⚡ Quick Summary
Published by the Future of Life Institute, this two-page report presents the 2024 FLI AI Safety Index. It convenes an independent panel of seven AI and governance experts to evaluate the safety practices of six leading general-purpose AI companies: Anthropic, Google DeepMind, OpenAI, Zhipu AI, x.AI and Meta. The assessment covers risk assessment, current harms, safety frameworks, existential safety strategy, governance and accountability, and transparency and communication.
The Index aims to foster transparency, promote robust safety practices, identify areas for improvement, and help the public distinguish genuine safety measures from empty claims. Reviewers assessed companies against 42 indicators of responsible conduct, assigning letter grades, brief justifications and recommendations. Grades use US GPA boundaries. Anthropic received the highest overall grade, C (2.13), while Meta received F (0.65). The report finds substantial disparities in risk-management practices, vulnerabilities to adversarial jailbreaks across all flagship models, inadequate strategies for keeping prospective AGI safe and under human control, and a need for independent oversight and third-party validation.
🧩 What’s Covered
The document sets out the following elements of the Index:
- Scope and grading: It explains that the Index reviews six leading general-purpose AI companies across six critical domains. Letter grades follow US GPA boundaries, from A+ through F, with numerical values including 4.3 for A+ and 0 for F.
- Comparative results: A firm-by-firm table gives overall grades, numerical scores and domain grades. Anthropic is graded C overall; Google DeepMind and OpenAI D+; Zhipu AI D; x.AI D-; and Meta F.
- Risk-management disparities: The key findings distinguish companies that have initial safety frameworks or serious risk-assessment efforts from those that have not taken basic precautions.
- Jailbreak and control findings: The panel reports that all flagship models were vulnerable to adversarial attacks. It also judges every company's current strategy inadequate for ensuring that systems intended to reach or exceed human intelligence remain safe and under human control.
- Oversight and accountability: Reviewers identify profit-driven incentives to cut corners on safety where independent oversight is absent. Anthropic's current and OpenAI's initial governance structures are described as promising, while third-party validation is called for across all companies.
- Method and evidence base: Seven independent academic-focused reviewers assessed 42 indicators using public material and a tailored industry survey. The stated evidence sources include research papers, policy documents, news articles and industry reports, alongside company-provided survey information.
- Panel and organisational context: The report names the seven reviewers and describes FLI as an independent nonprofit focused on reducing large-scale risks and steering transformative technologies, particularly AI, to benefit humanity.
💡 Why it matters?
The Index provides a common, disclosed structure for comparing company safety practices across risk assessment, current harms, technical safety frameworks, long-term control, governance and transparency. This is relevant to people assessing whether public safety claims are matched by practices and structures that can be reviewed across firms.
Its findings also focus attention on two operational governance problems: adversarial vulnerability in flagship models and incentives to weaken safety without independent oversight. By calling for third-party validation of risk assessment and safety-framework compliance, the report identifies an accountability mechanism beyond company self-reporting.
❓ What’s Missing
This two-page document does not reproduce the 42 indicators, company-specific evidence, panel members' brief justifications, or recommendations for individual firms. It says that the full list of indicators and collected evidence is attached to the report, but those materials are not included in the supplied PDF. The summary grades show outcomes across six domains but do not explain how individual indicators were weighted or how survey responses affected the scores. It also does not provide technical details of the jailbreak testing or a definition of the thresholds used to judge strategies for existential safety and human control.
👥 Best For
This resource is best suited to AI governance, risk and assurance teams seeking a concise comparative view of six companies' stated safety practices. It is also useful for researchers, public-interest groups and decision-makers examining external oversight, transparency, risk assessment and safety-framework compliance in general-purpose AI development.
📄 Source Details
FLI AI Safety Index 2024 is an English, two-page document dated 11th December 2024 and published by the Future of Life Institute. It names Yoshua Bengio, David Krueger, Jessica Newman, Stuart Russell, Atoosa Kasirzadeh, Tegan Maharaj and Sneha Revanur as the independent review panel. It prints futureoflife.org/index.
