Towards Socially and Environmentally Responsible AI

Pengfei Li, Yejia Liu, Jianyi Yang, Shaolei RenarXiv 2024 · OpenReview

Abstract

The sharply increasing sizes of artificial intelligence (AI) models come with significant energy consumption and environmental footprints, which can disproportionately impact certain (often marginalized) regions and hence create environmental inequity concerns. Moreover, concerns with social inequity have also emerged, as AI computing resources may not be equitably distributed across the globe and users from certain disadvantaged regions with severe resource constraints can consistently experience inferior model performance. Importantly, the inequity concerns that encompass both social and environmental dimensions still remain unexplored and have increasingly hindered responsible AI. In this paper, we leverage the spatial flexibility of AI inference workloads and propose equitable geographical load balancing (GLB) to fairly balance AI's regional social and environmental costs. Concretely, to penalize the disproportionately high social and environmental costs for equity, we introduce $L_q$ norms as novel regularization terms into the optimization objective for GLB decisions. Our empirical results based on real-world AI inference traces demonstrate that while the existing GLB algorithms result in disproportionately large social and environmental costs in certain regions, our proposed equitable GLB can fairly balance AI's negative social and environmental costs across all the regions.

Introduction

As large language models like Llama-3 and GPT-4 become ubiquitous, their physical footprints grow in tandem. Currently, AI providers utilize Geographical Load Balancing (GLB) to route user requests to data centers where electricity is cheapest or carbon intensity is lowest. While this "efficiency-first" approach minimizes operational costs, it overlooks two critical dimensions of inequity:

  1. Environmental Inequity: Standard GLB might minimize total carbon but concentrate the environmental burden—such as local water scarcity or air pollution—onto a single, often marginalized, region.
  2. Social Inequity: High-performance models are computationally expensive. Users in resource-constrained regions may be consistently served by smaller, less accurate models to save on costs, widening the global digital divide.

This research asks a fundamental question: How can we route AI workloads so that the negative ecological and social impacts are shared fairly across the globe?

Methodology: SE-GLB and $L_q$ Norms

We propose SE-GLB (Socially and Environmentally equitable GLB), a novel algorithm that redefines the "cost" of AI inference. Instead of merely minimizing energy bills, the optimization objective integrates three pillars:

  • Operational Cost: The direct price of electricity.
  • Social Inequity Cost: Measured as the performance degradation experienced by a user when assigned a smaller model (e.g., Llama-2-7B) instead of the state-of-the-art alternative (e.g., Llama-2-70B).
  • Environmental Inequity Cost: A weighted combination of carbon emissions and freshwater usage.

To achieve equity, we introduce $L_q$ norms (where $q \ge 1$) into the objective function. While standard optimization focuses on averages ($L_1$ norm), increasing the value of $q$ makes the algorithm hypersensitive to "worst-case" scenarios. If one region suffers from disproportionate pollution or poor model accuracy, the $L_q$ norm acts as a heavy penalty, forcing the system to redistribute the workload even if it increases total energy expenditure.

Experimental Setup

The study validates SE-GLB using a simulation of 10 geographically distributed data centers across the US, Europe, and Asia over an 18-day period. The evaluation utilized:

  • GPU Power Traces: Real-world utilization data for AI workloads.
  • Model Benchmarks: Performance scores for Llama-2 (7B, 13B, and 70B) across benchmarks like AI2 Reasoning Challenge and Truthful QA.
  • Environmental Data: Hourly energy prices, weather data (for cooling water loss calculations), and regional carbon intensity.

Key Results

Comparison of social and environmental metrics across different GLB algorithms.

As shown in the table above, SE-GLB significantly outperforms standard baselines in achieving equity across all metrics:

  • Superior Equity: SE-GLB achieved a Max/Avg ratio for normalized performance cost of 1.02, nearly reaching perfect equality (1.00), compared to 1.71 for standard Cost-GLB.
  • Improved Worst-Case Performance: In the most disadvantaged regions, the minimum performance score rose from 54.94 to 57.98, ensuring a more consistent user experience globally.
  • Water Conservation: SE-GLB slashed peak water usage in the most burdened regions by more than half, reducing it from 1410.92 $m^3$ (under Cost-GLB) to 649.41 $m^3$.
  • The "Fairness Tax": Achieving this equity comes with a trade-off; total energy costs rose by approximately 29% compared to the most aggressive cost-cutting algorithm. We argue this is a necessary investment to prevent localized socio-ecological harm.

Discussion and Conclusion

Our findings highlight that AI inference is "spatially flexible." Unlike physical manufacturing, a digital prompt can be processed almost anywhere, making GLB a powerful tool for social justice. By treating water consumption and model accuracy as vital components of an AI system's footprint, we move toward a more holistic definition of "Green AI."

Future work will address the limitations of this study, including the transition from "offline" optimization to real-time "online" systems that can handle weather and workload uncertainty. Additionally, we must consider data sovereignty regulations (like GDPR) that may limit the flexibility of global workload routing. While this paper focuses on the inference phase, addressing the inequities inherent in model training remains a critical next step for the field.

Summary updated

Abstract

The sharply increasing sizes of artificial intelligence (AI) models come with significant energy consumption and environmental footprints, which can disproportionately impact certain (often marginalized) regions and hence create environmental inequity concerns. Moreover, concerns with social inequity have also emerged, as AI computing resources may not be equitably distributed across the globe and users from certain disadvantaged regions with severe resource constraints can consistently experience inferior model performance. Importantly, the inequity concerns that encompass both social and environmental dimensions still remain unexplored and have increasingly hindered responsible AI. In this paper, we leverage the spatial flexibility of AI inference workloads and propose equitable geographical load balancing (GLB) to fairly balance AI's regional social and environmental costs. Concretely, to penalize the disproportionately high social and environmental costs for equity, we introduce $L_q$ norms as novel regularization terms into the optimization objective for GLB decisions. Our empirical results based on real-world AI inference traces demonstrate that while the existing GLB algorithms result in disproportionately large social and environmental costs in certain regions, our proposed equitable GLB can fairly balance AI's negative social and environmental costs across all the regions.
BibTeX
@misc{li2024towards,
  author = {Pengfei Li and Yejia Liu and Jianyi Yang and Shaolei Ren},
  title = {{Towards Socially and Environmentally Responsible AI}},
  year = {2024},
  eprint = {2407.05176},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.05176},
}
Figure