Two Camps in AI Governance -- Separation of powers is a diagram until a refusal can stop a bind

Abstract AI governance increasingly contains two distinct architectural bets. One seeks to make the artificial system itself reliably good through constitutions, alignment, training, values, and internal guardrails. The other seeks to make the institution around consequential action real by separating rule-setting, proposal generation, oversight, and the power to refuse a binding. Most practical systems may combine both approaches, but they answer different questions. This paper examines the second question: when an autonomous or agentic system is about to create an external, durable, or accountable consequence, where does authority live, can it still refuse, and can an independent party reconstruct why? The paper maps Gadi Singer's separation-of-powers framing for autonomous agents and Evelyne-Claudia Yantony's work on consequential reachability and correction onto the EFA Execution Boundary. It identifies four recurring institutional failure modes: a probabilistic guardian mistaken for independent authority; policy collapsed into the actor's own reasoning; oversight without causal reach; and authority defined at the wrong temporal or cumulative unit. It then describes a deliberately small mechanical boundary that separates ALLOW from BIND, preserves local-policy variability through Non-Formation under BLOCK, and requires evidence capable of independent replay. The paper also states costs and residual failure surfaces, distinguishes statistical evaluation of probabilistic controls from deterministic replay of a bounded gate, and proposes a falsifiable test protocol. The claim is intentionally narrow: separation of powers becomes operational only when a valid refusal can prevent the governed state transition and the evidence can show that it did. Keywords: AI governance; agentic AI; execution boundary; separation of powers; corrective reach; EFA; standing; authorization; BIND; NO BIND; Non-Formation; independent replay; Refuse Pack.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-10-31
DOI
https://doi.org/10.5281/zenodo.23839765
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

Two Camps in AI Governance -- Separation of powers is a diagram until a refusal can stop a bind

Charles Rupp
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

Two Camps in AI Governance -- Separation of powers is a diagram until a refusal can stop a bind

Charles Rupp
preprint en

Abstract

Abstract AI governance increasingly contains two distinct architectural bets. One seeks to make the artificial system itself reliably good through constitutions, alignment, training, values, and internal guardrails. The other seeks to make the institution around consequential action real by separating rule-setting, proposal generation, oversight, and the power to refuse a binding. Most practical systems may combine both approaches, but they answer different questions. This paper examines the second question: when an autonomous or agentic system is about to create an external, durable, or accountable consequence, where does authority live, can it still refuse, and can an independent party reconstruct why? The paper maps Gadi Singer's separation-of-powers framing for autonomous agents and Evelyne-Claudia Yantony's work on consequential reachability and correction onto the EFA Execution Boundary. It identifies four recurring institutional failure modes: a probabilistic guardian mistaken for independent authority; policy collapsed into the actor's own reasoning; oversight without causal reach; and authority defined at the wrong temporal or cumulative unit. It then describes a deliberately small mechanical boundary that separates ALLOW from BIND, preserves local-policy variability through Non-Formation under BLOCK, and requires evidence capable of independent replay. The paper also states costs and residual failure surfaces, distinguishes statistical evaluation of probabilistic controls from deterministic replay of a bounded gate, and proposes a falsifiable test protocol. The claim is intentionally narrow: separation of powers becomes operational only when a valid refusal can prevent the governed state transition and the evidence can show that it did. Keywords: AI governance; agentic AI; execution boundary; separation of powers; corrective reach; EFA; standing; authorization; BIND; NO BIND; Non-Formation; independent replay; Refuse Pack.

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon