Climate Change: What is Computing’s Responsibility?

Vint Cerf is a Vice President and Chief Internet Evangelist for Google. He is known as one of the “Fathers of the Internet” for his work on TCP/IP protocols and the architecture of the Internet. For this and his continuing work on Internet development and his efforts to increase access to the Internet for everyone, he has received numerous accolades worldwide. Notably, he is a recipient of the ACM Alan M. Turing Award. His keynote focused on the impact of developing technologies both in contributing to the climate crisis and, potentially, to providing solutions.

Cerf discussed the role of machine learning in understanding climate change through detecting complex patterns (e.g. “atmospheric rivers”) and forecasting future trends. He advocated for extensive data collection on temperature changes, atmospheric conditions, and other climate variables to improve predictions and develop more effective interventions. He further stressed the need to refine theoretical models to improve their predictive power and, thus, help policymakers craft effective, targeted responses.

Like Berners-Lee, Cerf raised alarm regarding the urgency of the climate crisis. He noted, however, that even if changes were implemented immediately to drastically reduce emissions, the climate has already altered in ways that require serious consideration of climate adaptation strategies. He suggested that computing could be particularly useful here in modeling risks and benefits of novel interventions through large-scale simulations, thereby helping to avoid unintended negative consequences. He also noted the potential for advanced computational tools to assess regional risks and guide planning.

Cerf also echoed Berners-Lee in calling for solutions to combat misinformation. Chief among these is the need for increased public education to be able to evaluate online sources and verify data. Computing could further promote fact-based climate discourse by using digital signatures to authenticate images and reports, labeling AI-generated content, and flagging informational inaccuracies.

Cerf noted the key role of industry in driving change. Acknowledging that responding proactively to climate change contributes to the long-term profitability of companies, he urged businesses to integrate sustainability into their operations (e.g. pursuing improved hardware efficiency and the use of renewable energy) before such changes are legally mandated. He also underscored the need to develop robust and adaptable technologies that can withstand disruptions that might arise from environmental or political instability to avoid catastrophic failures in critical infrastructures.

The keynote concluded by emphasising that developing effective solutions will require collaboration between environmental scientists, economists, technologists, and policymakers, and that computing can be the bridge between these disciplines through modeling, predicting, and assessing synergistic solutions. In this way, computing can help with navigating the complexities of climate change while simultaneously building the requisite infrastructure for a more resilient, sustainable future.

The first day involved considerable discussion of computing professionals’ responsibilities with respect to the climate crisis. One dimension of this responsibility is ensuring that the environmental footprint of computing technologies themselves are minimised through actively considering these impacts throughout the development process. Responsible innovation frameworks were seen as useful in guiding development of new technologies in line with environmental considerations. There was some debate as to whether existing codes of ethics, such as those put out by ACM [1] and IEEE [2], adequately address the environmental impacts of computing.

Resilient system design emerged as an important challenge for the sector. Many agreed that the field places too much emphasis on rapid iteration, often leading to unsustainable infrastructure. Participants emphasised the need to create systems that can function reliably over long periods without requiring frequent maintenance. They also stressed that overly complex systems tend to be resource-intensive, and suggested that change in the direction of simplicity would be environmentally beneficial.

Regulation was another important theme of the day’s discussions. There was a suggestion that computing professionals should adopt regulatory structures similar to those in the medical field, where ethical review processes are a prerequisite for new developments. While this could ensure greater accountability for the long-term impact of technological advancements, there were as yet unresolved questions regarding the practical implications of this suggestion. There was agreement, however, about the importance of incorporating environmental and social impact assessments into all technological developments.

There was also debate about stronger regulation of specific classes of computing technologies. Technologies such as blockchain, cryptocurrencies, and generative AI were debated extensively due to their outsized environmental impacts. Discussion highlighted that impacts, both positive and negative (environmental and otherwise), are highly dependent on their specific applications and implementations, complicating any potential regulation.

Participants also discussed the importance of systemic, industry-wide change driven by policy reforms. Ensuring accountability was noted as a particular challenge, and participants explored how companies could be prevented from bypassing ethical responsibilities, as they might do through strategic public relations efforts.

Emma Strubell is the Raj Reddy Assistant Professor in the Language Technologies Institute (within the School of Computer Science) at Carnegie Mellon University. Strubell’s research focuses on advancing machine learning and natural language processing methodology and measurement in order to promote environmentally sustainable development and deployment of AI. work has been recognised with a Madrona AI Impact Award, best paper awards at ACL and EMNLP, and in 2024 they were named one of the most powerful people in AI by Business Insider.

Modern AI approaches, powered by deep learning and large language models (LLMs), have the potential to accelerate progress by augmenting human intelligence in our efforts to overcome urgent societal challenges such as climate change. At the same time, training and deploying these increasingly capable models comes at an increasingly high computational cost, with corresponding energy demands and environmental impacts. This presentation characterises the complex relationship between AI and the environment through the lens of LLMs, with a focus on describing what we know about AI’s direct environmental impacts.

