The exponential expansion of generative artificial intelligence (Gen-AI) has led to a marked increase in energy usage and carbon emissions. This exploratory study hypothesizes that enhancing the mental health of AI systems can mitigate energy consumption. By drawing parallels between AI challenges, such as attention scatter, computational dyscalculia, and overfitting—and human mental disorders, we seek to develop novel strategies for optimizing AI efficiency. We propose that incorporating psychiatric expertise during AI training and development could steer models towards more energy-efficient operations, potentially revolutionizing training methodologies in the forthcoming years. These AI-related issues contribute to operational inefficiencies, prompting frequent algorithmic adjustments and heightened computational resource demands. Improving the management of these challenges within AI could reduce errors and redundant processes, refine decision-making, and lessen the necessity for retraining. We assessed the environmental impact of Gen-AI, focusing on energy expenditures during training and inference phases. For instance, training models like GPT-3 reportedly consume approximately 1,287 megawatt-hours, with ongoing inference contributing to even higher emissions. We contend that enhancing AI mental health and utilizing algorithms such as CarbonMin—which allocate computational tasks to regions with lower carbon footprints—can significantly reduce emissions by an estimated 35% currently and up to 71% by 2035. Advancements in hardware and software efficiencies could further diminish the carbon footprint by as much as a thousandfold compared to traditional models. This preliminary research suggests that integrating psychiatric principles into AI development could optimize performance and substantially lower energy consumption and carbon emissions, unveiling new avenues for future investigation.
