The parameter is a policy

Somewhere in a configuration file there is a number that decides how much of one person’s exposure buys how much of everyone else’s accuracy, and it was set once, by whoever was closest to the code. The received view is that this is tuning. Parameters are engineering detail, chosen by the people who understand the mathematics, adjusted when the output looks wrong, and not the sort of thing a governance conversation has any purchase on. For most parameters that is exactly right. For one class it is a category error, and the cleanest instance of that class is a Greek letter.

On June 9, 2021, the United States Census Bureau’s Data Stewardship Executive Policy Committee announced the settings for the system protecting the 2020 census redistricting file, and the headline setting was a privacy-loss budget of epsilon equal to 19.61, split as 17.14 for the persons file and 2.47 for housing units. The committee had raised the figure after data users objected that the earlier demonstration products distorted counts for small places, tribal areas, and race and ethnicity statistics. Alabama had already sued in March to stop the method, and a three-judge court denied its request for a preliminary injunction on June 29 before the case was dropped that September. A number in a statistical pipeline had become a federal lawsuit, and nobody involved was confused about why.

Epsilon is unusual only in making the trade explicit. In the formulation Cynthia Dwork, Frank McSherry, Kobbi Nissim and Adam Smith published in 2006, epsilon bounds how much any single person’s record can change what the released statistics say, and the bound is purchased with noise. A smaller epsilon means more noise, a stronger guarantee for the individual, and worse numbers for everyone who uses the results to draw a district or fund a school. Whoever picks the value is therefore ruling on a trade between two parties’ interests, both real and neither of them in the room, and the ruling arrives as a decimal rather than as a sentence somebody could disagree with.

That is the shape to look for elsewhere, because the same structure is everywhere and almost never labeled. A retention window is a ruling about whose past stays recoverable. A classification threshold is a ruling about how many false accusations buy how many catches, issued by whoever set the cutoff. A decay constant decides how long a judgment about a person keeps its force. Each is chosen once, numerically, and then governs everyone the system touches, and each looks from inside the codebase like a knob. Our corpus already has a position on constraints that nothing enforces and on the scope a claim quietly leaves out. This is a third thing and it is the opposite failure. The decision was genuinely made and is genuinely binding, and it was never written anywhere a person subject to it could read it.

The remedy has already been drafted by the people closest to the mathematics. Dwork, with Nitin Kohli and Deirdre Mulligan, argued in 2019 under a title that is itself the whole instruction, expose your epsilons, for a public registry of the values that real deployments actually use, on the grounds that a guarantee whose parameter is secret is not a guarantee anyone outside can evaluate. Generalize it past differential privacy and it becomes an ordinary disclosure test. For every number in a system that trades one group’s interest against another’s, name who gains, who pays, who chose, and where the choice is recorded. Most parameters fail the test by being boring, which is the point, because the handful that survive it were governing people all along, and a team that can name them out loud has turned an unreviewed default back into a decision it can be held to.