DeepMind warns model-level AI governance is insufficient
A new position paper from Google DeepMind and the Centre for the Governance of AI argues that modern AI governance frameworks focusing primarily on base models fail to account for "non-model gains." These gains include inference-time compute scaling, system scaffolds (e.g., agents, external tools), and restricted asset integration, all of which enhance a model's capabilities post-deployment. The authors propose shifting towards broader layers of governance—such as system, entity, agent, and cloud-level controls—complemented by efforts to build overall societal resilience.
Traditional safety checks at the training finish line are obsolete when the real leaps in model capabilities happen downstream in production.
* **The Model-Centric Blindspot:** Pre-deployment evaluations (like red-teaming raw weights) completely miss dynamic runtime enhancements like agentic workflows and test-time compute scaling.
* **Three Pillars of Post-Training Gains:** The paper formalizes "inference gains" (compute at test-time), "systems gains" (scaffolding and tools), and "asset gains" (access to restricted data or APIs) as key vectors rendering static assessments inadequate.
* **Decentralized Governance Mandate:** Policy makers must pivot from policing compute scale/training data to monitoring runtime infrastructure, API access, cloud providers, and downstream agent behaviors.
* **Societal Resilience as a Fail-Safe:** If technical controls fail, the last line of defense shifts to strengthening societal infrastructure against AI-driven threats.
DISCOVERED
1h ago
2026-06-03
PUBLISHED
1h ago
2026-06-03
RELEVANCE
AUTHOR
manaltdan
