeter's creative destruction and Christensen's disruption became mantras.
• Innovative regions produce inequality, see the Google bus protests.
• Three corrections: technology is bigger than innovation since most objects are old; infrastructure failure is a political symptom; innovators are about 1 percent of workers and most work is maintenance.
• Cowan's More Work for Mother: appliances raised standards and made more housework. Innovation is value-agnostic, crack cocaine was innovative. Ask who benefits and to what end.

Reading: Fang, "Silicon Valley Goes to War"

• The connect-the-world ethos hid Cold War roots. Now Y Combinator funds missile startups, defense VC grew tenfold in five years, Palantir's Karp says engineers were wasted on delivery apps.
• Three acts: Stanford as the Cold War university and Lockheed as Santa Clara's largest employer; counterculture repurposing military tools, peaking with Project Maven in 2018 when 4,000 Google employees killed a drone contract; the return with Palantir, Project Nimbus, and a saturated consumer market.
• Startups want to undo the 1993 Last Supper that consolidated five contractors. The China arms race story justifies data center spending. OpenAI, Anthropic, and Google dropped military limits.
• Democracy paradox: defending democracy with deportation surveillance and condemned wars. Illiberal means undermine the democratic end.

Reading: Crawford, "Labor"

• Workplace AI returns to old industrial logics: surveillance, time discipline, deskilling, now tracked to the micro-movement. The question is how humans are treated like robots.
• Amazon: 15 minutes off-task per shift, robots behind fences, braces and painkiller vending machines. Minnesota workers walked out and Amazon would discuss anything except the rate, "our business model."
• Lineage: Smith's division of labor, Marx's appendage of the machine, E. P. Thompson's time discipline. The panopticon began as a workplace, Samuel Bentham's factory before Jeremy's prison. Babbage saw the factory as a computer with labor as a source of error; Chicago's disassembly line realized it and The Jungle produced food law and nothing for workers. McDonald's algorithmic scheduling by the second.
• Ghost work, Potemkin AI (people paid to pretend to be AI), fauxtomation (self-checkout moves labor to the customer). Mechanical Turk hid a chess master in 1770; Bezos calls it artificial artificial intelligence. AI is "neither artificial nor intelligent." Google's TrueTime privatizes time. "We are all tech workers."

Reading: Sloane, "The Business of AI"

• AI sells prediction as an asset and its value is speculative: OpenAI valued at 30 times revenue while losing money. Prediction is hard to price, so AI is sold as subscriptions to infrastructure. Financialization.
• Three ideas it rests on. Data is infinite: false, data is the new oil only if someone works the rigs, participating in society means producing data, refining falls on Majority World workers, and AI is physically extractive. Status decides what gets automated: London cabbies' the Knowledge devalued by Uber overnight, strawberry farms rebuilt indoors for robots, coding loses status once AI codes. Prediction can replace judgment: recruiters lost Boolean search to black-box semantic search and game it, adding work and bias; a dermatologist uses the score to learn and persuade and ignores it when sure. Feminized recruiting is treated as automatable, high-status medicine gets careful design. Discretion is where humans interpret after AI predicts.

Connect the dots

HCE tool to best evidence

Tool	Define it as	Reach for
Power	asymmetric capacity to structure others' behavior and decisions	panopticon, NYPD DAS, choice architecture, the rate at Amazon, Air Force funding N/C
Sociotechnical systems	human and technical action intertwined, no separable technical part	AFST (callers, screeners, model, managers), COMPAS in courts, Diana's onlife world, platforms
Classification	sorting beings and knowledge into categories that shape and are shaped by social order	classification situations, nominal vs ordinal, gang database, risk tiers, PII as a category
Representation	one thing standing for another, doing political work	proxies and ground truth, data double, re-referral standing for maltreatment, the standard story of the Valley
Performativity	descriptions and predictions make the thing they describe	AFST feedback loop, U.S. News rankings, predictive policing hot spots, scaling-law bets, prediction as asset
Narratives	stories about who we are and what is possible	1984 and panopticon templates, Terminator, garage myth, arms race story, techno-optimism, carrier bag
Agency	capacity to act in sociotechnical relations	nudges and hypernudges, preemption, Ofqual students, recruiters gaming search, Luddites
Identity/Positionality	socially conditioned selfhood that shapes how you are treated	differential exposure, Havasupai, who bears COMPAS errors, TSA scanners, eigencapital
Vulnerability	uneven exposure to harm	Crisis Text Line, poverty profiling, gig and ghost workers, Jacob's care robot
Labor	socially organized use of bodies and time	Noble, Crawford, Sloane, maintainers, Kenyan annotators, deskilling
Co-production	technology and society make each other	privacy law and databases, objectivity and democracy, fairness and insurance and neoliberalism, VC and the Valley