The presentation begins by establishing some high level metrics for how much we know that AI is emitting in terms of direct GHG emissions, versus how much we should in fact be curbing those emissions. While there exist many potential environmentally beneficial applications of AI, such as energy optimisation, materials discovery, and policy analysis, many these benefits have yet to be demonstrated in practice while the negative environmental impacts are already quite clear. At the same time, AI’s energy consumption is growing, despite targets to reduce emissions due to ICT (including AI data centre emissions) by 50% by 2030 [1]. Self-reported GHG emissions from major tech companies are rising significantly due to the increasing development and deployment of generative AI such as LLMs [3, 2, 4], which are resource intensive during model development (training) and use (inference). While training is the most well-studied phase of the AI model lifecycle where data is most readily available, inference likely makes up the majority of AI energy use, and rapidly growing due to recent methodological trends such as DeepSeek.

The presentation then shows that estimates of data centre energy usage vary widely, and argues that this stems from a lack of available data. There is a need for greater transparency from major tech companies regarding their energy use and emissions to improve assessments of direct impacts. There is also a need for better benchmarking tools, following from first steps via the AI Energy Score project [5], to measure and compare the energy efficiency of different models, and to better understand alignment between AI methodology, capability, hardware, and energy requirements.

The talk explores in more detail in what ways AI directly impacts the environment (primarily: GHG emissions, water consumption, and waste production), and how those impacts arise (examples: fossil-fuel based energy powering data centres, evaporative cooling to cool hardware in data centres, and mineral extraction for hardware.)

The presentation advocates for the importance of mitigating AI’s direct impacts. The current, unsustainable, trajectory is a consequence of prioritising computational scaling over improved efficiency, or environmental sustainability. This could be spurred through enforcement of stricter emissions and energy-use guidelines, such as a carbon tax. The talk ends by calling for collaboration between policymakers, researchers, and industry leaders to put AI on a sustainable trajectory.

Tom Romanoff is the ACM Director of Policy. In his career he has led AI policy research and issued recommendations adopted in NIST governance, White House Executive Order requirements, and US legislation. He launched AI101.org to educate and inform congressional staff on AI background and issues. He has also directed research and developed recommendations on topics including cybersecurity, privacy, technology’s role in climate change mitigation, content moderation, data privacy, digital divide issues, and competition in the technology sector.

Romanoff’s keynote provided an overview of the global regulatory and legislative landscape, current policy challenges, and reflections on strategy for effective advocacy in technology policy discussions, e.g. by organisations such as ACM. He noted that governments are investing billions in AI research and development for strategic advantage. They are also taking different approaches to regulating AI in different regions, with the US taking a market-driven approach with a slow roll-out of safeguards, the EU implementing stricter regulations such as the AI Act, and China prioritising AI efficiency and large-scale deployment, focusing on AI-driven infrastructure and national security applications. This has led to global regulatory fragmentation, hindering progress in curbing AI’s environmental impacts.

Romanoff identified several challenges in AI governance as it relates to climate change. The first, echoing Strubell, is the lack of transparency around AI models and their environmental impacts. Another is the influence of corporate interests in shaping AI policy and the resulting prioritisation of profits over ethical considerations. To address these challenges, he advocated mandating transparency and developing clearer, and more proactive (rather than reactionary), AI accountability frameworks; greater involvement of independent experts in influencing policy; and development of mechanisms to ensure compliance and assess long-term policy effectiveness.

The remainder of the keynote explored strategy for influencing technology policy. He recommended pursuing both insider and outsider strategies, i.e. establishing relationships with policymakers while simultaneously adding pressure through media influence. He also stressed that materials presented to policymakers must be backed by robust data to enhance credibility and framed in an accessible manner to be consumed by busy staffers; and that they are strategically timed to coincide with policy cycles and current priorities. This often means being ready to respond to calls for comment at short notice. He ended by providing some insights into the ways ACM currently works to influence policy, and how workshop participants could leverage these mechanisms to amplify the group’s manifesto.

A major theme of the second day was the growing trend toward regulating of AI. The European Union’s AI Act was discussed as a key example of such efforts, with participants exploring its implications for technology companies inside Europe and beyond. Of concern was the apparent pattern of big tech firms using their lobbying power to shape regulations to suit their financial interests. A noted corporate strategy was framing AI as a necessary tool for progress, even progress on environmental issues, while downplaying its various negative impacts. The lack of transparency, e.g. regarding data usage and carbon emissions, was seen as enabling overly positive framings of AI. Aware of the challenges in measuring AIs’ environmental impacts, participants proposed creating standardised benchmarking tools to help with assessment of energy consumption. Participants also explored various strategies for mitigating AIs’ energy consumption, including developing more efficient algorithms, optimising model architectures, and exploring alternative computing paradigms such as neuromorphic computing. There was some discussion of incentives for industry leading a transformation along these lines, given that rising costs of AI infrastructure may make large-scale models financially unsustainable in the long run, even potentially catalysing a new “AI winter”.

The other major theme of the day was an exploration of the role of computing professionals in influencing policy. It was broadly agreed that researchers and industry experts should take a more proactive approach to influencing policy. This was seen as critical given that policymakers lack the technical knowledge needed to craft effective policies. Strategies for policy engagement were explored at length. Those noted as especially relevant were capitalising on opportunities for providing expert testimony and publishing research on AI’s impacts, but also included advocating for policies that promote transparency and sustainability, collaborating with advocacy groups, and educating the public on the long-term consequences of AI-driven energy consumption.