Quantifying Privacy

Shrey Shah

This paper is part of a collection of unpublished economics papers.

Blackwell’s (1953) theorem implies that more information always benefits a decision-maker. We show this logic reverses in labor markets with statistical discrimination: identity-typed signals—revealing who people are rather than what they can do—reduce aggregate welfare even though each employer individually prefers them. We call this the Anti-Blackwell property. A COINTELPRO triple-difference ($N = 9.9$M) estimates that the FBI’s surveillance of Black organizations widened the racial income gap by 7.4 log points ($p = 0.001$) in targeted cities, with effects growing through 2021—50 years after the program ended. Cross-country and Stasi evidence is consistent with the same mechanism. The income penalty concentrates in background-check-intensive occupations, indicating that surveillance-generated records compound through modern screening infrastructure.

1. Introduction

Does surveillance reduce economic output? The microeconomic literature on privacy typically finds small individual valuations: people will sell personal data for pennies (Acquisti, John, and Loewenstein, 2013), and the efficiency costs of privacy regulation appear modest (Goldfarb and Tucker, 2011). This paper provides evidence for a different view. We argue that surveillance—state and algorithmic—imposes large, persistent costs on economies by generating information about who people are rather than what they can do.

The costs operate through multiple channels simultaneously. Surveillance generates identity-correlated information that enables statistical discrimination, sorting talented workers into the wrong occupations (Hsieh, Hurst, Jones, and Klenow, 2019). It deters experimentation and entrepreneurship, especially among targeted groups (Lichter, Loffler, and Siegloch, 2021). It erodes interpersonal and institutional trust (Algan and Cahuc, 2010). And it creates records—criminal histories, credit flags, algorithmic risk scores—that follow individuals across markets and generations, compounding over time. No single channel accounts for the full cost, but they interact: an economy where surveillance enables discrimination, chills innovation, and destroys trust simultaneously will underperform an otherwise identical economy with privacy protection. Historical evidence from six settings spanning four continents and eight decades—from the Stasi to COINTELPRO to post-9/11 algorithmic policing—illustrates this pattern (Appendix D).

We contribute in three ways. First, we extend the Coate and Loury (1993) statistical discrimination framework by introducing a signal technology parameterized by identity-specificity. Where Coate and Loury show discriminatory equilibria sustained by beliefs, we show they are sustained—and amplified—by a technological infrastructure that generates and distributes identity signals. The central result—the Anti-Blackwell property—is that identity-typed signals reduce aggregate welfare even though each employer individually prefers them.

Second, we provide causal evidence that this mechanism operates empirically. A COINTELPRO triple-difference ($N = 9.9$M, 173 metros) estimates that the FBI’s surveillance of Black organizations widened the racial income gap by 7.4 log points ($p = 0.001$) in targeted cities, bridging the micro privacy literature (Acquisti, Taylor, and Wagman, 2016) and the misallocation literature (Hsieh et al., 2019). Cross-country descriptive patterns from 174 countries and the Stasi natural experiment (Lichter et al., 2021) are consistent with the same mechanism operating across settings.

Third, we document that the effects persist and compound. ACS extension through 2021 shows the COINTELPRO effect growing 50 years after the program ended (−0.109**, $p = 0.002$), with the income penalty concentrated in background-check-intensive occupations (2.0$\times$ differential). This is consistent with surveillance-generated records entering permanent screening infrastructure, producing the runaway feedback the model predicts (Proposition 2).

Related literature. The paper connects several literatures. In statistical discrimination theory, Arrow (1973), Phelps (1972), and Coate and Loury (1993) establish the foundations; Bohren, Imas, and Rosenberg (2019) study dynamics of discrimination; Bohren and Hull (2023) develop measurement theory for systemic discrimination. We add an explicit information technology that generates and distributes the identity signals enabling these equilibria. In privacy economics, Acquisti, Taylor, and Wagman (2016) survey micro evidence; Goldfarb and Tucker (2011) and Jia, Jin, and Wagman (2021) study regulation costs; Jones and Tonetti (2020), Bergemann and Bonatti (2022), and Acemoglu et al. (2022) analyze data markets. We complement this literature with causal evidence on macro-level effects. Lichter, Loffler, and Siegloch (2021) provide the closest empirical antecedent, identifying surveillance effects in East Germany; Beraja, Yang, and Yuchtman (2023) study China’s surveillance-AI feedback loop; Klarl et al. (2023) examine surveillance and innovation. We extend the evidence to cross-country and within-democracy settings and connect micro identification to aggregate welfare.

The remainder of the paper proceeds as follows. Section 2 develops the theoretical framework. Section 3 summarizes supporting evidence from cross-country data and the Stasi natural experiment. Section 4 presents US evidence, centered on the COINTELPRO triple-difference and the background-check channel. Section 5 discusses policy implications and limitations. Section 6 concludes.

2. Theoretical Framework: The Talent Misallocation Channel

We formalize one key channel through which surveillance reduces economic output: talent misallocation. This is the channel most amenable to clean theoretical treatment, but the total cost of surveillance includes innovation, trust, extraction, and human capital channels that the model does not capture (Section 2.7). The empirical tests in Sections 3–4 measure total effects.

2.1 Two Types of Information

Standard information economics treats more information as welfare-enhancing: better signals improve matching, reduce adverse selection, and increase surplus (Stiglitz, 1975; Grossman and Stiglitz, 1980). We argue this logic depends on the type of information.

