To Be, Or Not To Be?: Regulating Impossible AI in the United States

TL;DR

This paper examines the challenges of regulating 'Impossible AI' systems in the U.S., emphasizing their fundamental limitations and advocating for a functionality-first regulatory approach.

Contribution

It provides an integrated analysis of Impossible AI, tracking specific examples and proposing a new advocacy focus on system validity and deployment decisions.

Findings

Impossible AI systems often cannot perform claimed tasks.

Current regulations lack focus on fundamental system impossibility.

A functionality-first approach can better guide regulation and advocacy.

Abstract

Many AI systems are deployed even when they do not work. Some AI will simply never be able to perform the task it claims to perform. We call such systems Impossible AI. This paper seeks to provide an integrated introduction to Impossible AI in the United States and guide advocates, both technical and policy, to push forward regulation of Impossible AI in the U.S. The paper tracks three examples of Impossible AI through their development, deployment, criticism, and government regulation (or lack thereof). We combine this with an analysis of the fundamental barriers in the way of current calls for Impossible AI regulation and then offer areas and directions in which to focus advocacy. In particular, we advance a functionality-first approach that centers the fundamental impossibility of these systems and caution against criti-hype. This work is part of a broader shift in the community to focus on validity challenges to AI, the decision not to deploy technical systems, and connecting technical work with advocacy.