When does AI strengthen democracy — and when does it undermine it?
The same technologies are deployed in Beijing, Berlin, and Brasília — with very different consequences for citizens.
The answer, this lecture argues, lies in institutions.
“Institutions” - mechanisms that determine whether state power can be challenged: courts, elections, free media, civil society, etc
Pro-privacy protesters rally in Hong Kong in support of Edward Snowden, 2013. Source: VOA / Wikimedia Commons (public domain).
Learning Objectives
By the end of this lecture, you should be able to:
Explain how AI lowers the cost of monitoring and what this means for state capacity
Apply the privacy-as-power framework to analyze citizen–institution asymmetries
Recognize the distributional consequences of AI surveillance for marginalized groups
Predict variation in AI’s democratic effects across regime types and identify the institutions that determine the outcome
The Master Causal Framework
%%{init: {'theme':'base','themeVariables':{'fontSize':'20px','primaryColor':'#e8eeeb','primaryTextColor':'#1e293b','primaryBorderColor':'#4a7c6f','lineColor':'#334155','secondaryColor':'#f5f0e8','tertiaryColor':'#fdf2e9'},'flowchart':{'useMaxWidth':true,'width':1150,'height':650,'nodeSpacing':40,'rankSpacing':50,'htmlLabels':true,'curve':'basis','diagramPadding':8}}}%%
flowchart LR
A["AI<br/>Capability"] --> B["Data<br/>Extraction"]
B --> C["Surveillance &<br/>Decision Systems"]
C --> D["Power Shift:<br/>State vs. Citizens"]
D --> I{"Institutions<br/>(strong or weak?)"}
I -->|strong checks| R["Re-equilibration:<br/>reform &<br/>accountability"]
I -->|weak checks| X["De-democratization:<br/>repression &<br/>exclusion"]
Author’s illustration. Institutions are the moderator: the same AI capability produces opposite outcomes depending on the strength of judicial, media, civic, and regulatory checks.
Four Anchor Scholars
James Scott — Seeing Like a State (1998)
States simplify society to make it “legible” for control
Charles Tilly — Coercion, Capital, and European States (1990)
State capacity = extraction + monitoring + coercion
Julie Cohen — Between Truth and Power (2019)
Privacy is a power resource, not mere secrecy
Part 2: The Mechanism — AI Lowers the Cost of State Surveillance
AI as Task Automation
Prediction: Given inputs, forecast likely outcomes
Credit default, recidivism, disease outbreaks
Classification: Given inputs, assign categories
Face to identity; text to sentiment; behavior to threat
Generation: Given prompts, produce content
Text, images, synthetic media for disinformation
AI Reduces State Monitoring Costs
Figure 1
Key Technical Capabilities for State Control
Each AI capability serves a distinct state goal — making populations legible, predictable, or governable.
Facial Recognition — Match faces across databases; identify in crowds → Goal: Identify who is where, in real time
Natural Language Processing — Analyze text at scale; detect sentiment → Goal: Detect dissent and shape public discourse
CCTV cameras: the visible infrastructure of state monitoring. AI transforms passive recording into active, real-time identification and tracking. Source: Wikimedia Commons.
Key Technical Capabilities for State Control
Each AI capability serves a distinct state goal — making populations legible, predictable, or governable.
Predictive Analytics — Risk-score individuals; forecast protests → Goal: Pre-empt threats before they materialize
Behavioral Biometrics — Identify by gait, typing, voice patterns → Goal: Track individuals even when faces are hidden
CCTV cameras: the visible infrastructure of state monitoring. AI transforms passive recording into active, real-time identification and tracking. Source: Wikimedia Commons.
Tilly’s State Capacity Framework
Charles Tilly identified four core state-making activities:
War making — eliminating rivals outside the territory
State making — eliminating rivals inside the territory
Protection — eliminating rivals of the state’s clients
Extraction — acquiring the resources to do the other three
AI enhances all four, but especially extraction (data as a new resource) and state making (internal control via legibility — Scott’s bridge to Tilly).
