An American Privacy Emergency

Scott’s foreword: Cynthia Dwork is Gordon McKay Professor of Computer Science at Harvard, and a pioneer in the fields of differential privacy and algorithmic fairness. On my recent travels to the SigmaWest science camp and then STOC, there was much talk about a recent Trump administration action that would ban not only differential privacy, but essentially all modern techniques for preserving privacy in large datasets, for example in the 2030 US Census. I realize that many of us have “outrage fatigue,” but this particular outrage hits extremely close to home for the CS theory community. So when Cynthia approached me at STOC to propose a guest post on the issue, of course I said yes. The post that she sent me, below, is cosigned by many other leaders in the field.

On June 4, 2026, the U.S. Secretary of Commerce issued a directive (DAO 216-26) relegating confidentiality protection in all Bureau of Economic Analysis (BEA) and U.S. Census Bureau publications to techniques dating back to the early 1970s, turning its back on over half a century of progress and protections for data subjects. Advances in confidentiality provision had enabled the Census Bureau to share increasing quantities of data at more granular detail. The order will result in less useful (or fewer available) statistics, weaker protection, or both. We write to illustrate the danger posed by the order and to mobilize the scientific community to speak out against it.

The acting force behind this order is political interest, not scientific merit. DAO 216-26 bypassed legally required administrative procedures. It fulfills a promise made by the architects of the Heritage Foundation’s Project 2025, and reflects both the rhetoric and misunderstandings of representatives of the Center for Renewing America (CRA), an organization founded by OMB Director Russell Vought. CRA’s explainer on the use of differential privacy in the 2020 Census is up-front about the stakes: “Even if the citizenship question is added to the Census, it will be impossible to ascertain the status of individuals so long as differential privacy is used.” But masking this sort of personal characteristics data is legally required by the Census Act (13 U.S. Code Section 9), which makes it a crime to “make any publication whereby the data furnished by any particular [individual] can be identified.” Confidentiality is also widely understood as critical to ensuring that people respond to the census.

DAO-216-26 bans differential privacy and other modern (and not so modern) techniques. It restricts disclosure avoidance techniques to “coarsening,” which it describes as “reducing the level of detail or specificity of published statistics, such as through rounding, aggregating (grouping), and/or the use of ranges.” “Suppression” (“expressly redacting certain values”) may also be used, but only as a “last resort.” DAO-216-26 forbids “noise infusion”, described as “methods that involve modifying a dataset by adding random values, or noise.”

Noise infusion was invented precisely to address the increasing demand for granular data in the face of confidentiality laws that forbid publishing reidentifiable data. Coarsening and suppression were satisfactory for most national, aggregate statistical series, like the Principal Federal Economic Indicators. However, these techniques failed when applied to business and demographic data at fine geographic or industrial detail. By forbidding noise infusion, the directive bans the disclosure avoidance techniques at the core of dozens of data releases over the last three decades. It bans input noise infusion, used in the Quarterly Workforce Indicators since 2002 and, until now, planned for the Bureau of Economic Analysis statistics [1]. It bans swapping, used for decennial census publications since 1990. It also bans differential privacy, the best currently known approach for obtaining the most data utility for any given level of privacy. Differential privacy was used for sharing data on commuting patterns (OnTheMap) since 2008 and for publications based on the 2020 Census. Until the recent directive, differential privacy was planned for the 2030 Census too. Many other products and procedures are implicated as well.

1. Illustrations

DAO-216-26 is incompatible with the Census Bureau’s dual mandate to provide confidentiality and fitness for use. To illustrate this, we recall and expand on an example due to Nathan Goldschlag, inspired by the County Business Patterns (CBP) data, which provides statistics on business activity broken down by industry and geography. Goldschlag describes three scenarios, illustrating the tension between providing useful information and maintaining confidentiality of responses as required by the Census Act.

· “There is only one brewery in a small county. If the CBP published the exact count of brewery employees in that county, it would be disclosing the information of one business (how many workers it employs), a clear violation of the law.

2· “There are two breweries in a small county, and the CBP again publishes the exact count of brewery employees. If I own one of those breweries, I could learn how many employees my competitor has, again violating the law.

· “There are more than two breweries in a small county, but the CBP chooses not to publish the total number of brewery employees out of concern that it might compromise the privacy of the businesses. If I’m a prospective brewery owner, I may deem the project too risky to pursue without information about the market I’m entering.”

In Goldschlag’s example, coarsening makes the published statistics useless. We now add a fourth scenario, showing that it also fails to maintain confidentiality. To keep things simple, assume none of us owns any of the businesses in the new example. The County has two towns with one brewery each, North Bend and South Bend. Furthermore, North Bend has a mobile bottling company and South Bend has a stationary bottling company. That’s a total of four beer-related business entities in the County. Two of these businesses, the North-Bend brewery and the South Bend bottling company, are publicly-owned.

• The CBP publishes five statistics:

• (A) The total number of employees in beer-related businesses in North Bend: Because there is only one brewing company in North Bend and only one bottling company in North Bend, the category is coarsened to “beer-related”.

• (B) The total number of employees in beer-related businesses in South Bend: Because there is only one brewing company in South Bend and only one bottling company in South Bend, the category is coarsened to “beer-related”.

• (C) The total number of employees in brewing only: Because there is only one brewing company in each of North Bend and South Bend, the statistic is coarsened to the total number of employees in brewing only in the County.

• (D) The total number of employees in bottling only: Because there is only one bottling company in each of North Bend and South Bend, the statistic is coarsened to the total number of employees in bottling only in the County.

• (E) The total number of employees at publicly owned companies: Because there is only one publicly owned company in each of North Bend and South Bend, the statistic is coarsened to the total number of employees in publicly owned companies in the County.

• (A) The total number of employees in beer-related businesses in North Bend: Because there is only one brewing company in North Bend and only one bottling company in North Bend, the category is coarsened to “beer-related”.

We now have 5 equations in 4 unknowns. Using only 4 of these (A, B, C, and E), we can solve for the exact number of employees at each of the four companies with high school algebra.

In the above (fictional but realistic) scenario, the County Business Patterns were released with good-faith coarsenings for the geographical, business, and ownership categories. Nonetheless, even without inside kn