Author to author

• Foucault to Vasquez to Fourcade and Healy to Crawford. Discipline as a diagram, its digital version in NYC, its sorting version in scores, and its origin as a workplace.
• Porter to Lecture 16 to Eubanks. Numbers as impersonal fairness under pressure; fairness as accuracy; a county that chose a score to deflect accusations of bias and got poverty profiling.
• Malkin to Pressly to Eubanks. Flows judged by context; the information already created is the problem; mandated reporting as a flow that turns help into surveillance.
• Lecture 18 to Winner to Sloane. Actuarial fairness as inherently political, the artifact that requires a certain social order; prediction as an asset is the business version.
• Hildebrandt to Lecture 15 to Lecture 24. Design regulates without law; choice architecture; the platform nudge loop.
• Le Guin to Andreessen to Lecture 20. Hero story versus bag; techno-capital conquest; doomers and boomers both technologically determinist.
• Noble to Nicholas to Fang. The Air Force chose N/C; the military bought half the chips; the Valley goes back to war.
• Vinsel and Russell to Crawford to Lecture 21. Maintenance is most work; ghost work and the rate; automation as capital.

Week 1 to 4 material worth carrying in

• Winner, "Do Artifacts Have Politics": artifacts have politics either because design settles a social question (Moses's overpasses) or because the technology is inherently political, requiring or strongly favoring a form of social order (the atom bomb needs hierarchy, solar favors decentralization). Against technological determinism and against naive social determinism.
• Bowker and Star: classifications are invisible until they break; apartheid race categories. Hacking: classifications make up people, looping effects.
• Scott: legibility, the state simplifies forests and people into an administrative grid for control from a distance.
• Kuhn and Collins and Pinch: knowledge is made by communities, and experts failed the Cumbrian shepherds by ignoring local knowledge.
• Lecture 8 to 9: biopower, Quetelet's average man, statistics co-produced with eugenics, Terman and Stanford.

Predicted short answers (2 x 10 points, 30 minutes each)

Midterm 1 short answers had one shape: take an author's argument, explain it with its distinctions, then apply it to a case the author did not discuss. The study guide says the same for Midterm 2 and adds "explain in greater depth." Below are the five patterns, the variants most likely on the alternate exam, and the bullets to reuse. Learn the bullets, not the questions.

Pattern A. "How would author X analyze phenomenon Y?" (the Winner-on-neoliberalism type)

Structure: paragraph 1, X's argument with its key distinction. Paragraph 2, define Y as the course defines it, with its history. Paragraph 3, run X's distinction over Y and conclude.

A1. Winner on neoliberalism, actuarial fairness, or risk assessment.

Winner: artifacts have politics either by design that settles a social question, or because a technology is inherently political, requiring or favoring a social order. Neoliberalism (L18): erosion of the welfare state, risk moved from shared pools to individuals, automated welfare, the prediction paradigm. Application: actuarial fairness is not a neutral tool that neoliberal politics later picked up. Individualized risk assessment requires granular surveillance, ranked populations, and a market in which each person carries their own risk, so it is inherently political in Winner's strong sense. The unisex insurance debate shows the design-choice version too: pricing by sex settled a question about equality. Close with Winner's warning against technological determinism: neoliberalism was a choice, not what computers made inevitable.

A2. Foucault on the AFST, COMPAS, or the NYPD.

Start with the panopticon's features, then show which hold (seen but not seeing, unverifiable, self-policing through mandated reporting, depersonalized power in the score) and which go further (the sort allocates life chances, data is durable and circulates).

A3. Porter on COMPAS, Ofqual, or the AFST.

Weak elites under suspicion adopt mechanical objectivity. Allegheny adopted a score after being accused of bias; Ofqual needed grades nobody could call personal; the judge becomes an automaton. Cost: the standardized method standardizes values and strips local judgment, the same failure as Cumbria.

A4. Noble on generative AI in classrooms, coding, or recruiting (Sloane).

Horizontal and vertical relations; who funds, who gains control; desire versus satisfaction; the intelligence of production stays with the workforce.

