Major artificial intelligence developers are falling short on concrete safety protocols, independent oversight, and long-term risk management, according to the Winter 2025 AI Safety Index released by the nonprofit Future of Life Institute (FLI). The evaluation examines eight leading AI companies across 35 safety indicators spanning six domains, including risk-assessment practices, information-sharing protocols, whistleblowing protections, and support for independent safety research.
Winter 2025 AI Safety Index evaluation
Sabina Nong, an AI safety investigator at the Future of Life Institute, stated during the San Diego Alignment Workshop that the findings expose a distinct divide in industry approaches. The assessment establishes two operational clusters: a higher tier led by Anthropic, OpenAI, and Google DeepMind in descending order, and a lower tier comprising five other firms. Anthropic secured the highest overall position with a C+ grade, whereas lower-tier operators received D and F marks.
Industry tier divide and specific firm evaluations
The lower tier of evaluated organizations includes xAI and Meta, alongside Chinese developers Z.ai, DeepSeek, and Alibaba Cloud. While Chinese models have gained widespread adoption in Silicon Valley due to rapid capability growth and open-source availability, the index ranked Alibaba Cloud lowest with a D- grade. The evaluation panel noted that DeepSeek, which recently released a cutting-edge model matching Gemini 3 capabilities across several benchmarks, ranked second-to-last overall.
According to the report, DeepSeek does not publish safety-minded evaluation frameworks or disclose formal whistleblowing policies. FLI President Max Tegmark noted that second-tier companies have focused heavily on matching the technical frontier but failed to establish equivalent safety standards. Tegmark argued that the abundance of low marks stems from an absence of comprehensive AI legislation comparable to established food-safety regulations.
Methodology and institutional response
The index was graded by an independent panel of eight AI experts, including Massachusetts Institute of Technology professor Dylan Hadfield-Menell and Chinese Academy of Sciences professor Yi Zeng. The report recommends that developers increase internal transparency, adopt independent safety evaluators, implement safeguards against systemic harm, and curtail intensive lobbying efforts.
Representing industry perspectives, an OpenAI spokesperson emphasized that safety remains central to their development process, noting substantial investments in frontier research, internal testing, and independent expert evaluations. Similarly, a Google representative highlighted their Frontier Safety Framework, noting that protocols exist to identify and mitigate severe risks prior to model deployment as technological capabilities continue to advance rapidly.