AI Companies Can't Regulate Themselves. They Should Regulate Each Other (www.lawfaremedia.org)

Recent discussions in the AI community highlight a significant concern: the inability of AI companies to self-regulate effectively due to competitive pressures, which often compromise model safety. Notably, Anthropic recently withdrew its safety guarantee for new releases, citing the rapid advancement of competitors as a key factor. Similarly, OpenAI shortened its safety testing phases, illustrating a troubling trend where prioritizing speed over safety could elevate AI risks. As a solution, experts propose the establishment of an industry self-regulatory organization (SRO) akin to those in finance, which could mitigate risks through binding rules endorsed by government oversight. The proposed SRO for AI aims to address four critical challenges: the race to the bottom on safety; information asymmetry regarding AI training and safety evaluation; the rapid pace of AI advancements outstripping existing legal frameworks; and the need for pre-emptive measures against irreversible AI harms. The model would draw on successful financial regulation structures, requiring mandatory participation from major AI companies and facilitating a coordinated approach to safety standards. This would not only encourage investment in safety without competitive disadvantage but also streamline updates to safety protocols to keep pace with technological developments, thus potentially safeguarding against catastrophic risks associated with advanced AI systems.