Safeguarding the safeguards: How best to promote AI alignment in the public interest

TL;DR

This paper discusses how public institutions can effectively support AI alignment efforts to reduce risks of accidents and misuse, emphasizing systematic approaches and mitigation of potential counterproductive issues.

Contribution

It identifies four problems that can hinder AI alignment efforts and proposes mitigations, offering a systematic framework for institutions to enhance effectiveness.

Findings

Support for alignment should focus on reducing accident and misuse risks.

Four key problems can undermine alignment efforts and require mitigation.

Systematic planning increases the likelihood of beneficial AI safety outcomes.

Abstract

AI alignment work is important from both a commercial and a safety lens. With this paper, we aim to help actors who support alignment efforts to make these efforts as effective as possible, and to avoid potential adverse effects. We begin by suggesting that institutions that are trying to act in the public interest (such as governments) should aim to support specifically alignment work that reduces accident or misuse risks. We then describe four problems which might cause alignment efforts to be counterproductive, increasing large-scale AI risks. We suggest mitigations for each problem. Finally, we make a broader recommendation that institutions trying to act in the public interest should think systematically about how to make their alignment efforts as effective, and as likely to be beneficial, as possible.