Productivity information reveals an individual’s ability to perform specific tasks. A job interview, a skills test, or a work sample tells an employer how productive a particular worker will be. Productivity information improves the match between workers and tasks and unambiguously increases efficiency.

Identity information reveals an individual’s membership in a demographic, geographic, or social group. A name, an address, a photograph, a criminal record, a credit score—these tell an employer who the person is and where they come from, not what they can do. Identity information is the input to statistical discrimination: it allows employers to condition offers on group averages rather than individual characteristics.

The distinction matters because surveillance systems predominantly generate identity information. Criminal records databases, facial recognition, predictive policing algorithms, credit scoring systems, and social media monitoring all classify people into groups. They do not identify which particular member of a group is a talented scientist, entrepreneur, or surgeon.

The model’s key assumption is $\delta_g > \delta_\theta$: surveillance reduces net signal quality for targeted groups because identity information dominates productivity information. Published evidence on individual signal quality supports this: criminal records predict race (AUC $\approx$ 0.85) far better than they predict recidivism (AUC $\approx$ 0.65; Dressel and Farid, 2018); facial recognition error rates are 10–100 times higher for dark-skinned faces (Buolamwini and Gebru, 2018); and credit scores correlate with race at $r > 0.3$ but explain less than 5% of job performance variance (Bernerth et al., 2012). The assumption is not directly testable in our data, but these magnitudes indicate that common surveillance outputs carry substantially more identity than productivity content.

2.2 Setup

Workers. An economy has $G$ demographic groups indexed by $g \in {1, \ldots, G}$, with population shares $(p_1, \ldots, p_G)$ satisfying $\sum_g p_g = 1$. Define the Herfindahl index of group diversity $H = 1 - \sum_g p_g^2$. Each worker $i$ in group $g$ has talent $\theta_i$ drawn from a common distribution $F(\theta)$ on $[0, 1]$, independent of group membership.

Occupations. There are $J$ occupations. Occupation $j$ has output $y_j = \omega_j \theta$ for a worker with talent $\theta$, where $\omega_1 > \omega_2 > \ldots > \omega_J$.

Investment. Workers invest in human capital at cost $c_i \sim U[0, 1]$, independent of group. Investment yields qualification for occupation 1 with certainty.

Signal technology. Employers observe a noisy signal $s_i \in {0, 1}$:

\[P(s = 1 \mid \text{qualified}, g) = \alpha_g, \qquad P(s = 1 \mid \text{unqualified}, g) = \beta\]

Surveillance. Surveillance intensity $\delta \geq 0$ decomposes as:

\[\alpha_g = \alpha + \underbrace{\delta_\theta}_{\text{productivity info}} - \underbrace{\delta_g(\delta)}_{\text{identity penalty}}\]

where $\delta_\theta \geq 0$ is the improvement from productivity information and $\delta_g(\delta) \geq 0$ is the identity penalty for group $g$. The key assumption is that for targeted groups, $\delta_g > \delta_\theta$: surveillance reduces effective signal quality because it generates more identity information than productivity information. For the non-targeted group, $\delta_1 = 0$; for targeted groups, $\delta_g = \delta \cdot \phi_g$ with $\sum_g p_g \phi_g = 1$.

Employers. Competitive employers observe group membership $g_i$ and signal $s_i$, assigning workers with $s_i = 1$ to the high-return occupation at wage $w_g = \omega_1 \cdot P(\text{qualified} \mid s = 1, g)$.

2.3 The Anti-Blackwell Property

Definition (Information type). An improvement to the signal structure is productivity-typed if it increases $\alpha_g$ equally for all $g$; identity-typed if it decreases $\alpha_g$ for targeted groups while leaving $\alpha_1$ unchanged.

Proposition 0 (Anti-Blackwell). Consider a Blackwell-improving signal structure $\tilde{S}$ more informative than $S$ about the joint variable (qualification, group).

(i) If the improvement is purely productivity-typed, aggregate output and welfare both increase.

(ii) If the improvement is purely identity-typed, aggregate output and welfare both decrease.

(iii) The welfare effect of a general improvement is negative whenever $\delta_g > \delta_\theta$ for all targeted groups.

(Proof: Appendix A.0.)

Blackwell’s (1953) theorem establishes that a decision-maker always weakly prefers more information. Proposition 0 shows this logic inverts in the presence of statistical discrimination. The reason is that Blackwell’s theorem applies to a single decision-maker. In a labor market with competitive employers who observe group membership, more information about identity enables group-conditional updating. The employers are individually rational, but the aggregate outcome is worse because identity information reduces targeted workers’ incentive to invest. The information is Blackwell-improving for each employer but welfare-reducing for the economy.

This identifies a specific informational structure—identity-typed signals—that reverses the welfare implications of better information. The Anti-Blackwell property predicts that expanding surveillance technologies reduces aggregate output, a prediction we test in Sections 3–4.

2.4 Equilibrium and Aggregate Output

A static equilibrium is a vector $(\pi_1, \ldots, \pi_G)$ where $\pi_g$ is the fraction of group $g$ workers who invest. Interior equilibria satisfy:

\[\pi_g^\ast = \alpha_g \omega_1 - \frac{\beta}{\alpha_g - \beta}\]

Lemma 1 (Group-Specific Investment). For $\alpha_g > \alpha_{g’}$, the interior equilibrium satisfies $\pi_g^\ast > \pi_{g’}^\ast$ and $w_g^\ast > w_{g’}^\ast$. Any group targeted by surveillance invests less than the non-targeted group. (Proof: Appendix A.1.)

