The non-profit organization Future of Life Institute (FLI) recently released the "AI Safety Index (Winter 2025)," which provides an independent assessment of the safety practices of eight leading global AI companies: Anthropic, OpenAI, Google DeepMind, xAI, Z.ai, Meta, DeepSeek, and Alibaba Cloud.

This assessment covers six major domains—Risk Assessment, Current Harms, Safety Frameworks, Existential Safety, Governance & Accountability, and Information Sharing—comprising a total of 35 indicators. The report results show that Anthropic, OpenAI, and Google DeepMind occupy the top three positions. However, not a single company achieved a grade of B or above; the highest-rated company only received a C+, indicating that the industry's overall safety practices significantly lag behind the pace of its capability advancement.

The report points out that a clear divide persists between the top performers (Anthropic, OpenAI, and Google DeepMind) and the other companies reviewed (Z.ai, xAI, Meta, Alibaba Cloud, DeepSeek). The most substantial gaps exist in the domains of risk assessment, safety frameworks, and information sharing, caused by limited disclosure, insufficient evidence of systematic safety processes, and uneven adoption of robust evaluation practices.

Despite public commitments from various companies, their practices still fall short of emerging global standards. While many companies partially align with frameworks such as the EU's Code of Practice on General-Purpose AI, the depth, specificity, and quality of implementation vary significantly. Current safety practices have yet to meet the rigor, measurability, or transparency levels envisioned by these frameworks.

In terms of Risk Assessment, Anthropic, OpenAI, and Google DeepMind continue to lead. The report notes that although more companies have begun conducting internal and external evaluations of frontier AI risks, the overall depth and independence remain inadequate. Specific manifestations include: a narrow scope of risk assessments, such as failing to cover emerging areas like climate risks; no company has conducted "Human Uplift Trials" (used to measure whether AI enhances users' ability to carry out harmful actions); no company has quantified the probability of extreme scenarios like AGI loss of control; and the independence of external reviews remains insufficient.

Regarding Current Harms, Anthropic scored highest in this dimension, while xAI performed the weakest. Overall, companies consistently scored poorly in this domain, with frequent safety failures, weak robustness, and inadequate control of serious harms being common issues. Performance in trustworthiness benchmarks such as truthfulness, fairness, and harmful-content generation was uniformly low. Furthermore, the practice of all models training on user interaction data by default reflects inadequate enforcement of privacy principles by companies. Notably, Chinese companies performed relatively better in content watermarking due to policy requirements.

For Safety Frameworks, Anthropic, Google DeepMind, Meta, OpenAI, and xAI have published safety frameworks. Among them, Anthropic, Google DeepMind, and OpenAI offer the most structured approaches, outlining risk areas, qualitative thresholds, and mitigation measures. On the other hand, DeepSeek, Z.ai, and Alibaba Cloud have not yet released any form of safety framework.

In the domain of Existential Safety, this was the weakest-performing area for all companies. Despite active advancement in AGI and superintelligence research and development by various firms, none have presented credible plans to prevent potential large-scale misuse or loss of control risks. The report states that these most consequential risks remain effectively unaddressed, representing one of the industry's most severe weaknesses.

Overall, AI technology is rapidly approaching higher levels of general intelligence, while the industry's preparedness in safety governance is woefully inadequate. Only by synchronously elevating safety governance standards and regulatory requirements on a global scale can risks be kept manageable amidst accelerating technological development.

Author：Qinger