Open problem 1
The Governable Object Problem
- current AI
- frontier AI
- transformative AI
- AGI
Should governance target models, systems, deployments, compute, data, companies, capabilities, behaviours, supply chains, or socio-technical arrangements?
Why it’s foundational
Almost every governance proposal smuggles in an answer. A use-case regulator governs deployment contexts. A frontier regulator governs model developers. Compute governance targets infrastructure. Liability law targets harms. These produce different enforcement points and different blind spots.
Why it’s difficult
AI risk does not sit neatly in one object. A base model may be harmless alone but dangerous when scaffolded with tools, agents, memory, APIs, fine-tuning, or user workflows. Conversely, regulating only deployment ignores training-time risks, internal lab use, model theft, and downstream diffusion.
Hidden assumptions
The common hidden assumption is that “an AI system” is a stable regulatory unit. It often is not. The real object may be a shifting stack: model, weights, tools, prompts, data, deployment interface, organisational incentives, cloud provider, and user context.
Competing positions
- Use-case regulation
- Model-level regulation
- General-purpose/foundation-model regulation
- Compute governance
- Organisation-level licensing
- Post-harm liability
- Sectoral regulation
- Public-infrastructure governance
What could make progress
Detailed causal maps of AI incidents; comparative studies of whether harms originate in model properties, deployment choices, user misuse, or organisational failures; audit pilots comparing model-level versus system-level assurance; technical work on tracing model provenance and capability transfer.
What it would change
It would determine whether regulators need AI agencies, sectoral supervisors, compute registries, auditor accreditation, lab licensing, cloud-provider duties, or all of these.
Sub-agenda
- When do model-level properties predict deployment-level risk?
- Which AI harms are best governed upstream versus downstream?
- Can “frontier model” be defined without creating loopholes or overreach?
- How do tool use, agents, and fine-tuning change the regulatory object?
- What is the minimum viable object for enforceable international rules?
Priority (editor scoring)
If the object is wrong, the regime fails regardless of good intentions.
- Importance
- 5/5
- Neglected
- 4/5
- Difficulty
- 5/5
- Actionable
- 5/5
- Robust
- 5/5
- Nat’l+int’l
- 5/5
Where the catalog bears on this
No current catalog instrument resolves this puzzle — which is the point: it is a foundational question the existing rules leave open. Browse the coverage catalog for what the instruments do and don’t say.
Editorial content — a human-authored agenda question, rendered verbatim. No part of this analysis is AI-generated (see the charter).