Proposition 1 (Surveillance Reduces Aggregate Output). Suppose $\delta > 0$ and $\delta_g > \delta_\theta$ for targeted groups. Then $Y(\delta) < Y^\ast$, with TFP loss:

\[\Delta \log Y \approx -\delta \cdot \left(\sum_{g=2}^{G} p_g \phi_g \right) \cdot \left[\omega_1 + \frac{\beta}{(\alpha - \beta)^2}\right]\]

This is increasing in surveillance intensity, diversity, and occupational complexity. (Proof: Appendix A.2.)

The TFP loss has two components: underinvestment by targeted groups and signal degradation among those who do invest. Both reduce the economy’s effective use of its talent pool. Hsieh et al. (2019) attribute 20–40 percent of US growth from 1960–2010 to reduced barriers for women and Black Americans. Our model provides one mechanism for these barriers—surveillance generates the identity information that enables occupation-specific discrimination—but surveillance also creates barriers through other channels, so the TFP expression above is a conservative lower bound.

2.5 Dynamic Extension

We endogenize surveillance:

\[\delta(t+1) = \gamma \cdot \left[\max_g \pi_g^\ast(\delta(t)) - \min_g \pi_g^\ast(\delta(t))\right]\]

where $\gamma > 0$ governs how outcome disparities translate into differential surveillance.

Proposition 2 (Runaway Feedback). (i) $\delta = 0$ is a steady state. (ii) If $\gamma > \gamma^\ast \equiv [\omega_1 + \beta/(\alpha - \beta)^2]^{-1}$, it is unstable. (iii) There exists a unique discriminatory steady state $\delta^\ast > 0$. (Proof: Appendix A.3.)

Under calibrated parameters ($\alpha = 0.9$, $\beta = 0.2$, $\omega_1 = 1.2$), $\gamma^\ast \approx 0.49$. Even modest responsiveness destabilizes equality: any shock that introduces differential surveillance triggers self-reinforcing divergence. This extends Coate and Loury (1993): in their model, discriminatory equilibria are sustained by beliefs; here, they are sustained by a technological infrastructure that continuously regenerates confirming data.

2.6 Welfare Analysis and Extensions

Proposition 3 (Social Welfare Decreasing in Surveillance). $W(\delta)$ is strictly decreasing in $\delta$ for all $\delta > 0$ such that interior equilibria exist. (Proof: Appendix A.4.)

Proposition 4 (Surveillance as Implicit Tax). The implicit tax $\tau_g(\delta) = 1 - \pi_g^\ast(\delta) \alpha_g(\delta) / (\pi_1^\ast \alpha)$ satisfies $\tau_g(0) = 0$ and $\tau_g > 0$ for $\delta > 0$. (Proof: Appendix A.5.)

Table 1: Calibrated Implicit Surveillance Tax by Surveillance Intensity

$\delta$	$\pi_{\text{targeted}}^\ast$	$\pi_{\text{non-targeted}}^\ast$	$\tau_{\text{targeted}}$	Interpretation
0	0.794	0.794	0%	No surveillance bias
0.05	0.713	0.794	14.5%	Mild (credit score noise)
0.10	0.627	0.794	29.8%	Moderate (over-policing + credit)
0.15	0.536	0.794	43.8%	Severe (full surveillance stack)

Parameters: $\alpha = 0.9$, $\beta = 0.2$, $\omega_1 = 1.2$.

Proposition 5 (Cross-Market Amplification). When the same identity data feeds into $M$ markets, $\tau_g^{\text{total}} > \sum_m \tau_{g,m}$. (Proof: Appendix A.6.)

Proposition 6 (Over-Surveillance). When surveillance generates genuine public benefits $b(\delta)$, the competitive equilibrium over-surveils: $\delta^{CE} > \delta^{SP}$ because the surveillance industry externalizes misallocation costs. (Proof: Appendix A.7.)

2.7 Beyond Misallocation

The misallocation model formalizes one channel. Surveillance harms economies through at least five additional channels, each empirically documented:

Chilling effects on innovation. Surveillance deters experimentation and risk-taking. Lichter et al. (2021) find reduced self-employment and patenting from Stasi exposure; Penney (2016) documents self-censorship following the Snowden revelations; Bernstein (2012) shows productivity increases when workers are shielded from observation.
Trust destruction. Surveillance erodes interpersonal and institutional trust. Algan and Cahuc (2010) estimate that one SD of trust raises GDP growth by 0.5 percentage points per year. Lichter et al. show Stasi exposure persistently reduced trust.
Direct extraction. Surveillance data enables algorithmic price discrimination (Bartlett et al., 2022), predatory lending, and municipal fine extraction—the DOJ’s Ferguson Report documented 90,000 citations in a city of 21,000 residents.
Incarceration. The US incarcerates 2 million people; 20% of Black men born 1965–69 served prison time by age 35 (Western and Pettit, 2010). Criminal records reduce callbacks by 50% (Pager, 2003).
Organizational destruction. COINTELPRO was explicitly designed to destroy Black economic organizations, using surveillance to identify donors, disrupt cooperatives, and manufacture prosecutions against organizational leaders.

The TFP decomposition (Section 3) indicates that misallocation accounts for approximately one-fifth of the total GDP effect, with the remaining four-fifths operating through the channels above.

2.8 Testable Predictions

The model generates predictions about the misallocation channel specifically; the empirical tests measure total effects, which should be at least as large. The model yields 12 testable predictions—4 quantitative and 8 directional—summarized in Appendix Table A18. The evidence is consistent with all 12: four quantitative matches fall within confidence intervals and all eight directional signs are correct. The complexity prediction provides the cleanest structural test: the TFP $\times$ complexity interaction (−0.089***, Section 3) directly tests whether misallocation costs are amplified by economic complexity, exploiting variation in $\omega_1$ independent of surveillance intensity.

3. Supporting Evidence: Cross-Country and Stasi

Cross-country data and the Stasi natural experiment provide supporting evidence consistent with the model’s predictions. Neither resolves the fundamental challenge of isolating surveillance from broader institutional quality, but both inform magnitudes and mechanisms. The paper’s causal identification rests on the COINTELPRO triple-difference in Section 4.

Cross-country patterns. In a decade panel of 174 countries (1960–2019), within-country surveillance predicts GDP per worker (−0.185***, country FE) and TFP (−0.039**; Appendix Tables A2, A4). Controlling for democracy renders the GDP coefficient insignificant, because surveillance and democracy are collinear at $r = -0.86$ within countries—a collinearity no instrument resolves (Appendix C.12–C.13). Three results nevertheless provide structural specificity beyond regime type. First, the TFP $\times$ complexity interaction (−0.089***, Appendix Table A5) is robust to all democracy specifications—it exploits variation in economic structure rather than regime type, directly testing whether misallocation costs increase with occupational complexity (Prediction C3). Second, in a horse race with media censorship, corruption, and rule of law, surveillance retains significance for TFP while media censorship drops to insignificance (Appendix Table A4). Third, restricting to countries above-median democracy throughout the panel, the coefficient is −0.196*** ($N = 267$). The TFP coefficient (−0.039) is approximately 21% of the GDP coefficient (−0.182), indicating that misallocation accounts for roughly one-fifth of the total cost; the remaining four-fifths operate through factor accumulation channels (Section 2.7).

The Stasi natural experiment. Lichter, Loffler, and Siegloch (2021) provide the cleanest causal evidence in the existing literature, exploiting county-level variation in informer density across East Germany (1 informer per 63 citizens). One standard deviation of Stasi exposure reduced monthly income by 84 euros ($\approx -5\%$), with simultaneous negative effects on self-employment, patenting, and interpersonal trust—consistent with the misallocation, chilling, and trust channels operating simultaneously (Section 2.7). The Stasi was a pure identity-surveillance system: the information collected (Western contacts, church attendance, political opinions) had no productivity content ($\delta_\theta \approx 0$). Effects persist more than two decades after dissolution. Across 29 post-communist countries, pre-transition surveillance intensity predicts 2010s GDP (−0.193**, Appendix Table A11).

4. United States: COINTELPRO and the Background-Check Channel

4.1 COINTELPRO Triple-Difference

Identification. The FBI’s COINTELPRO-BLACK program (1967–1971) targeted Black political and economic organizations in 11 metropolitan areas with documented active operations: Chicago, Oakland, Los Angeles, New York, Philadelphia, Memphis, Detroit, Cleveland, Baltimore, Washington DC, and St. Louis. We estimate the causal effect using a triple-difference design on individual-level Census microdata from IPUMS USA (1940–2000):

\[\log y_{igct} = \alpha_{ct} + \delta_{rt} + \gamma_{cr} + \beta(\text{Black}_i \times \text{Treated}_c \times \text{Post}_t) + X_i'\Gamma + \varepsilon_{igct}\]

where $\alpha_{ct}$ are metro $\times$ year fixed effects (absorbing any city-specific economic shocks affecting both races, such as deindustrialization or aerospace growth), $\delta_{rt}$ are race $\times$ year fixed effects (absorbing national trends in the racial gap, such as Civil Rights Act effects and affirmative action), $\gamma_{cr}$ are metro $\times$ race fixed effects (absorbing time-invariant city-specific racial gaps), and $X_i$ includes individual controls (age, age$^2$, years of education, marital status, sex). The coefficient $\beta$ captures whether the Black-White log income gap widened more in COINTELPRO-targeted cities after 1967 than in comparison cities—a triple-difference that differences out both race-neutral city shocks and city-neutral racial trends. Because all treated cities share a single treatment date (1967), the design is not subject to the staggered-timing bias identified by Sun and Abraham (2021) and de Chaisemartin and d’Haultfoeuille (2020).

Data. IPUMS USA Census microdata: working-age adults (25–64) with positive wage income, Black and White only, in identifiable metropolitan areas. The comparison group includes 121 metropolitan areas meeting data-driven population thresholds, after excluding 6 metros with documented COINTELPRO field offices from the control group. Total observations: 11,387,749 across 173 metros and 7 decades (1940–2000). Person-weighted throughout; standard errors clustered at the metropolitan area level (132 clusters).

