Environmental Costs
The computing power used in training, testing, and deploying generative AI systems, especially large scale systems, uses substantial energy resources and thereby contributes to the global climate crisis by emitting greenhouse gasses.
ENTITY
1 - Human
INTENT
2 - Unintentional
TIMING
3 - Other
Risk ID
mit172
Domain lineage
6. Socioeconomic and Environmental
6.6 > Environmental harm
Mitigation strategy
1. Model Optimization and Inference Efficiency: Implement mandatory optimization techniques, such as model compression (e.g., quantization and distillation) and the systematic deployment of smaller, task-specific models in lieu of large, general-purpose models for inference. This is prioritized due to inference accounting for over 90 percent of a model's lifetime operational emissions and the potential for energy reductions up to 90 percent per task. 2. Decarbonization and Location-Aware Computing: Strategically locate and migrate data center workloads to geographic regions powered predominantly by certified renewable or low-carbon electricity grids (e.g., hydroelectric, geothermal, nuclear). Furthermore, employ carbon-aware workload scheduling to align high-demand computing tasks with periods of peak renewable energy availability to minimize the emissions factor per kilowatt-hour. 3. Mandatory Transparency and Full Life-Cycle Assessment (LCA) Reporting: Institute a standardized requirement for all AI model developers to publicly disclose a comprehensive, auditable environmental footprint. This assessment must encompass operational emissions (training, inference, and pre-training steps), water consumption for cooling, and embodied emissions associated with the manufacturing, lifespan, and disposal of specialized hardware.
ADDITIONAL EVIDENCE
While the environmental costs of compute has become an area of active research, with workshops dedicated to the question, the environmental costs of manufacturing hardware remains under-explored. One potential reason for this discrepancy may be that estimating compute and energy costs, while complex, is a comparably transparent task compared to tracing the emissions of the of emissions throughout the manufacturing process. However, recent estimates suggest that the manufacturing process have substantial environmental costs [96]. Overall, information about emissions is scarce and there is no consensus for what constitutes the total carbon footprint of AI systems.