Green IA - book review
Gilles Babinet’s book, Green IA[1], explores the intersection of sustainability and artificial intelligence (AI), highlighting both their potential and limitations. AI's impact and usage are growing rapidly, but so is its resource consumption. Meanwhile, sustainability seeks to address the overuse of our finite resources. These two topics present complementary facets of humanity's genius and blind sides.
Babinet takes a technocentric approach, focusing on a few key points within his expertise. He begins by discussing the current debates on reducing and taxing pollution. AI is portrayed as both a problem—due to its significant and increasing environmental footprint—and a solution, offering tools for generating insights and making decisions. Babinet emphasizes themes such as improving transportation, rejuvenating industry, and enhancing energy systems. He concludes with a broader perspective on AI's deployment in daily life and society, envisioning autonomous personal and centralized collective modules.
Upon finishing the book, readers might believe that a mix of technology, policy, and small habit changes can ensure a safe future. However, this conclusion is a missed opportunity to present the fundamental decisions ahead of us, and the true potential of AI in shaping these decisions.
AI alone cannot save the planet. Cultural, societal, and policy changes are essential for reducing emissions. While data and AI provide valuable insights, they are not a cure-all. Using AI to optimize resource use, whether for cars or tools, operates only at the margins and overlooks the fundamental limits we face in a world with finite resources. Experts have already identified many solution pathways, such as those in The Drawdown Project[2] or Bill Gates' How to avoid a climate disaster[3]. The Climate Change AI association's groundbreaking report[4] also offers possibilities across all fundamental activities. Unfortunately, Babinet's focus on technocentric solutions with limited impact is a missed opportunity. For instance, using dams as energy reservoirs is limited to specific sites, and water scarcity and dam-related greenhouse gas emissions concerns further limit their potential.
The EU and France have led efforts to create decarbonized economies. However, the book glosses over a crucial reality: the EU has outsourced its polluting industries and is now figuring out how to rebuild a competitive, decarbonized industry through regulations and border taxes. France, for example, has mining potential for rare earths but closed its industries due to environmental concerns, leaving China as the global leader[[@pitronGuerreMetauxRares2018]]. Making AI chips is complex and highly polluting, and the EU must fully address the financial and environmental costs of developing a robust indigenous AI industry.
Regarding AI itself, the book presents a binary view: making hardware is very polluting, but AI chips are so efficient that their usage has minimal impact. This oversimplification shifts responsibility away from major decision-makers like large cloud companies (e.g., GAFA). In reality, a significant portion of emissions occurs during the use of data center chips when running language models (as noted by Sasha Luccioni[5]). Additionally, AI drives rapid turnover in consumer electronics, with manufacturers embedding more AI in phones to stand out. For instance, the latest iPhone performs full scene understanding to optimize photo quality, sometimes resulting in odd aberrations[6]. Without these features, a decade-old phone would still suffice for internet navigation and running Google Maps. It is time to debate the marginal benefits of having an AI supercomputer in our pockets as the social and environmental costs of AI rise.
AI can play a crucial role in addressing more distant but decisive challenges. For example, our developed world relies on refining crude oil into various products like plastics, oils, and medicine precursors. Physics dictates that a refinery using only electricity, water, and captured CO2 would need 5GW to operate—equivalent to four EPRs just to run a refinery[7]. The design of existing refineries has been refined over a century. We now have only 30 years to find an alternative, starting from basic chemistry. AI can profoundly impact exploring and validating numerous design possibilities in this area. Other topics that could benefit from AI include more accurate and trustworthy carbon credits and designing sustainable cement replacements.
In summary, Green AI initiates a timely and important discussion but falls short of fully addressing the scope and complexity of the challenges ahead.
- https://www.cultura.com/p-green-ia-l-intelligence-artificielle-au-service-du-climat-9782415008291.html ↩︎
- https://drawdown.org/ ↩︎
- https://www.penguinrandomhouse.com/books/633968/how-to-avoid-a-climate-disaster-by-bill-gates/ ↩︎
- https://dl.acm.org/doi/10.1145/3485128 ↩︎
- Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model ↩︎
- https://www.techradar.com/phones/mystery-solved-the-viral-glitch-in-the-matrix-iphone-wedding-photo-explained ↩︎
- https://www.nature.com/articles/s41586-024-07322-2 ↩︎