OpenAI to let outside groups vet AI models earlier

OpenAI announced Tuesday that it will let outside organizations conduct technical safety assessments of its AI models during earlier phases of training, evaluation, and rollout, a move aimed at addressing mounting concerns about the potential harms of frontier AI systems.

The ChatGPT maker outlined priorities for these reviews to function effectively, including "strong independence mechanisms, scientific rigor, robust security practices and clear responsibilities," according to Bloomberg. The announcement, made via a company blog post, comes during a period of intense scrutiny of the AI industry's safety practices.bloomberg

A Wave of Safety Proposals

The move follows a rapid succession of safety-related announcements from the leading AI labs. On Monday, OpenAI published a separate set of proposals focused on alignment research and a computing technique known as recursive self-improvement, or RSI, calling for international cooperation to develop frontier standards. "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely," the company wrote.cnbc

Last week, rival Anthropic published its own proposals for safer development of frontier models, after researcher Jacob Coxon ignited a global debate by resigning from the company and warning that AI labs were "gambling with our lives". Anthropic CEO Dario Amodei subsequently called for AI companies to slow their pace of development and proposed embedding third-party evaluators inside their organizations.explainx+1

Trust Gap Remains

Despite the flurry of pledges, skeptics question whether meaningful oversight is possible amid the industry's breakneck pace. More than $7 trillion is expected to be spent scaling AI over the next five years, according to Goldman Sachs ↘ 1.38% estimates cited by Axios. Both OpenAI and Anthropic face pressure to deliver increasingly capable models as they prepare for public offerings at multi-trillion-dollar valuations.axios

A coalition of more than 100 AI evaluators recently urged frontier model makers to adopt "minimum conditions" for independent audits, including deeper access and protections against retaliation for publishing unfavorable findings. Daniel Kokotajlo, a former OpenAI researcher and executive director of the AI Futures Project, told Axios he does not believe AI can be safely scaled, arguing that top companies should avoid directing their vast computing resources toward model improvement.cnbc+2

"The companies are just having this continuous process of conflicted feelings and internal discussion about like, what are we doing," Kokotajlo said. "Should we stop? Are we the good guys?"axios