Harnessing AI For Sustainability

January 24, 2024

Insights

A global conversation is underway about using AI to address climate change challenges, bolstered by a new initiative from the UN Framework Convention on Climate Change (UNFCCC). Yet, there is growing recognition that AI uses an inordinate amount of energy, potentially exacerbating the very problems it is expected to solve.

A week before the UNFCCC’s COP28 meeting in Dubai, the Vector Institute brought together AI experts, sponsors, and Fastlane companies, for a roundtable discussion of how Canadian businesses can harness AI to meet and exceed their own environmental, social and governance (ESG) commitments. The conversation covered both AI as a tool for mitigating climate change and also how we can make it a greener technology.

Panelists included Katharine Preston, Vice President, Sustainable Investing at OMERS; Alik Sokolov, Co-Founder and CEO, Responsibli AI; Eric Morrow, Managing Director of Data Science and AI, BMO; Katerina Kindyni, Senior Manager, Financial Services Risk Management at EY; Arthur Berrill, Chief Technology Officer, RBC Data and Analytics; Iyngaran Panchacharam, Senior Manager, AI and Analytics in Sustainability; Olga Kravtsova, Innovations and Accelerators Portfolio Manager, PwC; and David Crane, Director of Product Development, AltaML.

AI will change how we approach ESG commitments

To kick off the discussions, Vector’s VP of AI Engineering, Deval Pandya, provided a practical overview of the intersection of AI and sustainability for organizations. He described how few companies are ready for mandatory ESG reporting standards. The lack of preparedness is due, in part, to a lack of consensus on what comprehensive ESG reporting requires.

Pandya also said that companies need better tools to meet their ESG commitments, including AI-based products and services. Pandya acknowledged the problematic power consumption and water usage of AI machinery, but he maintained that it can have an overwhelmingly net positive impact in our collective efforts to address climate change. Increasingly, he said, the AI community itself is stepping up specifically to address ESG. For example, the Responsible AI Institute recently formed the world’s first global leadership group to address “the positive and negative implications of AI for sustainability” (the Vector Institute is a founding member and Pandya resides on the working group). Pandya described how generative AI can help organizations amass, process, and report their environmental impact. Looking ahead, he anticipated that there will be many useful ESG applications of multimodal models and that AI as a tool for accelerating materials science could lead to a wide range of sustainability solutions.

Greener, more efficient AI

Gennady Pekhimenko–Vector Faculty Member, U of T  computer science professor, and CEO & Co-founder CentML–gave a fascinating presentation on his work making AI systems more efficient.

Blob

“We can massively reduce the carbon footprint required for training LLMs.”

Gennady Pekhimenko

Vector Faculty Member, University of Toronto Computer Science Professor, and CEO & Co-founder CentML

Pekhimenko tracks inefficiencies in how AI models are trained, to reduce their energy consumption. He and his team found that the widespread use of the fastest and most expensive GPUs has created bottlenecks in other parts of AI hardware systems, such as CPUs and memory. As a result, the power-hungry GPUs sit idle for one third of the time–all the while drawing electricity. The inefficiency of these systems makes a significant contribution to their high carbon footprint.

What we need, Pekhimenko said, is countermeasures: tools that track how and why energy is being used and that optimize AI systems and make them maximally efficient. While there are a range of existing tools purporting to profile how systems use power, Pekhimenko found that they were nowhere near granular enough in their analysis.

Pekhimenko’s research enabled CentML to develop better tools to identify and analyze bottlenecks, accurately quantify energy consumption, and measure environmental impacts of training AI models. CentML also predicts training execution times using different GPUs and recommends the most efficient solutions for any given job. They even quantify emissions across different cloud providers, including Google, AWS, and Azure, allowing clients to choose the option that runs as fast, cheap, and clean as possible for their specific use case.

Panel and roundtable discussions

The panel and roundtable discussions drew on the keynote address and technical presentation to explore concrete ESG challenges and solutions that organizations are exploring today. The first panel discussion focused on ESG concerns and climate change risks and the second panel covered AI-based solutions. Over the course of these discussions, the following themes emerged.

LLMs for ESG

AI is getting better at automating primary research activities required for tracking ESG commitments, according to one panelist. With the right models and data, not only can research be made more efficient, AI could ultimately provide a better understanding of a given scenario. Another panelist agreed that there is much AI can do in this space and a lot of room for innovation to gather data and create genuinely transformative AI tools. According to this panelist, a survey of more than 500 companies around the world revealed that large language models (LLMs) are now the focus of activity. Today’s LLM-based systems can already perform 80 per cent of the labour involved in research, a panelist said–and this is just the beginning. The hope is to be able to gather data from reports and other documents quickly and efficiently. Automating processes and improving data analytics can reduce the costs of meeting ESG commitments.

