Lawrence Person's BattleSwarm Blog 10/12/2025
Here’s a provocative Substack essay that argues that the 2020 Census was systemically, algorithmically polluted by a single data scientist.
The 2020 census was marketed as an “actual enumeration,” a neutral count of people for apportionment and funding. It was not. The same official who helped block a basic citizenship question in 2018, John M. Abowd, then the Census Bureau’s Chief Scientist, pushed through a new, opaque methodology in 2020 called differential privacy. The new system deliberately injected mathematical noise into every block count in America, turning the census from a headcount into a model with knobs. The knob that mattered most was a single parameter, epsilon, a secrecy shroud known only to a small inner circle. Abowd argued that a single added question about citizenship posed an intolerable risk to data quality because there was, he said, not enough time to test it. Then he rushed an untested algorithm that altered every count in every neighborhood. The irony is so sharp it cuts: the man who warned that one question might distort the census approved a method that guaranteed distortion.
Start with the record. On January 19, 2018, Abowd sent Commerce a technical memo urging rejection of a citizenship question. He then testified for several days in federal court. The transcript, nearly 700 pages, cemented a narrative that any citizenship question would degrade data and impede participation. The courts cited this drumbeat of doubt, and the question was blocked. The administration lost the public fight. But the inside fight over how to publish the data was only beginning. Abowd immediately advanced a quiet revolution in disclosure avoidance, adopting differential privacy for the first time ever in a US census. That choice, made outside the glare that attended the citizenship question, had far more sweeping consequences.
Differential privacy sounds harmless. In truth, it is a mechanism that turns correct data into false data according to a secret recipe. Abowd did not merely suppress a few cells in tiny places. Instead, he ran an algorithm across the map that perturbed the population of every census block, and it postprocessed the results so the fabricated numbers looked tidy. The output retained familiar columns, but the counts were no longer the counts. Abowd convinced his colleagues in the Bureau that implementing differential privacy was merely compliance with 13 U.S.C. § 9, its duty to protect confidentiality. Privacy is important. But privacy, as a constitutional matter, follows the enumeration, it does not negate it. A 2021 Harvard analysis of Abowd’s manipulation showed what this means in real life. When researchers simulated the Abowd’s algorithm using public test data, they found that differential privacy moves people around on paper, shifting them from one neighborhood to another in ways that make communities look less diverse and change their apparent political makeup. In plain terms, the system can make a mixed neighborhood look whiter or more uniform, and a balanced district look more partisan than it is. The study also showed that the noise makes it impossible to meet the Supreme Court’s “One Person, One Vote” rule, which requires legislative districts to have nearly equal populations. If each district’s population count is warped by secret noise, some citizens’ votes end up weighing more than others. When a method, by design, destabilizes the precise block totals that redistricting depends on, it stops being disclosure avoidance and becomes statistical alteration. The framers mandated counting people, not blurring them.
-----
If all this is true, President Trump’s call for a mid-decade census is more than justified. The constitution calls for an enumeration of citizens, not an algorithmic approximation poisoned by partisan pollution. A new count is needed to restore accuracy and remove illegal aliens from the census.
More:
https://www.battleswarmblog.com/?p=68150