How States Use AI to Control Populations
Instrument
State Incentive
Rights Risk
Failure Mode
Mass Surveillance
Threat detection
Privacy, assembly
Function creep
Predictive Policing
Efficiency
Due process
Feedback loops
Welfare Fraud Detection
Cost reduction
Dignity
False positives
Social Credit Systems
Compliance
Autonomy
Arbitrariness
Border Control
Security
Asylum rights
Exclusion errors
Censorship
Narrative control
Expression
Overblocking
Author’s illustration based on comparative policy analysis.
Scott’s Legibility Concept
Core claim: States simplify complex social reality to govern
Standardized names, maps, censuses, registries
Legibility enables intervention:
What can be seen can be measured and targeted
AI vastly expands legibility:
Movement, social networks, emotions now trackable
AI is “Seeing Like a State 2.0”
Jeremy Bentham’s Panopticon (1791): a prison designed so inmates could be watched at any time without knowing when. AI-enabled surveillance generalizes this principle to entire populations. Source: Wikimedia Commons (public domain).
Surveillance Produces Chilling Effects
%%{init: {'theme':'base','themeVariables':{'fontSize':'22px','primaryColor':'#e8eeeb','primaryTextColor':'#1e293b','primaryBorderColor':'#4a7c6f','lineColor':'#334155','secondaryColor':'#f5f0e8','tertiaryColor':'#fdf2e9'},'flowchart':{'useMaxWidth':true,'width':1150,'height':650,'nodeSpacing':50,'rankSpacing':60,'htmlLabels':true,'curve':'basis','diagramPadding':8}}}%%
flowchart LR
A["AI-Enabled<br/>Surveillance"] --> B["Belief One<br/>Is Watched"]
B --> C["Self-<br/>Censorship"]
C --> D["Reduced Dissent<br/>& Participation"]
D --> E["Democratic<br/>Erosion"]
Author’s illustration. Even without punishment, belief in surveillance changes behavior — especially dissent (Penney, 2016).
Case: Predictive Policing
How it works:
Historical crime + demographic data produce risk scores
Police resources allocated to “high-risk” areas
The feedback loop:
More policing leads to more detected crime leads to higher scores
Case: Predictive Policing
What is actually being predicted?
The model is trained on arrest data, not crime data
Arrests reflect where police already go — not where crime actually occurs
Predictive policing predicts where policing has happened, not where crime happens (Lum & Isaac, 2016)
Distributional consequence:
Over-policed communities remain over-policed
The model can look “accurate” while reproducing historical discrimination
The Predictive Policing Feedback Loop
%%{init: {'theme':'base','themeVariables':{'fontSize':'22px','primaryColor':'#e8eeeb','primaryTextColor':'#1e293b','primaryBorderColor':'#4a7c6f','lineColor':'#334155','secondaryColor':'#f5f0e8','tertiaryColor':'#fdf2e9'},'flowchart':{'useMaxWidth':true,'width':1150,'height':650,'nodeSpacing':50,'rankSpacing':60,'htmlLabels':true,'curve':'basis','diagramPadding':8}}}%%
flowchart LR
A["Historical<br/>Crime Data"] --> B["ML Model<br/>Trains"]
B --> C["Predicts Risk<br/>in Same Areas"]
C --> D["Police Deployed<br/>to Those Areas"]
D --> E["More Arrests<br/>Recorded There"]
E --> A
Author’s illustration. The prediction changes the outcome it predicts (“performative prediction”).
Selective Enforcement and Legitimacy
The selective enforcement logic:
AI provides comprehensive information on violations
States cannot enforce all laws; discretion is political
Consequences:
Laws become weapons against opponents
Compliance with law does not guarantee safety
Acemoglu & Robinson: Extractive institutions use law instrumentally; inclusive institutions constrain discretion
Part 3: The Stakes — Privacy, Power, and Distributional Harm
Privacy as a Power Resource
Privacy is not secrecy (“nothing to hide, nothing to fear”). For Julie Cohen, it is a power resource that protects autonomy by limiting the information asymmetry between individuals and institutions.
