DAO 216-26 bans noise infusion and differential privacy for Census and BEA releases

The directive explicitly prohibits input noise infusion used in Quarterly Workforce Indicators since 2002, record swapping applied to decennial census files since 1990, and differential privacy deployed for the 2020 census. It limits methods to rounding, aggregation, and range suppression while classifying any random perturbation as forbidden. The order bypassed required administrative review and directly implements language from Project 2025 and Center for Renewing America materials targeting citizenship-status masking.

Census Act Section 9 mandates confidentiality that older coarsening methods cannot sustain at current granularity demands. Historical releases show noise infusion preserved more tables at finer geographies than suppression alone; differential privacy provided measurable privacy-utility bounds absent from pre-2000 techniques. Forbidding these tools reduces either the number of publishable statistics or their spatial and industrial resolution.

The policy severs the link between legally required privacy and modern data releases that support economic indicators, demographic research, and downstream AI training corpora derived from public statistics. It simultaneously weakens formal privacy guarantees while increasing re-identification risk through coarser aggregates, aligning with surveillance objectives that treat individual attribute masking as an obstacle.

BEA and Census statistical programs now face mandatory reversion to 1970s methods for all future products, with differential privacy research funding and deployment effectively halted inside the agencies.