Table 2: COINTELPRO Individual-Level Triple-Difference

Specification	DDD Coefficient	SE	p-value	N
Saturated FE + controls, drop 1960 (preferred)	−0.074***	0.023	0.001	9,954,271
Saturated FE + controls, full sample	−0.227*	0.094	0.016	11,387,749
Saturated FE, no controls	−0.347**	0.122	0.005	11,387,749
Intensity-weighted DDD	−0.111***	0.028	$<$0.001	11,387,749
Original 6 treated + all controls	−0.289*	0.114	0.011	11,387,749

Notes: DV = log(real wage income, 2020\$). Standard errors clustered at metropolitan area level. 11 treated cities, 121 clean controls (6 contaminated controls excluded). Preferred specification drops the 1960 Census wave, which contains a significant pre-trend (see text). *** $p<0.001$, ** $p<0.01$, * $p<0.05$.

The preferred specification drops the 1960 Census wave—which contains a significant positive pre-trend (discussed below)—yielding DDD $= -0.074$ (SE $= 0.023$, $p = 0.001$; $N = 9{,}954{,}271$). This estimate relies only on 1940 and 1950 as pre-treatment periods, neither of which shows a significant pre-trend. Including the 1960 wave yields −0.227 (SE $= 0.094$, $p = 0.016$), which exploits the full pre-post contrast but is sensitive to the 1960 convergence peak. The intensity specification shows a dose-response: a unit increase in COINTELPRO intensity widens the gap by 11.1 log points ($p < 0.001$).

Inference with few treated clusters. The design has 11 treated clusters, warranting careful inference. The preferred specification (drop 1960) yields $p = 0.001$ with conventional metro-clustered standard errors. For the full-sample specification, we additionally report: wild cluster bootstrap with Rademacher weights (Cameron, Gelbach, and Miller, 2008; $p = 0.003$), Fisher randomization inference with 5,000 permutations ($p = 0.038$), and the ACS extension with an independent sample of 221 metros ($p < 0.001$). All reject the null.

Six comparison cities were reclassified as contaminated controls after FBI archival records revealed active COINTELPRO field offices; excluding them increases the estimate from −0.198 to −0.227 (Appendix Table A13).

Event study. The year-specific DDD coefficients (reference: 1950) are as follows:

Year	Coefficient	SE	p-value	Pre/Post
1940	+0.020	0.038	0.594	Pre
1950	0 (ref)	—	—	Reference
1960	+0.212*	0.101	0.035	Pre
1970	+0.027	0.026	0.294	Post
1980	−0.050+	0.029	0.081	Post
1990	−0.073*	0.031	0.017	Post
2000	−0.137***	0.036	$<$0.001	Post

The 1940 coefficient is small and insignificant (+0.020, $p = 0.594$), supporting parallel trends before the civil rights movement. However, the 1960 coefficient (+0.212, $p = 0.035$) is a positive and significant pre-trend: treated cities experienced faster Black-White income convergence in the decade before COINTELPRO. This threatens the identification if the post-treatment decline merely reflects mean reversion from an anomalous 1960 peak.

Three results address this concern. First, re-normalizing the event study to use 1960 as the reference year, the post-treatment coefficients are all large and negative: −0.185 (1970), −0.262 (1980), −0.285 (1990), −0.349 (2000). Mean reversion from a peak predicts coefficients returning toward zero (i.e., back to pre-1960 levels); instead they diverge monotonically in the opposite direction. Under any mean-reversion model, the treatment effect is a lower bound. Second, re-estimating the DDD dropping the 1960 Census wave entirely yields −0.074 (SE = 0.023, $p = 0.001$; $N = 9{,}954{,}271$). The estimate is smaller but more precise, and the pre-trend plays no role. Third, Rambachan-Roth (2023) bounds exclude zero through $\bar{M} = 0.4$ for all post-treatment years. A trend-break model estimates a slope reversal from +0.039/decade (pre) to −0.042/decade (post, $p < 0.001$), consistent with an abrupt structural change rather than gradual mean reversion (Appendix Table A13b).

The 1960 coefficient is also consistent with historical non-randomness: COINTELPRO targeted cities where Black organizations were most active, and organizational activity plausibly drove convergence. But the robustness evidence does not depend on this interpretation.

ACS persistence through 2021. We extend the analysis through 2021 using ACS 1-year samples, merged via a population-weighted MET2013-to-METAREA crosswalk. The extended sample includes 221 metros (11 treated, 210 clean controls) and $N = 15{,}112{,}394$. The ACS trajectory shows not just persistence but growth: −0.082* (2006), −0.102** (2016), −0.109** (SE = 0.036, $p = 0.002$) for 2021—50 years after the program ended. The 2021 coefficient is the largest in the entire event study. The persistence ratio (mean ACS / mean Census post-treatment) is 1.86. The full-sample DDD is −0.087*** (SE = 0.022, $p = 0.0001$; wild cluster bootstrap $p = 0.026$). Within the same extract, the DDD is stable at approximately −0.08 regardless of control group size (Appendix Table A13d), consistent with the preferred Census estimate (−0.074).

Falsification and robustness. Placebo outcomes support misallocation over city-level decline: the DDD on employment is positive (+0.030**, $p = 0.003$), and the education DDD is zero. The result survives dropping any single treated city ($p < 0.001$ in all cases). Heterogeneity, mechanism decomposition, and 9 additional specifications are in Appendix C.2.

Cross-sectional evidence from the Opportunity Atlas ($N = 27{,}792$ tracts) and county-level surveillance technology adoption ($N = 3{,}222$) corroborate these patterns; see Appendix C.4 and C.6.

