AI Verification - Mechanisms to Ensure AI Arms Control Compliance

While diverse in form and function, the common thread that unites the constellation of AI technologies is the unique power to grant autonomy and beyond-human insight to a range of systems and processes. [...] Former Deputy Secretary of Defense Paul Nitze summarized these realities by stating the goal of effective verification is to make sure that: [I]f the other side moves beyond the limits of the Treaty in any militarily significant way, we would be able to detect such violation in time to respond effectively and thereby deny the other side the benefit of the violation.2 A practical verification mecha. [...] This view is the root of many of the above challenges and has led many experts to contend that AI verification requires full access to a system’s software to establish the degree of AI control over it or its functions.6 Under this assumption, to concretely establish whether a lethal autonomous drone, for instance, allows AI to control the use of force, an inspector would require full access to all. [...] The use of these chips for AI processing and specifically AI * It is assumed that because the complexity and sensitive design details of hardware is often invisible to the naked eye an inspector could not easily steal or reveal substantial design secrets unless they had the time and resources to analyze the components in depth. [...] To use Van Eck radiation to sustain verification post-inspection, a system’s owner would need to consent to the installation of a theoretical “Van Eck radiation sensor” onto the system in such a way that it can detect the unshielded Van Eck emissions of the device.