The Model You Can Switch Off: What the Fable 5 Suspension Reveals About Where Governance Can Act
Year: 2026
Type: preprint
Abstract: On June 12, 2026, a frontier AI model was disabled worldwide by government order. This essay — the companion to Access, Not Containment: A Working Vocabulary for Frontier Model Safety (DOI: 10.5281/zenodo.20736864) — asks why the prohibition landed where it did, and what that reveals about where governance can and cannot act. It separates three things the episode fused into one: danger (how harmful a capability could be), governability (whether access can still be conditioned once a model is live), and remediability (whether the affected part can be repaired while the whole keeps running). It locates the measurable content of "danger" in the size of a domain's language — vast and uncheckable in literature, smallest and fully checkable in software — and argues that the directive struck the most checkable, most governable end of that gradient. Governability is shown to reduce to a single question: does a persistent, revisable control point over access exist? For a hosted model, yes; for an openly released model, no — not because open models are more dangerous, but because open release irreversibly forfeits the only lever. The essay closes constructively: the object that belongs between engine and user — conditioning access by counterparty, authorization, and budget — can be written in the engine's own probabilistic language as a measurable interface rather than a blunt switch, and the G7 "trusted partners" scheme is that object being improvised by hand, under pressure, after the fact.
Author: Pignatelli, Paolo
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7165203476
Cites:
Cited by:
Related to: