AI Policy Wiki
Dashboard

FLI AI Safety Index Winter 2025

high confidence · updated 2026-06-06

Future of Life Institute's Winter 2025 scorecard of 8 frontier AI labs across 6 safety domains — no company scored above C+; 'racing toward AGI/superintelligence without presenting explicit plans.'

The FLI AI Safety Index Winter 2025 is a scorecard published by the Future of Life Institute that graded eight frontier AI companies across six safety domains. Evidence collection closed on November 8, 2025. No company scored above C+, and the review panel concluded that the firms assessed were "racing toward AGI/superintelligence without presenting explicit plans" to control such systems.

Publisher: Future of Life Institute Published: Winter 2025 (evidence collection closed November 8, 2025)

Rankings

CompanyGradeScore
AnthropicC+2.67
OpenAIC+2.31
Google DeepMindC2.08
xAI(D range)
Z.ai (Zhipu AI)(D range)
MetaD1.10
DeepSeekD1.02
Alibaba CloudD–0.98

No company scored above C+. Anthropic (2.67) and OpenAI (2.31) led on a C+ grade; Google DeepMind received a C (2.08); xAI and Z.ai (Zhipu AI) fell in the D range; and Meta (D, 1.10), DeepSeek (D, 1.02), and Alibaba Cloud (D–, 0.98) ranked lowest.

Evaluation domains

The index scored companies across six domains: Risk Assessment, Current Harms, Safety Frameworks, Existential Safety, Governance and Accountability, and Information Sharing. The panel identified Existential Safety as the industry's weakest domain.

Methodology

The assessment drew on a panel of 8 independent AI experts applying 35 indicators to public materials and company surveys. Five of the eight labs participated in the survey component — Anthropic, OpenAI, Google DeepMind, xAI, and Z.ai — while Meta, DeepSeek, and Alibaba opted out.

Panel conclusion

The review panel summarized its finding as follows:

"All companies reviewed are racing toward AGI/superintelligence without presenting explicit plans" to control such systems.

The panel framed this conclusion as applying to the industry as a whole rather than to any single company, and treated existential safety as a first-class evaluation domain alongside more conventional categories such as current harms and risk assessment. The index graded the lab safety frameworks documented in Anthropic's Responsible Scaling Policy (Version 3.1), OpenAI Preparedness Framework V.2, and Frontier Compliance Framework (February 2026), and sits alongside the industry snapshot in Stanford HAI AI Index Report 2026.

Relationships

Sources