⚡ Quick Summary
Published by the Future of Life Institute (FLI), this report presents the Summer 2025 FLI AI Safety Index, in which distinguished experts assess the safety practices of leading AI companies. Its stated purpose is to foster transparency, promote robust safety practices, identify improvements, and help the public distinguish genuine measures from empty commitments. The index evaluates Anthropic, OpenAI, Google DeepMind, x.AI, Meta, Zhipu AI, and Deepseek.
The panel assessed 34 indicators of responsible conduct across six domains: Risk Assessment, Current Harms, Safety Frameworks, Existential Safety, Governance & Accountability, and Information Sharing. It assigned US-GPA-style letter grades and numerical overall scores. The executive summary reports modest progress in some areas but a striking lack of commitment to safety, especially control-loss risk. It concludes that no assessed company has robust, reliable plans to keep the public safe from its products, that current strategies are inadequate to ensure advanced systems remain under human control, and that external validation and legally binding oversight are needed.
🧩 What’s Covered
The supplied summary covers the following elements in its printed order:
- Scorecard and grading: A comparative table gives each of the seven companies an overall letter grade and numerical score, alongside grades for the six domains. Anthropic receives the highest overall score shown, 2.64 and C+, while Zhipu AI and Deepseek receive F grades, with scores of 0.62 and 0.37 respectively. The document explains that grade boundaries use the US GPA system, from A+ (4.3) to F (0).
- Risk Assessment: This domain contains six indicators. The table shows substantial variation: Anthropic is graded C+, OpenAI C, Google DeepMind C-, Meta D, and x.AI, Zhipu AI, and Deepseek F.
- Current Harms and Safety Frameworks: Eight indicators cover Current Harms, where the strongest grade displayed is Anthropic's B-. Six indicators cover Safety Frameworks; Anthropic and OpenAI receive C, Google DeepMind, x.AI, and Meta receive D+, and Zhipu AI and Deepseek receive F.
- Existential Safety and human control: Five indicators address Existential Safety. The executive summary characterises the control problem as unsolved and says the panel found every company's current strategy inadequate for ensuring that systems remain safe and under human control.
- Governance, accountability, and information sharing: The final domains contain five and four indicators respectively. The table records Anthropic's A- in both areas, while grades for other companies vary. The summary argues that companies have abandoned previous voluntary commitments and calls for third-party validation and legally binding oversight with enforcement mechanisms.
- Context, method, and panel: The document describes competitive pressure as a possible incentive for a race to the bottom. It says an independent expert panel used public materials and a tailored industry survey, contributing grades, brief justifications, and recommendations. It also names the six review-panel members and briefly describes FLI's nonprofit mission.
💡 Why it matters?
The index offers a structured comparison for people assessing whether public AI-safety commitments are supported by risk-management practices, safety frameworks, governance arrangements, and information sharing. Its use of 34 indicators and company-level grades makes differences across the named firms visible rather than treating safety claims as interchangeable.
For governance and policy work, the summary directly frames voluntary commitments as insufficient where they are abandoned or cannot be independently validated. Its recommendation for third-party validation of risk assessments and legally binding oversight identifies the accountability mechanisms it considers necessary, while its focus on control loss highlights the gap it finds in plans for keeping increasingly capable systems under human control.
❓ What’s Missing
This two-page summary does not provide the full list of 34 indicators, the company-specific evidence collected, or the brief justifications and improvement recommendations that the methodology says were produced by the panel. It also does not explain the calculation that converts domain assessments into the displayed numerical overall scores beyond identifying the US GPA grade boundaries. Although it says firms could use a tailored industry survey to increase transparency, it does not reproduce the survey, state which firms responded, or show responses. The document directs readers to a separate full report for the indicators and evidence, so the supplied PDF alone cannot support detailed examination of an individual grade.
👥 Best For
AI governance leads, safety researchers, policy teams, and assurance professionals who need a high-level comparison of the named GPAI companies' publicly evidenced safety practices. It is particularly useful for identifying which of the six assessed domains require closer review and for considering the report's case for independent validation of risk assessments.
📄 Source Details
FLI AI Safety Index: Summer 2025 is a two-page English-language publication issued by the Future of Life Institute in July 2025. It names Dylan Hadfield-Menell, Jessica Newman, David Krueger, Tegan Maharaj, Sneha Revanur, and Stuart Russell as members of the Independent Review Panel. The printed link to the separate full report is futureoflife.org/index-s25. The supplied input contains the complete two-page PDF.
