As AI systems continue to advance, researchers and policymakers are concerned about ensuring their safe and ethical use. The main issues center around the potential risks posed by ever-evolving and increasingly powerful AI systems. These risks involve potential misuse, ethical issues, and unexpected consequences stemming from AI’s expanding abilities. Several strategies are being explored by policymakers to minimize these risks, but accurately predicting and controlling any potential harm caused by AI systems as they develop is a significant challenge.
Commonly, governance strategies involve defining thresholds for the computational power used when training AI models. The idea is that higher computational power correlates with greater risk, and so by identifying and controlling AI systems exceeding certain levels of intensity, these can be regulated. Policies such as the White House Executive Orders on AI Safety and the EU AI Act incorporate these thresholds.
However, researchers from Cohere for AI critique the use of these computational thresholds as a means of governance. They argue that current implementations are inadequate and ineffective at risk minimization. While acknowledging the uncertain and fast-changing relationship between computation and risk, they push for a more detailed method to AI governance, considering various factors that influence AI’s risk profile.
Rather than relying on fixed computational thresholds, the researchers advocate for a comprehensive, dynamic evaluation of AI systems. This multi-faceted approach involves better defining metrics, considering additional risk dimensions and AI performance factors, and implementing adaptive thresholds that adjust to AI’s evolving abilities. There is also an emphasis on improving transparency and standardization in reporting AI risks and aligning governance practices with AI’s actual performance and potential harm. This method identifies factors such as the quality of the training data and specific applications of AI models to guarantee a more precise risk assessment.
Their research underlines how fixed computational thresholds often overlook significant risks connected with smaller, highly optimized AI models. Many current policies overlook the rapid advancements and optimization strategies that enable smaller AI models to have the same capacity and risk level as larger ones, for instance, models with less than 13 billion parameters outperform larger models with over 176 billion parameters. As a result, existing computational thresholds prove to be unreliable indicators of AI risks and need a significant overhaul to be effective.
In conclusion, the research suggests that computational thresholds, as they currently stand, are insufficient as a governance tool for AI systems. The fluctuating relationship between computation and risk is unpredictable, necessitating a more informed approach to regulation. Proposed solutions involve shifting towards dynamic thresholds and multi-faceted risk evaluations which are more capable of anticipating and mitigating the risks posed by advanced AI systems. Ultimately, policies must be adaptable and accurately reflect the complexities of AI development.
