Skip to content

AI behind closed doors: a primer on internal deployment governance

AI Behind Closed Doors: a Primer on The Governance of Internal Deployment

arXiv· arXiv:2504.12170· Charlotte Stix, Matteo Pistillo, Girish Sastry, Marius Hobbhahn, Alejandro Ortega, Mikita Balesni, Annika Hallensleben, Nix Goldowsky-Dill, Lee Sharkey·Indexed by this site · 21 days ago· Originally publishedGovernance
AI summary

An arXiv report notes that the most advanced future AI systems will first be deployed internally within the frontier AI companies that develop them, but governance for internal deployment remains lacking. The report reviews the concept of internal deployment, experience from other industries and the applicability of existing legal frameworks, and gives examples of scenarios of concern. It focuses on two types of risk: loss of control from misaligned AI being used in the AI R&D pipeline, and unconstrained and hard-to-detect concentration of power in closed-door settings. The report concludes with a small number of targeted recommendations as a preliminary blueprint for internal deployment governance.

Read originalarxiv.org