4.2 The Background-Check Channel

The Anti-Blackwell property predicts that identity-typed information reduces welfare through labor market sorting. If this mechanism drives the COINTELPRO effect’s persistence, the income penalty should concentrate in occupations where employers access surveillance-generated records—those requiring background checks.

We classify occupations by screening intensity. High-screening occupations (finance, government, healthcare, education, professional services) routinely require criminal background checks, credit checks, or security clearances. Low-screening occupations (construction, hospitality, manufacturing, agriculture) typically do not. The DDD is −0.136*** in high-screening occupations versus −0.069** in low-screening occupations—a 2.0$\times$ differential (Appendix C.2).

The growing differential is consistent with the expansion of background-check infrastructure: fewer than 10% of employers conducted criminal background checks in 1970; 94% do today. This pattern is diagnostic. If COINTELPRO harmed Black communities through generic economic disruption—destroying businesses, deterring investment, depressing local demand—the income penalty would not concentrate in background-check-intensive occupations specifically. The occupational heterogeneity indicates an information channel: surveillance-generated records (arrests, FBI files, credit disruptions) flow through the screening infrastructure that now mediates labor market access for 77 million adults with criminal records.

The pattern also explains the growing ACS trajectory. As screening infrastructure expands, identity records generated in the 1960s reach into more labor markets, consistent with the cross-market amplification of Proposition 5. The Ban-the-Box evidence (Appendix C.14) provides an independent test: Agan and Starr (2018) find that removing criminal history from applications increases racial discrimination against applicants without records—employers substitute to race as the remaining correlated identity signal. This is exactly the channel substitution the model predicts: restricting one identity channel shifts employer reliance to others, unless all correlated channels are restricted simultaneously.

5. Discussion

5.1 Policy Implications

The model implies three policy levers: reducing surveillance intensity $\delta$ (banning facial recognition, restricting data broker sales), decoupling identity from productivity signals (mandating blind hiring, restricting credit-based employment screening), and breaking the feedback loop (requiring algorithmic auditing, sunsetting surveillance programs absent demonstrated efficacy). The background-check channel evidence (Section 4.2) suggests that single-channel interventions (e.g., Ban-the-Box alone) are insufficient; effective policy must restrict correlated identity channels simultaneously.

5.2 Limitations

Several limitations warrant discussion. First, the model makes tractability-driven assumptions: binary signals, uniform costs, and a simplified occupational structure. Second, surveillance and democracy are collinear within countries ($r = -0.86$); the cross-country evidence (Section 3) is descriptive, and the paper’s causal identification rests on COINTELPRO. Third, the COINTELPRO identification has a significant 1960 pre-trend; the preferred specification drops 1960, yielding a smaller estimate (−0.074) than the full-sample estimate (−0.227). Fourth, extrapolation from COINTELPRO to modern surveillance requires assumptions about relative intensity for which we have no direct evidence. Fifth, surveillance generates genuine benefits (crime prevention, public health, administrative efficiency) that we model only through the general benefit function $b(\delta)$. Welfare cost estimates under various assumptions are in Appendix C.10–C.11.

6. Conclusion

Blackwell’s theorem says more information is always better for a decision-maker. We show this reverses in labor markets with statistical discrimination: identity-typed signals reduce aggregate welfare even though each employer individually prefers them. The Anti-Blackwell property formalizes why expanding surveillance technologies can reduce economic output.

The COINTELPRO triple-difference provides direct evidence. The FBI’s surveillance of Black organizations widened the racial income gap by 7.4 log points ($p = 0.001$) in targeted cities, with effects growing through 2021—50 years after the program ended. The income penalty concentrates in background-check-intensive occupations, consistent with surveillance-generated records compounding through modern screening infrastructure. Cross-country descriptive patterns and the Stasi natural experiment support the same mechanism from independent settings.

The findings suggest that the relevant policy question is not whether to regulate surveillance, but how to shift information systems from identity-typed toward productivity-typed signals: mandating blind hiring, restricting credit-based employment screening, and sunsetting surveillance programs absent demonstrated efficacy.

Data Availability Statement. Cross-country data are from V-Dem v13 (freely available at v-dem.net), Penn World Table 10.0 (rug.nl/ggdc/productivity/pwt), and World Bank Development Indicators. US individual-level data are from IPUMS USA (ipums.org); Census microdata (1940–2000) and ACS 1-year samples (2005–2021) require a free IPUMS account. Opportunity Atlas data are from opportunityatlas.org. All analysis code, variable construction procedures, and instructions for replicating IPUMS extracts are included in the replication package.

References

Abadie, A., A. Diamond, and J. Hainmueller. 2010. “Synthetic Control Methods for Comparative Case Studies.” Journal of the American Statistical Association 105(490): 493–505.

Acemoglu, D., T.A. Hassan, and J.A. Robinson. 2011. “Social Structure and Development: A Legacy of the Holocaust in Russia.” Quarterly Journal of Economics 126(2): 895–946.

Acemoglu, D., S. Naidu, P. Restrepo, and J.A. Robinson. 2019. “Democracy Does Cause Growth.” Journal of Political Economy 127(1): 47–100.

Acemoglu, D., A. Makhdoumi, A. Malekian, and A. Ozdaglar. 2022. “Too Much Data: Prices and Inefficiencies in Data Markets.” American Economic Journal: Microeconomics 14(4): 218–256.

Acquisti, A., C.R. Taylor, and L. Wagman. 2016. “The Economics of Privacy.” Journal of Economic Literature 54(2): 442–492.