A5. Le Guin on Andreessen or on transhumanism.

Killer story versus carrier bag; conquest language; what the bag would hold instead, maintainers and gatherers.

A6. Pressly on Crisis Text Line, the Havasupai, or DTC DNA testing.

Ideology of information; privacy as prevention of information, not control; privacy versus secrecy; oblivion as the right not to be made into data.

A7. Malkin on Crisis Text Line, Alexa reviews, or remote proctoring.

Name the five parameters, find the one that changed, then say why consent and anonymization do not repair the flow.

Pattern B. "Define concept C and apply it to case K from another unit"

B1. Contextual integrity applied to the AFST or NYPD gang database.

Flows from schools and clinics into a risk score; mandated reporting as the transmission principle that breaks the context of care.

B2. Panoptic sort or classification situation applied to the gig economy, credit, or admissions.

Nominal and ordinal, eigencapital's three forms, lumpenscoretariat.

B3. Performativity applied to U.S. News, predictive policing, the AFST, or scaling-law valuations.

Define performativity, then show how the measure or prediction helps produce the reality it claims to describe in each case.

B4. Differential exposure applied to TSA scanners, COMPAS errors, or Amazon warehouses.

Define differential exposure, then show which groups bear the burden of the system's errors and intrusions in each case.

B5. Technology as capital applied to ghost work, self-checkout, or AI tutors.

Substitution of capital for labor, displacement of control, division of labor embodied in the machine.

B6. Narrative template applied to the arms race story, the garage myth, or AI doom.

Define the narrative template, then show how each story casts heroes, enemies, and inevitability, and who it leaves out.

B7. Legibility or the data imperative applied to platforms or to Diana's onlife world.

Define legibility or the data imperative, then show how platforms and Diana's onlife world make people readable as data.

Pattern C. "Compare two readings"

C1. Malkin versus Pressly.

Both reject FIPs and consent. Malkin keeps the information frame and judges flows; Pressly says the frame itself is the problem. Test case: Crisis Text Line passes Malkin's parameters only if norms are read narrowly, and fails Pressly because the conversations should never have become a dataset.

C2. Eubanks versus Fourcade and Healy.

One case in depth versus a theory of the ordinal society. Both: data covers the surveilled, scores allocate life chances, the invisible are penalized. Difference: Eubanks' remedy is human discretion of the many; Fourcade and Healy show discretion itself is now scored.

C3. Andreessen versus Noble, or Andreessen versus Crawford.

Growth raises wages versus control and deskilling; who speaks, a VC or a historian of labor.

C4. Nicholas versus Fang, or Nicholas versus Vinsel and Russell.

Same military facts, different valuation; innovation as an ideology versus investment style.

C5. Foucault versus Hildebrandt.

Discipline through visibility versus regulation through design you cannot see or violate; Kafka versus Big Brother.

Pattern D. "Give the history of concept C and say why it matters"

D1. Privacy.

19th century photography and newspapers, 1970s databases, Privacy Act, FIPs, PII, terms of service, contextual integrity, oblivion. Point: the current form fits 1970s concerns, not inference and circulation.

D2. Fairness and risk.

Marine insurance, collectivized industrial risk, Social Security, actuarial fairness, 1970s unisex debates, neoliberal welfare automation, prediction paradigm. Point: fairness as accuracy truncates equality claims.

D3. Objectivity and trust in numbers.

Bookkeeping, Condorcet, mechanical objectivity, the Corps des Ponts, rule of law, U.S. News. Point: a political strategy for trust, not a property of numbers.

D4. Decision-making.

WWII logistics, cybernetics, behaviorism, game theory, bounded rationality, nudges, hypernudges. Point: agency framed as choice is governable.

D5. Silicon Valley.

Ohlone orchards, Navy radio, Terman's Stanford, military chip buying, Fairchild and Section 16600, VC styles, dot-com, platforms, scaling laws, defense return. Point: the garage myth hides the state and capital.

Pattern E. "Explain the argument of reading R" (the essay-reading type)

Likely targets, because they carry the week's argument. These are the official reading questions, so answer each one in your head before the exam.

E1. Nicholas.

Explain the three builders.

E2. Noble.

Explain the two reasons N/C won, and for whom.

E3. Eubanks.

Explain how the AFST is trained and used, and the criticisms.

E4. Fourcade and Healy.

