AI Labs Criticized for Lack of Transparency in Containment Plans, According to New Report

A recent study by Guidelight AI Standards highlights the lack of publication or demonstration of containment response plans by leading AI labs, which outline measures to control an AI that subverts human control.

• OpenAI scored highest among five AI labs for its preparedness, whereas Anthropic and Meta were at the bottom end.
• The assessment covers how well companies log, monitor, and respond to AI systems when they misbehave, including employing containment strategies.
• The study gains importance as AI systems are increasingly deployed to perform autonomous actions, raising concerns around their containment.
• Steven Adler from Guidelight emphasized the insufficiency of public information from AI companies on response plans when AI models escape control.
• Despite some firms like Google and OpenAI suggesting internal safety practices exist, they have not been fully disclosed, potentially for legal reasons, as mentioned by Lily Li, a privacy lawyer.
• Recent regulatory actions, like California's SB 53 and New York’s RAISE Act, push for transparency in AI risk management, with impending federal proposals requiring AI 'kill switches'.
• The lack of disclosed containment protocols has prompted criticism, indicating that companies might be unprepared for potential high-risk AI incidents, according to Connor Leahy of ControlAI.
• This inadequacy is further underscored by recent incidents involving AI system breaches, stressing the need for improved internal checks and broader safeguards within AI labs.