Agan, A. and S. Starr. 2018. “Ban the Box, Criminal Records, and Racial Discrimination: A Field Experiment.” Quarterly Journal of Economics 133(1): 191–235.

Alesina, A., A. Devleeschauwer, W. Easterly, S. Kurlat, and R. Wacziarg. 2003. “Fractionalization.” Journal of Economic Growth 8(2): 155–194.

Alesina, A. and N. Fuchs-Schundeln. 2007. “Good-bye Lenin (or Not?): The Effect of Communism on People’s Preferences.” American Economic Review 97(4): 1507–1528.

Algan, Y. and P. Cahuc. 2010. “Inherited Trust and Growth.” American Economic Review 100(5): 2060–2092.

Arrow, K.J. 1973. “The Theory of Discrimination.” In O. Ashenfelter and A. Rees (eds.), Discrimination in Labor Markets. Princeton University Press.

Bartlett, R., A. Morse, R. Stanton, and N. Wallace. 2022. “Consumer-Lending Discrimination in the FinTech Era.” Journal of Financial Economics 143(1): 30–56.

Becker, G.S. 1957. The Economics of Discrimination. University of Chicago Press.

Becker, S.O., L. Mergele, and L. Woessmann. 2020. “The Separation and Reunification of Germany.” Journal of Economic Perspectives 34(2): 143–171.

Beraja, M., A. Kao, D.Y. Yang, and N. Yuchtman. 2023. “Exporting the Surveillance State via Trade in AI.” NBER Working Paper 31676.

Beraja, M., D.Y. Yang, and N. Yuchtman. 2023. “Data-intensive Innovation and the State: Evidence from AI Firms in China.” Review of Economic Studies 90(4): 1701–1723.

Bergemann, D. and A. Bonatti. 2022. “Data, Competition, and Digital Platforms.” Working Paper.

Bernstein, E. 2012. “The Transparency Paradox: A Role for Privacy in Organizational Learning and Operational Control.” Administrative Science Quarterly 57(2): 181–216.

Bernerth, J.B., S.G. Taylor, H.J. Walker, and D.S. Whitman. 2012. “An Empirical Investigation of Dispositional Antecedents and Performance-Related Outcomes of Credit Scores.” Journal of Applied Psychology 97(2): 469–478.

Bertrand, M. and S. Mullainathan. 2004. “Are Emily and Greg More Employable than Lakisha and Jamal?” American Economic Review 94(4): 991–1013.

Blackwell, D. 1953. “Equivalent Comparisons of Experiments.” Annals of Mathematical Statistics 24(2): 265–272.

Bohren, J.A. and P. Hull. 2023. “Systemic Discrimination: Theory and Measurement.” Working Paper.

Bohren, J.A., A. Imas, and M. Rosenberg. 2019. “The Dynamics of Discrimination: Theory and Evidence.” American Economic Review 109(10): 3395–3436.

Buolamwini, J. and T. Gebru. 2018. “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification.” Proceedings of Machine Learning Research 81: 1–15.

Cameron, A.C., J.B. Gelbach, and D.L. Miller. 2008. “Bootstrap-Based Improvements for Inference with Clustered Errors.” Review of Economics and Statistics 90(3): 414–427.

Canay, I.A., J.P. Romano, and A.M. Shaikh. 2017. “Randomization Tests Under an Approximate Symmetry Assumption.” Econometrica 85(3): 1013–1030.

Chetty, R. 2009. “Sufficient Statistics for Welfare Analysis.” Annual Review of Economics 1: 451–488.

Chetty, R., N. Hendren, M.R. Jones, and S.R. Porter. 2020. “Race and Economic Opportunity in the United States: An Intergenerational Perspective.” Quarterly Journal of Economics 135(2): 711–783.

Coate, S. and G.C. Loury. 1993. “Will Affirmative-Action Policies Eliminate Negative Stereotypes?” American Economic Review 83(5): 1220–1240.

Coppedge, M., J. Gerring, C.H. Knutsen, et al. 2023. “V-Dem Dataset v13.” Varieties of Democracy (V-Dem) Project.

Cunningham, D. 2004. There’s Something Happening Here: The New Left, the Klan, and FBI Counterintelligence. University of California Press.

Crawford, N.C. 2021. “The U.S. Budgetary Costs of the Post-9/11 Wars.” Watson Institute, Brown University.

de Chaisemartin, C. and X. d’Haultfoeuille. 2020. “Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects.” American Economic Review 110(9): 2964–2996.

Derenoncourt, E., C.H. Kim, M. Kuhn, and M. Schularick. 2022. “Wealth of Two Nations: The U.S. Racial Wealth Gap, 1860–2020.” Quarterly Journal of Economics 137(4): 1325–1383.

Dressel, J. and H. Farid. 2018. “The Accuracy, Fairness, and Limits of Predicting Recidivism.” Science Advances 4(1): eaao5580.

Doleac, J.L. and B. Hansen. 2020. “The Unintended Consequences of ‘Ban the Box’: Statistical Discrimination and Employment Outcomes When Criminal Histories Are Hidden.” Journal of Labor Economics 38(2): 321–374.

Fang, H. and A. Moro. 2011. “Theories of Statistical Discrimination and Affirmative Action: A Survey.” In J. Benhabib, A. Bisin, and M.O. Jackson (eds.), Handbook of Social Economics. North-Holland.