Explain the eigencapital types, and classification situation versus demographics.

E5. Pressly.

Explain oblivion versus privacy.

E6. Malkin.

Explain what is contextual, the difference from FIPPs, and why it is hard to legislate.

E7. Foucault.

Explain how the panopticon works and why it models data surveillance.

E8. Porter.

Explain why we trust impersonal numbers and why weak elites use rules.

E9. Le Guin.

Explain the hero spear versus the carrier bag.

E10. Hildebrandt.

Explain Diana as futuring.

Predicted analytical essay (20 points, 60 minutes)

The official sample is a budget-driven proposal to replace or augment instructors with generative AI. The exam scenario will be shorter and will again be a policy choice about deploying a data or AI system on people, with a budget or efficiency pressure, two or more options, and a request for a recommendation. Everything in weeks 5 to 9 is built to analyze exactly that.

One template that fits any scenario

Use two or three lenses in depth rather than all four.

T1. Name the tension in one sentence.

Usually efficiency and scale against something the course values: care, discretion, due process, labor, context, or democratic control.

T2. Lens 1: who is made legible and what flows.

Contextual integrity parameters, Pressly's point that the data should perhaps not exist, differential exposure, panoptic discipline. Pick the people the scenario makes into data and say what the system now knows and preempts.

T3. Lens 2: what the prediction does.

Proxies and ground truth, feedback loops, performativity, stigma and durability, agency foreclosed, human in the loop trained by the score, fairness as accuracy obscuring the social question, actuarial individualization of risk.

T4. Lens 3: who does the work and who owns the machine.

Technology as capital, deskilling versus skill kept on the floor (Noble's two paths), ghost work and fauxtomation, status deciding what gets automated (Sloane), maintenance versus innovation, Baumol and the service sector.

T5. Lens 4, if it fits: the story being told.

Narrative template behind the proposal (abundance, arms race, inevitability, 2 Sigma as destiny), who tells it and who it leaves out, technological determinism, Winner's choice.

T6. Recommendation.

Choose, then condition it: keep skill with the practitioner, make the flow match the context, make scores contestable with due process, pay and count the hidden labor, and name what the choice costs.

Modules you can drop into any essay

M1. Design is regulation (Hildebrandt).

The proposal regulates students or workers through design nobody voted on and nobody can violate; ask for legal protection by design and contestability.

M2. Consent is not enough (L12, Malkin, Pressly).

Crisis Text Line consented, Havasupai consented. Judge the flow, not the click.

M3. Best case still profiles (Eubanks).

The AFST was open and well-meaning. Expect proxies for the outcome, oversampling of whoever is already in the system, and a loop that confirms itself.

M4. Two ways to automate (Noble).

Record-playback keeps skill with the worker, N/C moves it to the office. Proposal B is record-playback, Proposal A is N/C, and N/C won for control, not quality.

M5. Status decides automation (Sloane).

GSIs and adjuncts are low status so they are automated first, like recruiters and cabbies; tenured faculty keep careful design, like dermatologists.

M6. Ghost work (Crawford, L21).

The AI tutor's training data, moderation, and prompt maintenance are labor that the budget line hides; the rate becomes the business model.

M7. Panoptic sort (Fourcade and Healy).

The learning portfolio is a data double and an eigencapital account that follows the student.

M8. Weak elites reach for numbers (Porter).

A university under a deficit adopts auditable scores to show impartiality, trading judgment for a technology of distance.

M9. Narrative check (L19, L20, Le Guin).

2 Sigma plus LLM as a hero story of abundance; ask whose future, and what the bag would hold.

M10. Winner's test.

Does the system require a social order, centralized data, ranked students, deskilled staff, or merely favor one?

Predicted scenarios, with the lenses that score

For each, the template holds. Only the modules change.

S1. Campus or city budget crisis, AI replaces frontline workers (tutors, advisers, 311 operators, nurses' triage).

Lenses 3, 2, 1. Recommend augmentation that keeps skill on the floor plus data minimization and contestable scores.

S2. A county or school adopts a risk score (child welfare, student dropout, campus mental-health flags, housing priority).

Lenses 2, 1, Porter. AFST and Ofqual are the direct precedents; demand due process, expungement, and discretion of the many.

S3. A city buys surveillance tech for safety (cameras, plate readers, Flock, Palantir, drones at a protest or campus).