%%{init: {'theme':'base','themeVariables':{'fontSize':'22px','primaryColor':'#e8eeeb','primaryTextColor':'#1e293b','primaryBorderColor':'#4a7c6f','lineColor':'#334155','secondaryColor':'#f5f0e8','tertiaryColor':'#fdf2e9'},'flowchart':{'useMaxWidth':true,'width':1150,'height':650,'nodeSpacing':50,'rankSpacing':60,'htmlLabels':true,'curve':'basis','diagramPadding':8}}}%%
flowchart LR
A["Institution Knows<br/>Much About You"] --> C["Information<br/>Asymmetry"]
B["You Know Little<br/>About Institution"] --> C
C --> D["Power<br/>Imbalance"]
D --> E["Privacy Protection =<br/>Reduce Asymmetry"]
Author’s illustration based on Cohen (2019). The individual is transparent; the institution is opaque.
Rights Tradeoffs in AI Governance
Security vs. Liberty
Surveillance may reduce crime — but restricts freedom
Automation vs. Accountability
Automated decisions are fast and consistent — but opaque and hard to contest
Rights Tradeoffs in AI Governance
False Positives vs. False Negatives
Every threshold trades one error for the other — who bears each kind?
The question is not whether tradeoffs exist, but who decides.
Who Bears the Harms?
Pattern: AI harms fall disproportionately on marginalized groups (O’Neil, 2016; Eubanks, 2018; Benjamin, 2019)
Welfare systems: poorest face most intrusive scrutiny (Eubanks, 2018)
Why this pattern?
Training data reflects historical discrimination (Barocas & Selbst, 2016)
Affected groups have less power to contest errors (Eubanks, 2018)
Harms to marginalized groups are politically cheaper (Benjamin, 2019)
Facial recognition scanning at an airport gate, 2018. The same technology produces disparate accuracy across demographic groups. Source: Delta News Hub / Wikimedia Commons (CC BY 2.0).
Facial Recognition Accuracy Disparities
Figure 2
Case Evidence: Wrongful Arrest by Facial Recognition
1. How much more likely was facial recognition to misidentify Black people than white people — and what causes this disparity? 2. The Police said FR matches are used “as a clue, not as proof.” Why does this safeguard still fail in Robert Williams’s case? 3. The startup proposes “synthetic digital humans” to fix bias. Are there limits of that fix?
Part 4: The Answer — Institutions Decide
Acemoglu & Robinson: Institutions Matter
Core distinction:
Inclusive institutions distribute power and constrain elites
Extractive institutions concentrate power for elite benefit
Applied to AI surveillance:
Inclusive institutions: AI constrained by law and oversight
Extractive institutions: AI amplifies elite control
East Germans storm the Stasi headquarters, Berlin, January 1990. The Stasi maintained files on 5.6 million people — a surveillance state dismantled by institutional change. Source: Bundesarchiv / Wikimedia Commons.
Acemoglu & Robinson: Institutions Matter
Key insight: Same technology, different outcomes by institutional context
East Germans storm the Stasi headquarters, Berlin, January 1990. The Stasi maintained files on 5.6 million people — a surveillance state dismantled by institutional change. Source: Bundesarchiv / Wikimedia Commons.
Regime Types and AI Outcomes
%%{init: {'theme':'base','themeVariables':{'fontSize':'20px','primaryColor':'#e8eeeb','primaryTextColor':'#1e293b','primaryBorderColor':'#4a7c6f','lineColor':'#334155','secondaryColor':'#f5f0e8','tertiaryColor':'#fdf2e9'},'flowchart':{'useMaxWidth':true,'width':1150,'height':650,'nodeSpacing':35,'rankSpacing':45,'htmlLabels':true,'curve':'basis','diagramPadding':8}}}%%
flowchart LR
A["Same AI<br/>Technology"] --> B["Democracy<br/>(Inclusive)"]
A --> C["Hybrid<br/>Regime"]
A --> D["Autocracy<br/>(Extractive)"]
B --> E["Constrained Use<br/>Courts, Press, Elections"]
C --> F["Volatile Outcomes<br/>Leader-Dependent"]
D --> G["Systematic<br/>Repression"]
Author’s illustration based on Acemoglu & Robinson (2012).