Feenstra, R.C., R. Inklaar, and M.P. Timmer. 2015. “The Next Generation of the Penn World Table.” American Economic Review 105(10): 3150–3182.

Feldstein, S. 2019. “The Global Expansion of AI Surveillance.” Carnegie Endowment for International Peace Working Paper.

Goldfarb, A. and C.E. Tucker. 2011. “Privacy Regulation and Online Advertising.” Management Science 57(1): 57–71.

Grossman, S.J. and J.E. Stiglitz. 1980. “On the Impossibility of Informationally Efficient Markets.” American Economic Review 70(3): 393–408.

Hendren, N. and B. Sprung-Keyser. 2020. “A Unified Welfare Analysis of Government Policies.” Quarterly Journal of Economics 135(3): 1209–1318.

Hipp, A., D. Czarnitzki, A. Hussinger, and J. Trunschke. 2022. “Comprehensive Patent Data of the German Democratic Republic 1949–1990.” GESIS Data Archive, DOI: 10.7802/2423.

Hsieh, C.-T., E. Hurst, C.I. Jones, and P.J. Klenow. 2019. “The Allocation of Talent and U.S. Economic Growth.” Econometrica 87(5): 1439–1474.

Hsieh, C.-T. and P.J. Klenow. 2009. “Misallocation and Manufacturing TFP in China and India.” Quarterly Journal of Economics 124(4): 1403–1448.

Ichihashi, S. 2020. “Online Privacy and Information Disclosure by Consumers.” American Economic Review 110(2): 569–595.

Jia, J., G.Z. Jin, and L. Wagman. 2021. “The Short-Run Effects of GDPR on Technology Venture Investment.” Marketing Science 40(4): 661–684.

Jones, C.I. and C. Tonetti. 2020. “Nonrivalry and the Economics of Data.” American Economic Review 110(9): 2819–2858.

Kaplan, D.E. 2009. “The Impact of Apartheid on Economic Growth.” Comparative Economic Studies 51(4).

Klarl, T., D.F. Karpa, M.E. Leusin, and M. Rochlitz. 2023. “Authoritarian Surveillance, Innovation and Growth.” SSRN Working Paper.

Knack, S. and P. Keefer. 1997. “Does Social Capital Have an Economic Payoff? A Cross-Country Investigation.” Quarterly Journal of Economics 112(4): 1251–1288.

Leshno, J. and P. Spector. 2022. “The Value of Information in Matching Markets.” Working Paper.

Lichter, A., M. Loffler, and S. Siegloch. 2021. “The Long-Term Costs of Government Surveillance: Insights from Stasi Spying in East Germany.” Journal of the European Economic Association 19(2): 741–789.

McLaughlin, M., C. Pettus-Davis, D. Brown, C. Veeh, and T. Renn. 2016. “The Economic Burden of Incarceration in the U.S.” Institute for Advancing Justice Research and Innovation, Washington University in St. Louis.

Mueller, J. and M.G. Stewart. 2014. “Evaluating Counterterrorism Spending.” Journal of Economic Perspectives 28(3): 237–248.

Mueller-Smith, M. 2015. “The Criminal and Labor Market Impacts of Incarceration.” Working Paper, University of Michigan.

Oster, E. 2019. “Unobservable Selection and Coefficient Stability: Theory and Evidence.” Journal of Business and Economic Statistics 37(2): 187–204.

Pager, D. 2003. “The Mark of a Criminal Record.” American Journal of Sociology 108(5): 937–975.

Penney, J.W. 2016. “Chilling Effects: Online Surveillance and Wikipedia Use.” Berkeley Technology Law Journal 31(1): 117–182.

Perry, A.M., J. Rothwell, and D. Harshbarger. 2018. “The Devaluation of Assets in Black Neighborhoods.” Brookings Institution.

Phelps, E.S. 1972. “The Statistical Theory of Racism and Sexism.” American Economic Review 62(4): 659–661.

Putnam, R.D. 2000. Bowling Alone: The Collapse and Revival of American Community. Simon and Schuster.

Rambachan, A. and J. Roth. 2023. “A More Credible Approach to Parallel Trends.” Review of Economic Studies (forthcoming).

Sun, L. and S. Abraham. 2021. “Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects.” Journal of Econometrics 225(2): 175–199.

Sunstein, C.R. 2003. “Terrorism and Probability Neglect.” Journal of Risk and Uncertainty 26: 121–136.

Tirole, J. 1996. “A Theory of Collective Reputations (with Applications to the Persistence of Corruption and to Firm Quality).” Review of Economic Studies 63(1): 1–22.

U.S. Department of Justice. 2015. “Investigation of the Ferguson Police Department.” Civil Rights Division.

U.S. Senate, Select Committee to Study Governmental Operations with Respect to Intelligence Activities. 1976. “Intelligence Activities and the Rights of Americans.” Final Report, Book III. 94th Congress, 2nd Session.

Waldinger, F. 2016. “Bombs, Brains, and Science: The Role of Human and Physical Capital for the Creation of Scientific Knowledge.” Review of Economics and Statistics 98(5): 811–831.

Western, B. and B. Pettit. 2010. “Incarceration and Social Inequality.” Daedalus 139(3): 8–19.

Zeng, Z. and J. Eisenman. 2018. “The Price of Persecution: The Long-Term Effects of the Anti-Rightist Campaign on Economic Performance in Post-Mao China.” World Development 109: 249–260.