by Christopher Steven Marcum, Meghan Maury, and Beth Jarosz

Late last week, the Department of Commerce quietly issued a sweeping new policy that could reshape how the Census Bureau and the Bureau of Economic Analysis (BEA) protect the privacy of people and businesses whose information they collect. 

The policy bars the agencies from using “noise infusion,” a family of privacy-protection techniques that make small, controlled changes to published data so that no individual person, household, or business can be reidentified from publicly accessible datasets. One such technique that has garnered some controversy in the past is known as “differential privacy.” That may sound technical. But the stakes are simple: federal statistical agencies are required to do two things at once. They must protect confidential information, and they must publish useful public data. By taking away one of the tools agencies use to do both, the likely result is not “better data.” It may be less data, or less useful data.

What Is Noise Infusion and Differential Privacy?

To understand this framework, it helps to think of public data as a portrait. A high-quality portrait does not explicitly highlight every single pore, scar, or private detail on a person’s face. Instead, it provides enough detail to clearly recognize and understand the subject, while softening the fine textures that do not need to be exposed. Public data works the same way. It gives us a picture of our communities and economy: where people live, where businesses are growing, where jobs are located, where resources are needed, and how conditions differ across places. However, that portrait should not be so sharp that it reveals confidential information about a particular person, household, or business.

Noise infusion is one way to balance protections with precision. Instead of publishing raw numbers exactly as collected, the agency makes small, controlled adjustments before releasing the data. The goal is to preserve the larger picture while making it much harder to trace a published number back to someone’s private information. 

Differential privacy is one especially strong version of this approach. It gives agencies a formal way to measure and limit how much any one person’s information can affect a published result. That matters more than ever in a world where public datasets can be combined with commercial data, administrative records, and powerful AI tools.

Noise infusion is not the same as “error.” It is privacy protection. It allows agencies to publish data that remain useful for understanding broad patterns, while reducing the risk that confidential information can be reconstructed. In the 2020 Decennial Census, this noise was injected in clearly-documented ways that preserved certain important characteristics without change (like state population totals used for apportionment and the number of housing units in each neighborhood) but blurred some details just enough that it would be nearly impossible to reidentify individuals. While this can complicate research on small populations, it ensured accurate counts for apportionment and ensured that public officials still received accurate data for policymaking. There’s a great cartoon explanation of these tradeoffs by Josh Neufeld here. 

This Is Bigger Than the Decennial Census

Most public debate about these techniques has focused on the 2020 Census and the redistricting data. Researchers and communities debated the tradeoffs between privacy and accuracy for specific characteristics. The Census Bureau adjusted their methods in response to public feedback. That debate matters. Decennial census data are used for redistricting, civil rights enforcement, funding formulas, emergency response, local planning, and countless public and private decisions. You can read more about Census’s history of use of differential privacy on the Internet Archive’s Wayback Machine capture of their page here.

But, the new Commerce policy is much broader than the Decennial census.

The Census Bureau has long used noise infusion in important economic data products. For example, the Quarterly Workforce Indicators, one of the flagship products from the Longitudinal Employer-Household Dynamics program, uses noise infusion to publish detailed information about employment, job creation, hiring, separations, earnings, industries, and local labor markets. These data help state and local officials, workforce boards, researchers, businesses, and community organizations understand how jobs are changing over time and across places. Furthermore, as recently as April 2026, BEA announced it was planning on adopting noise infusion for its economic survey products. Now, Commerce wants to end the use of the technique for all of those purposes. 

Less Privacy Means Less Data

Without techniques like noise infusion, agencies may have to fall back on cruder tools. They can publish data only at higher levels of aggregation. They can suppress individual cells. Or they can withhold entire products. In portrait terms, they can blur the whole image, black out parts of it, or take the portrait down.

None of those options serve the public good with better quality data. Instead, the options provide less data.

While researchers are often primary data users, that loss would be felt far beyond academic research. Detailed public data help local governments plan services, businesses understand markets, workforce boards identify labor trends, advocates document inequities, journalists hold institutions accountable, and communities make the case for resources. When those data disappear or become less detailed, the people and communities with the least access to private data are usually the ones who lose out the most.

Data Quality Concerns Are Real. A Ban Is Still the Wrong Answer.

To be clear, concerns about data quality are real. Some researchers raised legitimate concerns about how differential privacy affected certain 2020 Census products, especially for small geographies and small population groups. Those concerns deserve to be taken seriously. But the answer is not to prohibit an entire category of privacy tools. The answer is to improve the methods, test them publicly, document their effects, and help users understand when a data product is reliable for a particular use.

That was already the more sensible path and part of existing information policy. The Office of Management and Budget’s 2025 open data guidance did not ban privacy-protective methods. In footnote 89, the guidance said that when agencies use techniques like noise injection or differential privacy, they must be transparent. Agencies should explain the limits of the data, provide methodological information, describe the uncertainty introduced by the privacy method, make clear that the changes were made to protect privacy and confidentiality, and explain how qualified users can access raw data in secure settings. 

This is a reasonable approach that acknowledges community concerns with noise infusion while allowing statistical agencies to continue to ensure the confidentiality, accuracy, and objectivity of its statistical activities and to disseminate timely and relevant statistical products they publish using the technique. Commerce’s new policy, on the other hand, moves in the opposite direction. It imposes a blanket rule from above, rather than allowing statistical agencies to make product-specific decisions based on evidence, law, technical expertise, and public input.

That is especially troubling because federal law already gives statistical agencies strong responsibilities. They must produce timely, relevant, accurate, and objective statistics. They must protect confidentiality. And they must be able to make statistical decisions free from inappropriate interference. These responsibilities were formalized into the “Trust regulation” issued by OMB in 2024.

What Should Happen Now

A policy this consequential should not be implemented behind closed doors. And, it should not have been released without public comment. 

Commerce, Census, and BEA should publish an implementation plan in the Federal Register and request information from the public before making changes to any data products. They should explain which products will be affected, which methods will still be allowed, how much data may be suppressed or coarsened, and how users can comment before changes take effect. 

Public data depends on public trust. That trust requires privacy and confidentiality. It also requires access and transparency. A blanket ban on noise infusion puts both at risk.