What Constrains AI State Power?
Whether AI produces reform or repression depends on which institutions are strong enough to check it.
Safeguard
What It Prevents in AI Surveillance
What It Cannot Prevent
Judicial independence
Warrantless monitoring; AI evidence used without due-process review
Vendors invoke trade secrecy — private firms refuse to disclose how their models work
Affected citizens often don’t know they were scored, flagged, or matched
Regulators are under-resourced — data protection authorities lack the staff to enforce
“Human-in-the-loop” becomes rubber-stamping — officers trust the AI’s output rather than challenging it
AI and the Foundations of Democracy
Boix (2022): AI threatens democracy through three reinforcing channels
1. Economic: AI may widen inequality (capital vs. labor), weakening the social consensus that sustains democratic rule
2. Informational: AI empowers both opposition coordination and state repression—but authoritarian regimes have structural advantages in deploying repression
3. Cost of exclusion: AI lowers the cost of identifying, monitoring, and preempting opponents—making it cheaper to suppress dissent
AI and the Foundations of Democracy
Key insight: In strong democracies, elections can “re-equilibrate” policy in response to AI shocks. In emerging and peripheral economies, AI is more likely to be de-democratizing (Boix, 2022).
Source: Boix (2022), Oxford Handbook of AI Governance
Global Surveillance Camera Density
Figure 3
What Works Where?
Tool
Democracy
Hybrid Regime
Autocracy
Data minimization
Effective (enforceable)
Partial (inconsistent)
Ineffective (state ignores)
Algorithmic auditing
Effective (capacity exists)
Limited (expertise lacking)
Cosmetic (regime controls)
Judicial authorization
Effective (independent courts)
Variable (courts pressured)
Ineffective (courts captured)
International pressure
Moderate (less needed)
Potentially effective
Potentially effective (if costs imposed)
Author’s illustration. Effectiveness depends on institutional context, not technical design alone.
Summary: Four Key Takeaways
1. AI collapses the cost of monitoring — mass surveillance and predictive flagging are now cheap and scalable.
2. Privacy is a power resource — AI widens the asymmetry between transparent citizens and opaque institutions (Cohen).
3. Harms fall unequally — biased data and weak contestability concentrate costs on marginalized groups (Buolamwini & Gebru; Eubanks).
4. Institutions decide the outcome — strong democracies re-equilibrate; AI is most de-democratizing where checks are weakest (Boix; Acemoglu & Robinson).
References
Boix, C. (2022). AI and the economic and informational foundations of democracy. In J. B. Bullock et al. (Eds.), The Oxford Handbook of AI Governance. Oxford University Press.
Scott, J. C. (1998). Seeing like a state: How certain schemes to improve the human condition have failed. Yale University Press.
Tilly, C. (1990). Coercion, capital, and European states, AD 990–1990. Basil Blackwell.
Acemoglu, D., & Robinson, J. A. (2012). Why nations fail: The origins of power, prosperity, and poverty. Crown Business.
Cohen, J. E. (2019). Between truth and power: The legal constructions of informational capitalism. Oxford University Press.
Cohen, J. E. (2019). Turning privacy inside out. Theoretical Inquiries in Law, 20(1), 1–31.
Buolamwini, J., & Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of Machine Learning Research, 81, 77–91.
References
Barocas, S., & Selbst, A. D. (2016). Big data’s disparate impact. California Law Review, 104(3), 671–732.
Benjamin, R. (2019). Race after technology: Abolitionist tools for the new Jim code. Polity Press.
Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin’s Press.
Lum, K., & Isaac, W. (2016). To predict and serve? Significance, 13(5), 14–19.
O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Crown.
Penney, J. W. (2016). Chilling effects: Online surveillance and Wikipedia use. Berkeley Technology Law Journal, 31(1), 117–182.
Comparitech. (2021). Surveillance camera statistics: Which cities have the most CCTV cameras?
