# Efficiency is Not Enough: A Critical Perspective of Environmentally Sustainable AI

**Authors:** Dustin Wright, Christian Igel, Gabrielle Samuel, Raghavendra Selvan  
**arXiv:** [2309.02065](https://arxiv.org/abs/2309.02065)  
**Date:** 2023-09-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.02065v1/

## Key points

- Artificial Intelligence (AI) has made significant advancements through machine learning (ML) methods like deep learning (DL)
- ML methods often require substantial computational power, consume high amounts of energy, and contribute to carbon emissions
- Efficiency alone is insufficient to make ML environmentally sustainable
- Three key discrepancies between efficiency and the overall environmental sustainability of ML are presented
- Technical and non-technical reasons at multiple levels of granularity are highlighted
- Systems thinking is proposed as a viable approach to improve the environmental sustainability of ML
- Increasing efficiency is important but not enough on its own to reduce the environmental footprint of ML systems
- A broader perspective that considers all aspects of ML's interaction with the environment is necessary for true sustainability
- Systems thinking is needed to effectively address the environmental challenges posed by AI and ML technologies.

## Summary

Artificial Intelligence (AI) has made significant advancements through machine learning (ML) methods like deep learning (DL), which have enabled AI to tackle tasks previously considered impossible. However, these ML methods often require substantial computational power, consume high amounts of energy, and contribute to carbon emissions that drive climate change. Moreover, the platforms on which ML systems operate also have environmental impacts beyond carbon emissions. To address these concerns and improve the environmental sustainability of ML, the industry and ML community have focused on increasing efficiency in terms of compute and energy consumption. However, this perspective argues that efficiency alone is insufficient to make ML environmentally sustainable. The authors present three key discrepancies between efficiency and the overall environmental sustainability of ML. By considering the various variables that interact with efficiency, the authors comprehensively demonstrate why efficiency alone cannot fully mitigate the environmental impacts of ML. They highlight both technical and non-technical reasons at multiple levels of granularity. Based on their analysis, they propose systems thinking as a viable approach to holistically improve the environmental sustainability of ML. In conclusion, while increasing efficiency is an important step towards reducing the environmental footprint of ML systems, it is not enough on its own. A broader perspective that considers all aspects of ML's interaction with the environment is necessary to achieve true sustainability. This critical perspective calls for a shift towards systems thinking to address the environmental challenges posed by AI and ML technologies effectively.