As Generative AI reshapes industries, its overlooked environmental toll and social inequalities demand bold, collective action for a sustainable future.
Commentary: A social-environmental impact perspective of generative artificial intelligence. Image Credit: Gorodenkoff / Shutterstock
A recent commentary article by researchers from Northwestern University, Harvard University, and The University of Texas at San Antonio highlights the significant but overlooked environmental and social impacts of Generative Artificial Intelligence (GenAI). Published in the journal Environmental Science and Ecotechnology, the research underscores the urgent need for sustainable practices and ethical governance as GenAI technologies proliferate.
The study reveals the environmental toll of GenAI development, with hardware production, such as GPUs and data centers, consuming vast resources and contributing to substantial environmental degradation. For example, the manufacturing site of Taiwan Semiconductor Manufacturing Company (TSMC) in Taichung is projected to consume 25% of the city's electricity and 6% of its water, raising concerns among local residents. Mining rare metals like cobalt and tantalum for these systems contributes to deforestation, water pollution, and soil degradation. In addition, training large GenAI models like GPT-3 typically consumes about 1,287 MWh of electricity and emits approximately 552 tons of CO₂, demonstrating the extensive energy demands of these systems. Data centers, essential for GenAI operations, are projected to consume over 8% of U.S. electricity by 2030, further straining energy grids. Additionally, GenAI systems generate substantial e-waste, exacerbating global pollution challenges.
On the social front, the study highlights inequities in GenAI's production and use. Labor concerns range from child exploitation in cobalt mining to underpaid workers training AI systems under precarious conditions. For instance, in Kolwezi, a mining hub in the Democratic Republic of Congo, children involved in cobalt mining operations face significant risks to their health and safety. Unequal access to GenAI deepens the global digital divide, privileging industrialized nations and English speakers over marginalized communities. Training tasks for GenAI models also suffer from systemic biases, as underrepresented groups such as women, older adults, and speakers of non-English languages are often excluded, leading to models that fail to address the needs of diverse populations.
The researchers advocate for immediate action to mitigate these impacts. Proposed measures include parameter-efficient fine-tuning, knowledge distillation to optimize model performance while reducing resource use, energy-efficient AI training, sustainable hardware designs, improved labor conditions, and inclusive governance frameworks. Transparency from developers and policymakers is essential, with recommendations for mandatory reporting of GenAI's environmental and social footprint in annual reports.
"This study sheds light on the hidden costs of GenAI and calls for collective action to address them," said lead author Mohammad Hosseini. The paper emphasizes the importance of a lifecycle assessment model to evaluate the full social and environmental impact of GenAI, from design to deployment, maintenance, and eventual decommissioning. The findings provide a roadmap for fostering responsible and equitable AI development globally.
Source:
Journal reference: