91�鶹������Ʒ����

Authors: David Rolnick (McGill University, Mila - Quebec AI Institute), Priya L. Donti (Massachusetts Institute of Technology), Maria João Sousa (Climate Change AI), Sara Beery (Massachusetts Institute of Technology), Lynn Kaack (Hertie School), Amanda Sessim Parisenti, Sebastian Ruf 

AI can be one such tool, and is already being used in many applications that accelerate climate action. At the same time, AI is also being used in ways that impede climate action through its immediate applications and broader systemic effects, as well as its computational footprint. As national governments refine AI Strategies and governance structures it is critical that such strategies be cognizant of AI’s implications for climate action. In this piece, we analyze the ways that AI and climate change intersect and offer recommendations for decision-makers.  

Fostering uses of AI for climate action: , including electricity systems, buildings and cities, transportation, heavy industry, agriculture, forestry and other land use, climate science, societal resilience, biodiversity, climate finance, and climate policy. In each of these various sectors, there are several mechanisms by which AI can be useful: 

• Distilling raw data into actionable information. AI can identify useful information within large amounts of unstructured data, often by scaling up annotations that humans could provide more laboriously. For example, AI can analyze satellite imagery to monitor greenhouse gas emissions, pinpoint deforestation, or identify areas vulnerable to coastal flooding. Such decision-relevant information may help in establishing policies as well as with monitoring, reporting, and verification of ongoing efforts. 
• Improving predictions. AI can use past data to predict what will happen in the future. For example, AI can provide minute-level forecasts of solar power generation to help balance the electrical grid, or help governments predict agricultural yield after extreme weather events. 
• Optimizing complex systems. AI methods are good at optimizing for a specific objective given a complicated system with many variables that can be controlled simultaneously. For example, AI can be used to reduce the energy needed to heat and cool a building, to improve the efficiency of steel or cement manufacture, or to optimize freight transportation schedules. 
• Accelerating scientific modeling and discovery. AI can accelerate the process of scientific modeling, often by blending known physics-based constraints with approximations learned from data. For example, AI can suggest promising candidate materials for batteries and catalysts to speed up experimentation, and can quickly simulate portions of climate and weather models to make them more computationally tractable. 

It is important to note that AI is not a silver bullet in addressing climate change, and therefore should only be employed in contexts where it is actually needed and truly impactful. In addition, deployment of AI-for-climate applications must be underpinned by fundamental values of responsibility, ethics, and equity (in line with principles described within the UNESCO Recommendation on the Ethics of Artificial Intelligence and the Aarhus Convention on Access to Information, Public Participation in Decision-Making and Access to Justice in Environmental Matters). Close partnerships between experts in AI and relevant application areas are key to ensure the effectiveness of technological tools, avoid pitfalls, and establish a pathway to impact. 

To enable these benefits, in fostering access to data and digital infrastructure, targeting research and innovation funding to enable interdisciplinary and cross-sectoral collaboration, supporting deployment and systems integration of AI-for-climate applications, and establishing standards for meaningful stakeholder engagement and participatory design. 

Reducing negative impacts of AI applications: Every application of AI affects the climate, and does not only entail facilitating beneficial applications of AI. It also means shaping the space of AI overall so that business-as-usual applications are more climate-aligned, as well as reducing explicitly negative impacts. In particular, there are two principal ways in which applications of AI can adversely affect decarbonization strategies and result in an increase in greenhouse gas emissions [K+2022]. 

• Acceleration of high-emissions activities. For example, AI has been used extensively in oil and gas exploration and extraction, and .More subtly, perhaps, AI has become ubiquitous in digital recommender systems, resulting in highly personalized advertising (e.g. for fast fashion). This likely increases consumption and therefore comes with a significant (though unattributed) carbon footprint. 
• System-level societal impacts that affect the climate. These impacts are the hardest to quantify, but may be very significant. For example, self-driving personal cars are expected to lower the barrier to driving and may therefore cause a net increase in transportation-related emissions. AI-generated misinformation and the AI algorithms governing social media may have a significant impact on climate-related discourse and may influence support for decarbonization measures. More broadly, the growth of AI is inducing societal power shifts due to concentration of resources and expertise, with potentially negative ramifications for equity and climate justice. 

, national governments and international organizations should require climate-cognizant impact assessment and audits of AI applications under the purview of governmental regulation, funding mechanisms, and procurement programs. 

Addressing AI’s computational and hardware footprint: Some AI algorithms also have a non-negligible “direct” impact on greenhouse gas emissions, water, and materials through both computational energy and the production and disposal of underlying hardware. Exact numbers on such impacts do not yet exist. , of which AI accounts for an unknown fraction. The landscape of these impacts is shifting due to changing trends in the types of AI models used and potential changes in the pace of hardware efficiency improvements. For an individual AI algorithm, computation occurs both during development and training (infrequent but compute-intensive) and at inference time (less compute-intensive but more frequent). based on the type of model and data used; for instance, of state-of-the-art models trained on different computer vision and natural language processing datasets found that dynamic electricity consumption for training ranged from just 0.02 kWh to over 400,000 kWh.  

include requiring transparency on energy use, embodied emissions, and water use (e.g. from data center operators and AI solutions providers) and supporting RD&D on efficient and sustainable AI compute and hardware. 

Cross-cutting recommendations: International organizations and national governments can play a key role in building institutional capabilities across a wide range of organizations to enable the responsible implementation, evaluation, and governance of AI in the context of climate change. It is critical that AI ethics and governance bodies include impartial and climate-focused experts when shaping relevant AI policies and strategies, to ensure such developments align with climate goals. 

_{The ideas and opinions expressed in this article are those of the author and do not necessarily represent the views of UNESCO. The designations employed and the presentation of material throughout the publication do not imply the expression of any opinion whatsoever on the part of UNESCO concerning the legal status of any country, city or area or of its authorities, or concerning its frontiers or boundaries.}