Lenses 1 and 2 plus Foucault and Vasquez. Differential exposure, chilling effect, repurposing, state-corporate partnership; recommend bans on collection rather than storage rules.

S4. A health or insurance program personalizes premiums or care using wearables or telematics.

Lens 2 via L18, actuarial fairness against collectivized risk, Pressly on the data that should not exist, Winner.

S5. The state partners with defense or frontier AI firms, or a university takes military money.

Fang, Nicholas, Noble, Eisenhower, Project Maven; the arms race narrative; the democracy paradox.

S6. A platform or gig scheduling system is proposed for public workers.

Crawford's time discipline, algorithmic scheduling, nudge loop, we are all tech workers.

S7. A dataset is shared or sold for a good cause (crisis texts, student records, tribal or genetic data).

Lens 1 in depth: parameters, oblivion, relational and collective identity, sovereignty.

Memorize page

Authors and titles

• Malkin, "Contextual Integrity, Explained." Pressly, The Right to Oblivion.
• Foucault, "Panopticism." Vasquez, "The N.Y.P.D. Is Teaching America How to Track Everyone Every Day Forever." Porter, "Objectivity and Authority" (Trust in Numbers).
• Eubanks, "The Allegheny Algorithm" (Automating Inequality). Fourcade and Healy, "Classification Situations" (The Ordinal Society).
• Hildebrandt, Smart Technologies and the End(s) of Law. Le Guin, "The Carrier Bag Theory of Fiction." Andreessen, "The Techno-Optimist Manifesto." Noble, "Social Choice in Machine Design."
• Nicholas, "Silicon Valley and the Emergence of Investment Styles." Vinsel and Russell, "Hail the Maintainers." Fang, "Silicon Valley Goes to War." Crawford, "Labor" (Atlas of AI). Sloane, "The Business of AI."
• From before: Winner, "Do Artifacts Have Politics?"; Scott on legibility; Bowker and Star; Hacking; Kuhn; Collins and Pinch.

Numbers

• 87 percent re-identified by zip, birth date, gender. Privacy Act 1974. Crisis Text Line 2022. Snowden 2013, 120 million Verizon subscribers.
• NYPD: 3 billion dollars, 13,000 in the gang database, 99 percent people of color, 24,000 phones in 2024.
• COMPAS: 44.9 versus 23.5 percent false high risk, 28.0 versus 47.7 percent false low risk, calibrated around 60 percent.
• AFST: 131 of 287 variables, AUC 76 percent, 76,964 referrals, about 3,633 errors in 2016, Black children 48 percent of foster care and 18 percent of county kids, scores 5 and 14.
• Noble: Air Force over 62 million dollars. Nicholas: military over half of ICs until 1967, 31.60 to 2.33 dollars in five years, Section 16600. Fang: Project Maven 2018, 4,000 employees, 1993 Last Supper, defense VC up tenfold. Sloane: OpenAI at 30 times revenue. Kenyan workers under 2 dollars an hour. Amazon 15 minutes off task. Vinsel: innovators about 1 percent of workers. Eisenhower 1961.

Lines to quote

• "Visibility is a trap." The inmate is "the object of information, never a subject in communication." Power "visible and unverifiable."
• Quantification is a "technology of distance" and a "weapon of the weak." Objectivity is "an adaptation to the suspicions of powerful outsiders."
• "Parenting while poor" versus "poor parenting." "We definitely oversample the poor." "Models are opinions embedded in mathematics." Privacy and due process are the price of services.
• Eigencapital: embodied, objectified, institutionalized. Lumpenscoretariat.
• Pressly: privacy prevents information from being created; privacy is not secrecy. Malkin: appropriate flows, five parameters.
• Hildebrandt: law can be violated and contested, design cannot. Kafka not Big Brother.
• Le Guin: the Hero in a bag looks like a potato. Andreessen: everything good is downstream of growth.
• Noble: record-playback versus N/C, skill on the floor versus skill in the office, horizontal and vertical relations.
• Nicholas: universities, military spending, culture and legal climate. Rock people, Perkins technology, Valentine market.
• Crawford: AI is "neither artificial nor intelligent." Potemkin AI, fauxtomation, the rate is "our business model." Sloane: prediction as an asset, data is not infinite, status decides automation.
• Vinsel and Russell: innovation is an ideology, most work is maintenance.
• L18: actuarial fairness truncates more radical claims to social equality.
• L22: the standard story naturalizes the Valley so winners ride the next wave.