On Monday, the Commerce Department issued Departmental Administrative Order 216-26, a document whose title is as dry as its consequences are explosive. The order bans “noise infusion” — the practice of adding small, random perturbations to published statistics — from all statistical products released by the Census Bureau and the Bureau of Economic Analysis. Any use of noise infusion, the order states flatly, “is inconsistent with the Department’s policies.”

The framing in the accompanying press push was straightforward: the Trump administration is cutting bureaucratic artifice, letting Americans see the real numbers, ending a practice that muddied the statistical waters in the name of privacy theater. Who could object to cleaner data?

Quite a few people, as it turns out. But not for the reasons you might expect.

The Liability Shield Nobody Talks About

The public conversation about disclosure avoidance has been captured by the privacy-versus-transparency frame, and that frame is wrong. Noise infusion was never primarily about protecting citizens from prying eyes. It was about protecting the Census Bureau from its own statute.

Title 13 of the U.S. Code makes it a felony for any Census employee to disclose individually identifiable information. Penalties include up to five years in prison and a $250,000 fine. This is not a guideline. This is not a vibes-based policy preference. This is a law with teeth, and every career attorney in the Commerce Department’s Office of General Counsel knows exactly how sharp those teeth are.

The differential privacy framework that the Bureau adopted for the 2020 Census — and the noise infusion techniques the Bureau of Economic Analysis has been rolling out for its international accounts data — existed for one overriding purpose: to let the lawyers say yes. Without a mathematically demonstrable disclosure-avoidance system, the Bureau’s own legal review process would have to default to suppression. Cell suppression. Row suppression. Table suppression. The wholesale withholding of data that, in a world without noise, could theoretically be reverse-engineered to identify a respondent.

One statistician at a federal statistical agency — reached by Slack DM on Tuesday, and not authorized to speak publicly — put it bluntly: “The noise was never about privacy. It was about plausible deniability for the release process.”

What the Bureau’s Own Research Actually Found

The Census Bureau spent years evaluating disclosure risk after researchers demonstrated that older methods — principally data swapping, in which values are exchanged between records — could be defeated by modern database reconstruction attacks. The Bureau’s own 2021 decision memo on differential privacy noted that swapping was designed for an era when you had to physically visit a data center to run queries. Today, an attacker with public Census tables and a commodity GPU can reconstruct individual-level records with alarming precision.

Noise infusion wasn’t an academic indulgence. It was the Bureau’s answer to a demonstrated vulnerability — one that, if exploited, would expose the agency to exactly the kind of Title 13 liability no general counsel wants to explain to a congressional oversight committee.

DAO 216-26 bans the answer without addressing the vulnerability. It tells the Bureau: you may no longer add noise to protect confidentiality, but you remain legally obligated to protect confidentiality. The contradiction is not subtle.

The Practical Outcome: Less Data, Not Cleaner Data

Here is what the order does not do: it does not repeal Title 13. It does not indemnify Census employees against prosecution. It does not provide an alternative disclosure-avoidance method that satisfies the Bureau’s statutory obligations.

Here is what it likely does: it forces the Bureau’s disclosure-review officers to fall back on the bluntest instruments available — suppressing entire cells, collapsing geographic categories, withholding tables that can’t be released safely. The result will not be a glorious new era of unsullied statistics. The result will be holes. Black boxes where small-county data used to be. Coarsened categories that make granular analysis impossible. The kind of data degradation that businesses, local planners, and academic researchers have been dreading since the privacy debate first heated up in 2018.

A small-city planning director in Ohio, asked about the order at a regional data-users meeting this week, reportedly said: “I don’t need pure data. I need any data.”

The irony is that the constituencies most likely to applaud this order — transparency advocates, open-data activists, businesses that rely on Census products for market research — are the ones most likely to be burned by it. Privacy advocates, meanwhile, will find no comfort here either: suppressing data doesn’t protect anyone; it just makes everyone equally ignorant.

The Hard Conversation Nobody Wants to Have

The underlying problem is not noise infusion. The underlying problem is that the United States has never had an honest public conversation about the tension at the heart of its statistical system. Title 13 demands absolute confidentiality. The data users demand granular precision. Those two demands are in direct conflict, and for decades the Bureau has navigated that conflict through a series of increasingly elaborate technical compromises. Noise infusion was the latest — and, by the Bureau’s own technical assessment, the most defensible — compromise yet devised.

Banning the compromise does not resolve the conflict. It simply removes the Bureau’s ability to manage it. What remains is a legal framework that punishes disclosure, a policy framework that demands release, and no mechanism to bridge the gap.

Someone will eventually have to decide: either give the Bureau the legal cover to release data without noise, or accept that large portions of the statistical products Americans rely on are going dark. Pretending that DAO 216-26 solves anything is a comforting fiction. The Bureau’s lawyers do not have the luxury of believing it.

Sources