Predictive analytics for climate data

Another arena where AI is expected to help is data analytics. According to one panelist, there are three areas where AI will have the greatest impact in the efforts to mitigate climate change over the next few years. First, AI will help companies make better use of vast amounts of unstructured climate and emissions data by giving it structure and making it useful. Second, AI will improve predictive analytics, enhancing the ability of organizations to set emissions targets and de-risk their activities. Third, computer vision and real time analytics will help to automate how companies track key performance indicators.  Panelists also discussed how AI can help to integrate different kinds of climate models, e.g. by connecting models for heat distributions with models for flooding. This could also improve our ability to predict changes and risks we face in specific locations around the world.

Parsing data, protecting privacy

The ability of AI to help parse troves of data can be enormously beneficial. A participant described how large, 70-year-old organizations may have massive amounts of data from hundreds of legacy systems. AI can help to automate the work of pulling that data together and utilizing it. Another participant talked about the challenges of leveraging AI to efficiently track and calculate the environmental impact of business travel in their organization. While there are clear benefits of using AI to automate this work, the participant noted that it also raised implications for privacy. The question is: at what point does a system for tracking employee activity, such as travel, become a form of surveillance?

Measuring footprints

Panelists discussed challenges that the financial industry faces in capturing an organization’s worldwide environmental footprint, including its entire portfolio. This can be difficult, especially when there are long or unpredictable lags in receiving ESG data from portfolio companies. Another problem is that the fragmentation and inconsistency of practices and methodologies for tracking ESG across different companies and industries make it challenging to tally up a complete picture for an entire portfolio. Asked about where the greatest opportunities for change are, one participant said the focus should be on high emitting activities. For many companies, this means not scope 1 or 2 but rather scope 3 emissions. To address this, many multinationals are focusing on their supply chains to reduce their carbon footprint.

Embedding ESG commitments

A recurring topic among panelists and participants was the need to embed ESG commitments in an organization’s DNA. One panelist described how, in the financial industry, there is often a disconnect between investment and sustainability teams. The lack of cohesion tends to create inefficiency by setting climate issues aside as an external constraint on business activity. However, other panelists noted that things are starting to change. Companies are increasingly moving beyond treating ESG as a matter of meeting compliance obligations, recognizing ESG instead as part of their core values, and as part of their value creation.

Transition and Collaboration

Making progress on ESG commitments requires leadership from the top and hiring what one panelist called “internal muscle” to advance the organization toward meeting its goals. Strategies discussed included creating an internal climate institute to engage with governments, clients, and internal teams. The key, one participant said, is to get all the key people, teams, and disciplines represented in one room, set goals together, and ensure this interdisciplinary group can keep communicating. Another panelist emphasized the importance of including the end users of any new tool that is to be developed, as an opportunity to build trust and ensure the tool is used effectively. (As another participant pointed out, it is a common occurrence for end users to simply ignore new tools that companies develop or acquire.)

In addition to building effective teams, multiple panelists observed that the key to broader progress is collaboration on multiple scales–not only within organizations but also across industries and between the private and public sectors. Similarly, a panelist noted that developing better climate models would be difficult for any organization to do alone, and that the best approach is to establish conglomerates that pool climate data, resources and expertise. Panelists also noted that Canada faces particular challenges because the country continues to have a heavily resource based economy. The government is stepping up to support the transition to more sustainable practices, they said, and industry groups are now working with the government to create markets for investing in green technologies and practices.

Behavioral change

At the end of the roundtable discussion, the topic turned to the question of climate action. If the ultimate goal of climate and sustainability data analysis is to effect behavioural change in organizations and individuals, then what kinds of AI models do we need? Looking at the challenge this way again illuminates the importance of trust. As one participant noted, the more momentous a decision, the more critical it is that we trust the model. Since the outputs of simpler models are generally easier to explain, they are often easier to trust, and are therefore can be more effective tools for driving behavioural change. 

Healthcare AI applications–where decisions can be literally a matter of life or death–may provide guidance for building trust in AI models for ESG. Participants discussed a recent medical application of AI involving an algorithm designed to recommend radiotherapy treatments tailored to the needs of individual patients. Over a period of six years, the AI model’s recommendations were rigorously reviewed and compared to those of the best human doctors, and the health outcomes of human vs AI recommendations were carefully tracked. This careful process ultimately demonstrated that the algorithm reliably made the best possible recommendations for treatment, creating a new, improved–and trusted–standard of care. Building trust in AI models that automate ESG functions and guide our attempts to address the challenges of climate change could follow a similar path.

Next steps

The more useful and reliable AI-powered ESG tools become, the more Canadian organizations will embrace them to meet their commitments. 

From emissions to sustainability related research and application, Vector is working with partners and researchers using interdisciplinary approaches and open source solutions – fostering real results for the planet – exploring climate change from every angle through Vector’s collaborative projects.Check the Vector events page for new opportunities to join the conversation or contact us today or follow our GitHub here.

Related:

Headshot of Vector Faculty Member Xi He.
Insights
Trustworthy AI

How Vector Researcher Xi He uses differential privacy to help keep data private

Three people stare at a laptop with a Vector logo on it
Insights
Research

Vector Research Blog: Structured Neural Networks for Density Estimation and Causal Inference

Insights
Research

Vector Research Blog: Causal Effect Estimation Using Machine Learning