Leading AI companies fail on Safety Index

In FLI’s latest Safety Report, Google DeepMind and OpenAI got a D+, Zhipu AI got a D, while x.AI got a D-. Meta failed, while Anthropic was the only survivor with a C.

An independent panel of seven AI and governance specialists assessed the safety practices of these six AI companies across six parameters – Risk Assessment, Current Harms, Safety Frameworks, Existential Safety Strategy, Governance and Accountability, and Transparency and Communication.

Findings revealed that only a few companies have established adequate safety frameworks and done risk assessments. All flagship models are vulnerable to attacks, and all artificial general intelligence (AGI) models are far from safe or under human control. Companies often cut corners on safety parameters to succumb to profiteering.

FLI was founded by MIT cosmologist Max Tegmark, Skype co-founder Jaan Tallinn, and others in March 2014.

Anthropic’s existing structure and OpenAI’s initial governance structure were deemed promising.