The Quiet Politics of File Formats

The infrastructures that produce our most consequential disputes are rarely the ones we argue about. A model card is contested; a JSON schema is not. A risk register is debated in the press; the file format it lives in moves through every stage of the policymaking process untouched. This is not an accident.

For three years I have been collecting, with a colleague, the small documents that surround machine‑learning systems in the public sector — procurement attachments, vendor questionnaires, the JSON appendices to risk frameworks. The pattern is consistent: where policy demands ambiguity, the schema demands choice. The schema wins.¹

1. The schema as policy

Consider a deceptively boring field: the risk_level enum in a municipal AI procurement form. It admits three values — minimal, limited, high — and one of them is preselected. The default, of course, is limited. Most procurement officers report never having changed it.

The default is not a setting. It is a policy decision deferred until the moment of friction makes it permanent.

We can read this two ways. The charitable reading: defaults are guardrails, useful in volume. The less charitable: defaults are the residue of an argument someone had at a working group, transcribed into permanent shape and shipped to thirty thousand offices.

2. EXIF, or, the camera that wrote the law

EXIF — the metadata appended to every photograph your camera takes — was, for most of its life, a curiosity. Until the day a court admitted it as evidence and the day after that, when an evidentiary standard quietly attached itself to a 1995 specification authored by a Japanese trade group. The trade group was not consulted.

[ figure · schematic ]

This is the quiet politics of file formats: they accrete authority by being adjacent to other things that already have it. A camera ships with a schema; the schema attends a trial; the trial cites the schema; the schema returns to the camera, now slightly more sovereign than it was the day before.

3. A small example

{
  "system_type": "decision_support",
  "risk_level": "limited",     // <- the default
  "human_in_the_loop": true,   // <- always true; nobody checks
  "data_sources": ["internal"],
  "model_card_uri": null
}

Each of these fields is a hinge. human_in_the_loop is the most consequential of them, and the least audited: in a 2024 study of 412 procurement filings across six U.S. cities, the boolean was set to true in every single one.

---

A version of this essay appears in the Spring 2026 issue of an academic journal whose title I will add when the embargo lifts.

Footnotes

1. With apologies to those who have made the same point in other words. The schema-wins-over-policy intuition is at least as old as Suchman (1995) and arguably older.