A new assessment by the nonprofit Guidelight reveals that no major AI company fully applies basic control measures to its own internal systems. The report, which evaluates companies like Anthropic, OpenAI, Google, xAI, and Meta, found that none meet proposed safety standards. Guidelight's analysis focused on six core practices, including logging internal AI activity and emergency shutdown mechanisms. The findings highlight a critical gap in AI safety practices across the industry.
Anthropic and OpenAI scored the highest with a C+, while Google received a D+ with a detailed roadmap. xAI and Meta performed the worst, with xAI getting a D− and Meta an F. The report emphasizes that companies excel at detecting misbehavior but fall short in prevention and containment strategies. Guidelight's assessment is based on public sources like system cards, safety reports, and blog posts, without direct access to internal data.
Guidelight is an independent nonprofit founded by former OpenAI safety leads Page Hedley and Steven Adler. The organization aims to provide unbiased assessments of AI safety practices. The report underscores the need for stronger internal controls and transparency in AI development. The findings are part of a broader effort to address safety concerns in the rapidly evolving AI landscape.
Source: thedecoder
