[BODY]
As the world's major artificial intelligence (AI) companies accelerate their competition over AI performance, the safety commitments they set for themselves are instead weakening, according to an assessment by international experts. With AI rapidly spreading into cyberattacks and the military domain, they warn that corporate self-regulation alone is insufficient to control the risks.
The Future of Life Institute (FLI), a US-based nonprofit think tank focused on AI safety, released its "2026 First-Half AI Safety Index" on the 7th (local time), evaluating the safety practices of nine major AI companies worldwide. The companies assessed included Anthropic, OpenAI, Google DeepMind, Meta, Z.ai, Alibaba Cloud, xAI, DeepSeek, and Mistral. A panel of independent experts evaluated these companies across six categories: risk assessment, current harms, safety frameworks, existential safety for humanity, governance and accountability, and information disclosure and communication.
The results showed Anthropic ranked first on the safety index, but its actual grade was only C+. OpenAI and Google DeepMind each received a C, while Meta earned a D+. Z.ai and Alibaba Cloud received a D-, and xAI, DeepSeek, and France's Mistral received an F. None of the major big tech firms leading the global AI market received an A or B in the safety assessment.
The FLI expressed concern, saying, "Companies are backing away from the 'red line' commitments they previously made." Anthropic, OpenAI, Google DeepMind, and Meta had earlier presented safety policies pledging to halt development or restrict releases on their own if they approached certain risk levels, but these companies have recently weakened or abandoned those commitments internally. The point is that while AI capabilities are rapidly growing, the mechanisms to halt or control them have instead loosened.
In fact, in February, Anthropic withdrew its previous pledge not to train AI systems unless it could guarantee in advance that its safety measures were sufficient. The assessors who participated in compiling the FLI safety index argued that "this move needs to be reversed." Although Anthropic ranked first in this assessment, the FLI concluded that "even the top-ranked companies' current safety measures are completely inadequate relative to the pace of AI capability advancement."
The FLI also pointed out that "in recent years, major AI companies have softened their opposition to military applications." Citing cases linked to real-world harm—such as AI providing information needed for mass shootings or abetting suicide—it assessed that "AI risks are no longer a distant future hypothesis but are shifting into present-day social and security problems." Max Tegmark, co-founder and president of the FLI, noted, "Companies themselves warn of the dangers of superintelligent AI, yet they do not stop competing to build it," adding that "voluntary commitments alone are not enough."
Meanwhile, the FLI is a US-based nonprofit think tank in the AI safety field, founded in 2014 by MIT physics professor Max Tegmark, Skype co-founder Jaan Tallinn, and UC Santa Cruz professor Anthony Aguirre, among others. It aims to reduce the large-scale risks that advanced AI could pose to humanity and to promote related policy discussions. In 2023, it led an open letter calling for at least a six-month halt to training AI systems more powerful than GPT-4, sparking global debate. At the time, the letter was signed by Tesla CEO Elon Musk, Apple co-founder Steve Wozniak, University of Montreal professor Yoshua Bengio, and Yuval Harari, among others.
