R-02 · 7 min

Why AI Needs Machine-Readable Policy

Abstract

Capable models with no authority layer are not governed systems. They are fast improvisation with an institution's name attached.

Learning objectives

01Explain why prose policy fails at machine speed and volume.
02Separate model capability from institutional permission.
03Describe the risk profile of unrecorded interpretation at scale.

Concept

The current pattern for governing AI systems is to describe the rules in prose and place that prose in a prompt, a policy handbook, or a review meeting. Each of those places the interpretive burden back on something that cannot be held accountable for it.

Prompted rules are not rules

A rule pasted into a context window is an input to a probabilistic process. It can be outweighed, paraphrased, partially applied, or silently dropped when the context is truncated. It produces different outcomes on different days. An institution that cannot reproduce its own decisions has not delegated authority — it has abandoned it.

Three failures that follow

Non-reproducibility: the same case reaches two different outcomes and neither can be defended.
Non-attribution: when a decision is challenged, there is no artifact to point to that says why the answer was what it was.
Non-revocability: when a rule changes, there is no reliable way to know which past decisions were made under the old one.

The separation that fixes it

Machine-readable policy separates the model's job from the institution's job. The model handles language, extraction, drafting and interpretation of unstructured input. The authority artifact decides what is permitted. The model proposes; the artifact disposes. Everything the institution is accountable for lives in the artifact, where it can be reviewed, versioned and verified.

model output  →  candidate action
authority artifact  →  permit | deny | escalate
runtime  →  decision record + proof

This is why machine-readable policy is not an AI feature. It is the precondition for letting AI operate inside an institution at all.

Framework

The four pressures

01Speed — decisions outpace human review
02Volume — sampling replaces inspection
03Opacity — reasoning is generated, not recorded
04Drift — the model changes, the policy did not

Case study

A model reading regulation

An assistant is asked whether a system falls under a high-risk category. It produces a fluent, well-cited answer. Asked again a week later, with a newer model, it produces a different one. Neither answer is bound to a version of the regulation, and neither can be replayed.

The failure is not the model's accuracy. It is that no artifact ever held the institution's position, so there was nothing for the model to execute — only text to interpret.

Exercise

Audit one AI-assisted decision path for unrecorded interpretation.

01Trace the path from input to outcome.
02Mark every step where a human or model interprets a rule.
03For each, ask what artifact would have made the step deterministic.
04Estimate how many such decisions occur per month.

Artifact

execution.trace

policy: eu_ai_act@2024.06.13
input: system.profile#4471
evaluated:
  annex_iii.point_5b -> true
  article_6.exemption -> false
outcome: HIGH_RISK
verifiable_by: [operator, notified_body, regulator]