According to two new studies released on Thursday, major artificial intelligence firms worldwide exhibit “unacceptable” risk management practices and a “notable absence of dedication to numerous safety domains.”
Even current AI systems pose risks, as acknowledged by several leading companies themselves, such as potentially assisting malicious actors in conducting cyberattacks or developing bioweapons. Leading scientists are concerned that advanced AI models in the future might become entirely uncontrollable by humans.
These investigations were conducted by the nonprofit organizations SaferAI and the . Each represented the , as part of what these groups intend to be an ongoing series designed to encourage leading AI companies to enhance their operational methods.
Max Tegmark, president of the FLI, stated, “Our aim is to simplify the process for individuals to discern which entities are genuinely implementing their promises, rather than merely verbalizing them.”
SaferAI evaluated the risk management protocols (also known as responsible scaling policies) of leading AI companies to assign scores based on each company’s approach to identifying and alleviating AI risks.
In SaferAI’s evaluation of risk management maturity, no AI firm achieved a rating higher than “weak.” Anthropic led with the highest score at 35%, with OpenAI following at 33%, Meta at 22%, and Google DeepMind at 20%. Elon Musk’s xAI received an 18% score.
Anthropic and Google DeepMind both saw their scores decline compared to the initial study conducted in October 2024. Consequently, OpenAI has surpassed Google, now ranking second in SaferAI’s assessments.
Siméon Campos, founder of SaferAI, indicated that Google’s score was relatively low despite its commendable safety research, primarily due to the company’s limited concrete policy commitments. Furthermore, the company launched its frontier model, Gemini 2.5, earlier this year without disclosing safety details, which Campos described as a “flagrant oversight.” A Google DeepMind representative informed TIME: “Our commitment is to develop AI in a safe and secure manner for societal benefit. AI safety encompasses a broad array of potential risk reduction strategies. These latest reports do not fully capture all of Google DeepMind’s initiatives in AI safety, nor all prevailing industry standards. Our holistic strategy for AI safety and security significantly surpasses what is reflected.”
Anthropic also experienced a drop in its score since SaferAI’s assessment in October. This reduction was partially attributed to modifications the company made to its just days prior to the launch of Claude 4 models, wherein Anthropic removed pledges to address insider threats by the time models of that sophistication were released. Campos commented, “That constitutes a very poor process.” Anthropic did not provide an immediate response when asked for comment.
The authors of the study further noted that their methodology had been refined and made more comprehensive since the previous October, which explains some of the variations in the scores.
The companies demonstrating the most significant improvement in their scores were xAI, which advanced to 18% from 0% in October, and Meta, which increased to 22% from its prior 14% score.
The FLI’s research took a wider scope, examining not just risk management procedures, but also the companies’ strategies regarding existing harms, existential safety, governance frameworks, and data dissemination. A committee of six impartial experts assessed each company, basing their judgments on publicly accessible information like policies, scholarly articles, and media reports, alongside supplementary confidential data companies had the option to submit. Anthropic received the highest mark (a C plus). OpenAI was awarded a C, while Google obtained a C minus. (Both xAI and Meta received a D.)
Nevertheless, concerning FLI’s evaluations of each company’s commitment to “existential safety,” all companies received a score of D or lower. Tegmark commented, “They are all claiming a desire to construct superintelligent entities capable of surpassing human intellect in all aspects, yet they lack any discernible strategy for managing these creations.”