This article assembles Aristotle’s concepts into a four-pillar accountability framework for government AI systems, with an implementation checklist you can adapt into policy or procurement language.
The four pillars #
Pillar 1: Knowledge — make ignorance inexcusable #
Grounding: voluntary action and culpable ignorance (Book III; see Who Answers for the Algorithm?)
Responsibility attaches to what officials knew or should have known. The design goal is to make “we didn’t know” impossible to say honestly: mandatory system documentation, tested failure modes, performance monitoring disaggregated by population, and decision logs that record what the system recommended and what the human did.
Pillar 2: Named agency — attach responsibility to persons #
Grounding: praise and blame attach to choosers (Book III)
Every consequential system has one named accountable official who certifies deployment, owns incident response, and answers to oversight. Committees advise; a person signs.
Pillar 3: Redress — rectify wrongs #
Grounding: corrective justice and equity (Book V; see Aristotle on Justice)
Affected persons receive notice that an automated system was involved, plain-language reasons, a route of appeal to a human with authority to decide differently, and remedy with restoration when the system erred — on timelines fit for the stakes.
Pillar 4: Review — habituate institutional learning #
Grounding: virtue through repeated practice (Book II; see Virtue Is a Practice)
Accountability that fires once, at launch, is a ceremony. Post-deployment audits on a fixed cadence, blameless incident reviews, and tracked corrective actions are how an institution habituates good judgment.
Implementation checklist #
- System inventory: every AI system affecting rights, benefits, or enforcement is registered and public.
- Named accountable official recorded for each system before deployment.
- Documentation standard met: purpose, data provenance, known limitations, tested failure modes.
- Disaggregated performance monitoring live before launch, not promised after.
- Decision logging: recommendation, final decision, and deciding official retained for audit.
- Citizen notice and plain-language explanation for every consequential automated decision.
- Appeal to a human with override authority, with published response deadlines.
- Remedy procedures defined in advance, including correction of downstream records.
- Scheduled independent review with published findings.
- Incident review practice with tracked changes — evidence the institution learns.
How to use this article #
The checklist doubles as an audit instrument for existing systems and a requirements list for procurement. Pair it with Model Policy Language for draft clause text.
