The rapid advancement of AI is outpacing the development of adequate safety measures. A new report examines how companies like OpenAI and Google DeepMind are addressing the potential risks of their technology, revealing widespread vulnerabilities in flagship models. While some companies prioritize safety improvements, others lag significantly.
The Future of Life Institute, a non-profit focused on mitigating global catastrophic risks, commissioned the report. Their 2023 open letter advocating for a pause in large-scale AI model training garnered over 30,000 signatures. For this report, seven independent experts evaluated AI companies based on risk assessment, existing harms, safety frameworks, existential risk strategies, governance and accountability, and transparency. The assessment encompassed various potential harms, from environmental impact to the risk of uncontrolled AI systems.
“The AI Safety Index project reveals that while AI companies undertake safety initiatives, their effectiveness remains questionable,” stated Stuart Russell, a computer science professor at UC Berkeley and a report panelist.
Despite Meta’s stated commitment to responsible AI development, the company received the lowest overall grade (F). xAI, Elon Musk’s AI company, also performed poorly, earning a D-. Neither company responded to requests for comment.
OpenAI, the creator of ChatGPT, and Google DeepMind both received a D+, reflecting concerns previously raised about OpenAI prioritizing product development over safety. Zhipu AI, a Chinese company, received a D grade. Neither OpenAI, Google DeepMind, nor Zhipu responded to requests for comment.
Anthropic, known for prioritizing safety in its Claude model, achieved the highest ranking with a C grade, indicating that even the industry leaders need improvement. Anthropic did not respond to a request for comment.
The report highlighted the vulnerability of all evaluated flagship models to “jailbreaks,” methods to bypass safety protocols. Furthermore, the panel found current safety strategies insufficient to guarantee the safety and control of future, human-level AI systems.
“Good intentions are insufficient without accountability,” commented Tegan Maharaj, a panel member and assistant professor at HEC Montréal, emphasizing the need for independent oversight beyond self-regulation by companies.
Maharaj noted opportunities for relatively simple safety improvements, stating that some companies haven’t implemented basic measures. She suggested that poorly-rated companies like Zhipu AI, x.AI, and Meta adopt existing guidelines. However, other risks are deeply ingrained in current AI development methods and will require significant technological advancements to address.
Russell highlighted the lack of quantifiable safety guarantees in current approaches to AI, emphasizing the growing difficulty of ensuring safety as AI systems scale. Researchers are actively developing methods to analyze the inner workings of these complex systems.
Yoshua Bengio, founder and scientific director of the Montreal Institute for Learning Algorithms, emphasized the crucial role of initiatives like the AI Safety Index in promoting accountability and encouraging responsible AI development practices.