Literature Record: Frontier AI Regulation
What It Establishes
The paper defines frontier AI — highly capable foundation models that could possess dangerous capabilities sufficient to pose severe risks to public safety — and argues these pose a distinct regulatory challenge because the dangerous capabilities are hard to assess before deployment and because development is concentrated in a small number of firms.
Its contribution is a regulatory instrument set rather than a single proposal: standard-setting, registration and reporting, licensing or permitting for training runs above thresholds, third-party evaluation, and liability allocation. Twenty-four authors from across the field's institutions is itself a signal about where the policy consensus sat in 2023.
Rating Justification
- Textual Authority (B): retrieved at abstract level; preprint, not peer-reviewed.
- Corroboration (4): the instrument set it proposes is substantially the instrument set that was subsequently enacted — the EU AI Act's systemic-risk obligations (LIT-012) are recognisably this design. Policy uptake is independent corroboration.
What It Does Not Show
It does not settle which instrument binds. It is a menu, and the vault's proposal picks from a different menu — private law and cryptographic custody rather than licensing and evaluation. The paper therefore does not refute SYS-003; it demonstrates that the design space SYS-003 claims to occupy was already populated when the proposal was written.
Relation to the Vault
This is the document that makes gap 5 concrete. SYS-003 designs a legal-architectural regime — privilege, custody, node sovereignty — without engaging the regulatory architecture that was being built at the same time and for the same problem. See CNF-007.