On June 4, 2026, Commerce Department Disclosure Avoidance order DAO 216 26 stripped federal statistics of the 2020 Census privacy technique. Its co inventor, Cynthia Dwork, warns the 2030 count will be less accurate, less private, or both.
On June 4, 2026, the U.S. Secretary of Commerce signed an order, DAO 216-26, titled "Disclosure Avoidance for Statistical Products," that limits how the Census Bureau and the Bureau of Economic Analysis (BEA) protect the confidentiality of the data they publish. The directive bans a family of statistical techniques called "noise infusion," which includes the method known as differential privacy, and confines both agencies to older tools: aggregation, rounding, and data coarsening.
The order applies to "all" Census Bureau and BEA publications, according to a recent guest post on the computer scientist Scott Aaronson's blog by Cynthia Dwork of Harvard, one of the inventors of differential privacy, cosigned by more than 100 other researchers in differential privacy, statistics, and algorithmic fairness. They call the change "an American privacy emergency" and warn that the result will be "less useful (or fewer available) statistics, weaker protection, or both."
That trilemma is the heart of the dispute. Differential privacy, introduced by Dwork and her co-authors around 2006, works by adding calibrated statistical noise to query results so that no individual record can be reverse-engineered from a published table. It was first deployed at scale in the 2020 Decennial Census under a plan finalized in 2021. Critics on the right, organized around groups associated with the Center for Renewing America, a project of OMB Director Russell Vought, have argued for years that the noise injection distorted legitimate uses of the data, from block-level redistricting to enforcement of the Voting Rights Act. The new order, those critics say, simply returns the agencies to proven methods. The change also fulfills a long-standing policy promise associated with Project 2025, the Dwork group writes.
The researchers' objection is not nostalgia. It is that the older tools are weaker: aggregation, rounding, and coarsening can sometimes be reverse-engineered by linking published tables to other public records, a problem known as a re-identification attack. The Census Bureau's own re-identification experiments in the 2010s, NPR has reported, showed that several published 2010 Census tables could be unmasked at the household level. The 2020 methodology was designed to close that hole. Removing it without replacing it with anything equally rigorous is, the Dwork group argues, an unfilled gap.
The legal ground is also contested. A joint statement cosigned by COPAFS (the Census and Policy Advisory Caucus of the American Statistical Association), the Association of Public Data Users (APDU), the Inter-university Consortium for Political and Social Research (ICPSR), and the Population Association of America (PAA) argues that DAO 216-26 conflicts with the Census Act, the 1954 law codified at 13 U.S.C. § 9 that explicitly forbids the Census Bureau from publishing data that could be used to identify a specific person or business. The order, they write, "removes the only methodology capable of satisfying that statutory requirement at modern scale." It also, the researchers note, was issued without the notice-and-comment rulemaking that statutes typically require for changes of this scope.
The administration has begun to implement it. The Bureau of Economic Analysis has already updated its disclosure-limitation methodology to aggregation, rounding, and coarsening, and a working paper, WP2026-9, has disappeared from BEA's working-paper taxonomy page, a removal consistent with what the Dwork post describes as an in-house purge of differential-privacy research.
The downstream stakes are concrete and well-dated. The 2030 Decennial Census is already in its planning phase; any change to its disclosure-avoidance methodology affects apportionment, redistricting, and the enforcement of voting-rights and civil-rights statutes that depend on block-level accuracy. BEA's National Income and Product Accounts, regional and industry accounts, and international transactions all flow through the same disclosure-limitation pipeline. If those numbers become coarser, regional GDP, state-level personal income, and county-level business-formation data will become less granular, sometimes at the exact geographic resolution the data is now used to track.
The counter-argument deserves equal weight. Proponents of the order, including policy analysts associated with the Center for Renewing America and officials in the Office of Management and Budget, have argued since the 2020 cycle that differential privacy degraded the usability of census data for state and local planners, particularly in rural and small-minority jurisdictions, and that the legal mandate to protect confidentiality was being used to obscure accuracy losses. The order, in their telling, restores transparency and legal compliance.
Both stories can be true at once. Differential privacy does add noise. Older disclosure-limitation methods are also weaker against modern re-identification attacks. The question is whether the administration has offered, or is preparing, a replacement that closes the re-identification gap without restoring the accuracy complaints, or whether the order has simply traded one problem for a larger one. So far, no replacement methodology has been published.
What to watch: the Federal Register publication of the implementing rule; any congressional hearing; the Census Bureau's 2030 operational plan, due in updated form in late 2026 or 2027; and whether WP2026-9 or other removed BEA working papers reappear in archived form. A legal challenge, if filed, will likely turn on 13 U.S.C. § 9 and on whether the order required notice-and-comment rulemaking. Until those answers are public, the privacy math behind every table the federal government releases is a moving target.