A look at the fight over using differential privacy techniques in the US Census, which anonymizes datasets but some argue lowers the quality of the datasets
Todd Feathers / The Markup : Tweets: @juliaangwin and @toddfeathers Tweets: Julia Angwin / @juliaangwin : No one covers differential privacy like we do @themarkup! @ToddFeathers reports on the Census Bureaus' struggle to use fancy math to protect citizen data from being re-identified. https://themarkup.org/... Todd Feathers / @toddfeathers : In 2016 the Census Bureau conducted a test on the supposedly confidential statistics it publishes. By combining those statistics with commercial data, it was able to re-identify 52 million Americans by name and link them to their 2010 Census responses. https://themarkup.org/...
Context & Ripple Effects
The Census Bureau came to differential privacy defensively: a 2016 internal test showed that combining the statistics it routinely publishes with commercial data could re-identify 52 million Americans by name against their 2010 Census responses. That vulnerability was exposed just as the Bureau prepared its first online census, which experts already flagged as opening new attack surfaces.
The fix — injecting mathematical noise through a differential privacy framework — has triggered an open dispute: social scientists and other downstream users argue the anonymization lowers dataset quality enough to distort the very statistics the Census exists to produce. The Markup's report frames this as a collision between a legal secrecy obligation and a research constituency.
First-order effects
- The Census Bureau must publish 2020-round tabulations through the noisy differential privacy pipeline, directly affecting demographers, state officials, and redistricting bodies whose figures now carry built-in error the old tables did not have.
Second-order effects
- The commercial-data brokers profiled in the de-anonymization industry are the reason raw Census-style tabulations are untenable — their ability to unmask pseudonymized datasets means any statistical agency publishing exact counts hands them matching keys for free.
Third-order effects
- If the pattern holds, statistical agencies converge on a tiered model: noisy public releases for everyone, with vetted researchers getting cleaner access under controlled conditions — trading open-by-default publication for permissioned access as the price of protecting individual records.
The trend: Government statistics are moving from exact-count publication toward formally private, noise-injected releases, with researcher access increasingly gated behind verification rather than granted by default.