2024年AI安全指数报告英文版

FUTURE OF LIFE INSTITUTEIntroductionRapidly improving Al capabilities have increased interest in how companies report,assess and attempt tomitigate associated risks.The Future of Life Institute (FLI)therefore facilitated the Al Safety Index,a tooldesigned to evaluate and compare safety practices among leading Al companies.At the heart of the Index isan independent review panel,including some of the world's foremost Al experts.Reviewers were tasked withgrading companies'safety policies on the basis of a comprehensive evidence base collected by FLI.The indexaims to incentivize responsible Al development by promoting transparency,highlighting commendable efforts,and identifying areas of concem.ScorecardFirmOverallScoreRiskCurrentSafetyExistentialGovernanceTransparencyGradeAssessmentHamsFrameworksSafety StrategyAccountabilityCommunicationAnthropic2.13C+B-C+GoogleDeepMind1.55OpenAl1.32Zhipu Al1.11DX.A/0.75Meta0.65FGrading:Uses the US GPA system for grade boundaries:A+,A,AB+,[-]Flettervalues corresponding tonumerical values 4.3,40,3.7 3.3,[10.Key FindingsLarge risk management disparities:While some companies have established initial safety frameworks orconducted some serious risk assessment efforts,others have yet to take even the most basic precautions.Jailbreaks:All the flagship models were found to be vulnerable to adversarial attacks.Control-Problem:Despite their explicit ambitions to develop artificial general intelligence(AGI),capableof rivaling or exceeding human intelligence,the review panel deemed the current strategies of allcompanies inadequate for ensuring that these systems remain safe and under human control.External oversight:Reviewers consistently highlighted how companies were unable to resist profit-drivenincentives to cut corners on safety in the absence of independent oversight.While Anthropic's currentand OpenAl's initial governance structures were highlighted as promising,experts called for third-partyvalidation of risk assessment and safety framework compliance across all companies.