The Institutional Mechanics of Open-Weight AI Oversight: A Companion to "Thresholds, Not Timetables"
Year: 2026
Type: preprint
Abstract: This companion white paper provides a plain-language account of the structural argument behind a pre-registered, evidence-based protocol on the oversight of open-weight AI in the European Union. The paper explains what the protocol expects, why those expectations are held, how they will be scored, and what would count as evidence against them. It reports no outcome data and does not recommend any oversight mechanism, assess any statute, or forecast dates. The central question addressed is whether an authority can still inspect, interrupt, or modify a running open-weight AI system once its weights have been downloaded and deployed on an organization's own infrastructure. The paper distinguishes between two deployment architectures—closed, API-based services where the provider retains decision rights, and open-weight deployments where decision rights pass to the adopting organization upon release. The protocol records this shift directly through decision-rights indicators and tests whether these predict later oversight difficulty better than classical market measures such as concentration or price. The paper presents seven pre-registered predictions with stated confidence levels, including: direct oversight of deployed open-weight models will not be in place within 36 months (90% confidence); oversight through infrastructure intermediaries will not yet reach deployed models within 36 months (70% confidence); and the claim that decision-rights indicators outperform classical measures carries most of the evidential risk (registered at 40%, below even odds). The protocol fixes the claims, scoring rules, and data collection date in advance so that results cannot be shaped after the fact. The framework assesses institutional correction capacity across five dimensions—resource availability, decision authority, specialist capability, viable alternatives, and time to act—without combining them into a single score. It identifies the binding constraint as the dimension that most limits whether correction can be carried out in time, arguing that resources cannot substitute for authority. Two oversight scenarios are fixed in advance: direct oversight (Scenario S) and oversight through infrastructure intermediaries (Scenario Sβ), each with distinct expected binding constraints and registered expectations. Keywords: Open-weight AI, AI governance, Oversight mechanisms, Pre-registration, Decision rights, Infrastructure-mediated oversight, European Union AI policy, Institutional capacity, AI auditing Technology policy
Author: Katz, E.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225620747
Cites:
Cited by:
Related to: