Here's what AI for sustainability actually looks like

ChatGPT won't save the planet. Nor will Copilot, or Stability AI, or Claude. But there is still a sustainable promise to machine learning.

Apr 21, 2024 - 10:30

0 19

Here's what AI for sustainability actually looks like

Fast Facts

One of the oft-touted promises of AI involves addressing and mitigating the impacts of climate change.
The details of this effort, however, are complicated, particularly as AI consumes an enormous amount of energy.
TheStreet sat down with IBM executives to discuss the business of sustainability.

Microsoft (MSFT) CEO Satya Nadella said in his annual shareholder letter last year that climate change is the "defining issue of our generation," adding that artificial intelligence "can be a powerful accelerant in addressing the climate crisis."

Nvidia (NVDA) CEO Jensen Huang, echoing Nadella, has said that AI will lead to breakthroughs in climate science.

And Sam Altman, the former venture capitalist-turned-CEO of one of the most prominent names in the sector — OpenAI — has said that AI could lead to a host of benefits for mankind, such as addressing "climate change and curing cancer."

Despite these rosy perspectives, the reality of the relationship between AI and the climate is much more complex.

Part of that relationship involves the size and energy consumption of AI models.

Large Language Models (LLMs), which became popularized as the system behind ChatGPT, are known to suck up far more computing power than their traditional web search peers. The environmental cost of operating such models thus involves enormous quantities of electricity and water. And based on the location of the data centers hosting these models, in addition to the type of energy sources that power a given electrical grid, the carbon footprint of all that electricity consumption is likely quite large (though remains largely unknown).

Meta's latest model, Llama 3, for example, emitted 2,290 tons of carbon dioxide solely during its training. The U.S. Environmental Protection Agency (EPA) has found that the average gas-powered car emits one ton of carbon dioxide for every 2,500 miles driven.

It would take the average American (at 13,489 miles per year) around 420 years of driving to match that data point.

OpenAI and Google did not respond to requests for comment concerning the carbon footprint or electrical consumption of their AI models.

Google, which has plans to achieve net-zero carbon emissions by 2030 (partially through carbon offsets), in 2022 emitted 10.2 million metric tons of carbon dioxide equivalent. For comparison, Finland — with a population of around 5.5 million people, emitted 45.8 million metric tons of carbon dioxide equivalent that same year.

Google in 2022 also consumed 5.6 billion gallons of water, a 20% year-over-year increase. The bulk of that was consumed by its data centers. The company has committed to replenishing 120% of the water it consumes by 2030, but, according to its environmental report, replenished only 6% in 2022.

Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside, told Insider that Google's replenishment goal just "makes their water accounting look nicer," adding that "the water is still consumed."

Anthropic declined a request to comment.

More deep dives on AI:

Microsoft declined to provide details on the energy consumption or carbon footprint of its AI models, saying instead that it is "investing in research to measure the energy use and carbon impact of AI while working on ways to make large systems more efficient, in both training and application."

Bobby Hollis, vice president of energy at Microsoft, told TheStreet that the company is "continuing to invest in purchasing renewable energy and other efforts to meet our sustainability goals of being carbon negative, water positive and zero waste by 2030.”

In 2022, Microsoft emitted around 13 million metric tons of carbon dioxide equivalent, according to its environmental report. The company's global water usage spiked 34% to 1.7 billion gallons in 2022, an increase, according to Ren, that is due to its investments in AI.

A recent report from the International Energy Agency (IEA) found that data centers in the U.S. consumed 200 Terrawatt hours of electricity in 2022, equivalent to 4% of the country's total electricity demand.

The IEA expects that number to continue to rise.

The massive energy consumption posed by AI is well-known in the industry, with Altman himself noting in January that a new energy "breakthrough" will soon be needed to power AI.

"This is so symptomatic of the broken relationship between AI and the environment," Sasha Luccioni, an AI and sustainability researcher, said in response at the time. "We can't magically generate more energy ... We need to stop stuffing genAI into everything and reduce its energy use, right now."

The other side of that climate-AI relationship involves the legitimate applicability of certain technologies. The tech that Altman's OpenAI has become so well-known for is not the kind of tech that can help mitigate the impacts of climate change. ChatGPT and its ilk are generative AI models, designed to enhance corporate efficiency through instantaneous and synthetic textual output.

These LLMs, according to AI researcher and cognitive scientist Gary Marcus, "are not going to reinvent material science and save the climate."

"I feel that we are moving into a regime where the biggest benefit is efficiency," he told the New York Times last year. "I don’t have to type as much; I can be more productive. These tools might give us tremendous productivity benefits but also destroy the fabric of society."

But that's not to say that AI cannot be applied to help mitigate the impacts of climate change. Different, more specifically tuned iterations of AI and machine learning are already being applied to analyze vast amounts of data, providing pattern analysis that can in turn incite impactful action.

The climate promise of AI

The key difference between AI and machine learning, according to Konstantin Klemmer, a geospatial machine learning researcher and member of Climate Change AI, a global NGO working at the intersection of climate change and AI, has to do with size and specificity. Klemmer told TheStreet that the term "AI" is often used to describe very large, general-purpose models, while machine learning often refers to smaller models that are tuned to a specific task.

The smaller the model, the lesser the electricity demand.

And in applying AI or machine learning to the climate, Klemmer told TheStreet that focusing on the trade-offs is "essential."

"I think what is important is that we have a more systemic and formalized cost-benefit thinking when we think about these large models," he said, adding that the environmental cost of training and operating a very large model could be worthwhile if that model can then be leveraged to have an outsized positive climate impact. "But it's very unclear how this trade-off works."

"What we're seeing instead is just pushing forward on these models irrespective of their cost-benefit analysis," Klemmer added.

LLMs, he said, are not designed with the climate in mind. The impression that these language models could actually help humanity deal with and mitigate climate change, according to Klemmer, is based on the "far-fetched" argument that LLMs are the first step in the path to achieving artificial general intelligence (AGI), which refers to a hypothetical AI system that would have equal to or greater intelligence than a human. This is OpenAI's goal.

Many researchers, including Klemmer, are skeptical that AGI will ever be achieved. Others, like Marcus, have said that LLMs simply do not represent a pathway to AGI.

But LLMs are not the only iteration of AI out there. Klemmer said that geospatial models, which are getting "bigger and bigger," are allowing scientists to better understand the planet through the creation of better climate models. This, he said, is an example of an emission-intensive project that is likely "worth emissions," as such models can enable climate change adaptation.

The key word is 'efficiency'

For Klemmer, the promise of AI in climate change mitigation is all about efficiency, though not the same kind of email-writing efficiency that Marcus referenced.

"You can think of machine learning just as optimization," Klemmer said. "A lot of the high emission systems that we have on our planet are systems that could benefit from machine learning efforts to make them more efficient."

Inefficient power grids, buildings and homes all have an enormous carbon footprint that can be far-reduced through AI-fueled optimization, Klemmer said. Indeed, the National Renewable Energy Laboratory found in 2022 that air conditioning alone was responsible for about 4% of global emissions.

The key bridge between machine learning and climate mitigation is not AGI, Klemmer said. It is often simple pattern analysis and energy optimization, done at scale. His biggest wish at the moment is that there was more incentive for corporations to explore and integrate this kind of technology.

And even as the biggest tech companies in the world are spending more and more resources on building and integrating AI, many of these same companies are engaged in (largely small scale) efforts to apply AI not just to email writing or image creation, but for the greater good.

Google, for example, has its simply-named "AI for Social Good" project, whose mission is summed up on its website: "What if people could predict natural disasters before they happen? Track disease as it spreads, to eliminate it sooner? AI can help, but it’s not a silver bullet."

One of its climate-focused projects involves flood forecasting, where Google uses AI models to predict when and where a flood might occur, a model that enables people to evacuate before a flood.

Another involves the optimization of traffic light timing to reduce vehicular emissions.

More deep dives on AI:

Microsoft likewise has an "AI for Good" research lab, whose goal closely follows Google's own efforts in the space. One of its climate-focused projects involves the use of AI to closely track and map the Amazon rain forest to allow researchers to better prevent deforestation. Another provides for the tracking and measurement of global clean energy sources.

IBM told TheStreet last year that the big difference between it and other prominent tech giants is its focus on enterprise AI, rather than consumer-facing chatbots. But, like Microsoft and Google, IBM has also been exploring sustainable applications of AI.

TheStreet sat down with IBM's vice president and Chief Impact Officer Justina Nixon-Saintil, as well as Christina Shim, IBM's Global Head of product management and strategy, sustainability software, to discuss IBM's climate-positive explorations of AI, as well as the economic viability of sustainability.

IBM for the environment

IBM last year open-sourced its geospatial AI model, and has since announced a number of sustainability-focused projects and partnerships.

One of these such projects features a partnership with the Mohamed bin Zayed University of Artificial Intelligence; scientists at the university are using IBM's geospatial model to map and mitigate urban heat islands in Abu Dhabi. As of December, the project had resulted in the reduction of the impact of the heat island by more than 5.4 degrees Fahrenheit.

The company has also partnered with the Kenyan government to supercharge its reforestation efforts, and with groups across the U.K. to map urban areas and identify locations where trees can be planted to reduce future potential flood risks.

These partnerships came about alongside IBM's Sustainability Accelerator, a pro bono program that the company launched in 2022.

Through the Accelerator, IBM has also worked closely with farmers across the U.S., applying its AI models to help small farmers receive greater insights about water usage, a project that has led to increased crop yields in particularly arid regions like Texas.

IBM recently announced that it is investing $45 million in the Accelerator over the next five years. IBM reported a net income of $3.29 billion on $17.38 billion in revenue in the fourth quarter of 2023.

Internally, IBM applies its own technologies to enhance energy efficiency and track energy usage; still, the company used nearly 2.3 million megawatt hours of energy in 2023, according to a recent environmental report. But IBM said that 70.6% of that energy consumption was made up of renewable electricity — it emitted 364,000 metric tons of carbon dioxide for the year.

IBM has a goal of sourcing 90% of its electricity from renewable sources by 2030 and hopes to achieve net-zero emissions by then as well.

Sustainability, Nixon-Saintil said, is a "part of our culture. It's something where we have a leadership role and it's something we've been invested in for quite some time."

Beyond that, Shim explained that profit and purpose don't have to be mutually exclusive.

Still, at the same time as IBM is pushing non-profit sustainability projects, it is working with the oil, gas and chemical industries, offering its technology to help reduce operational costs, launch new products, enter new markets, pursue mergers and acquisitions and pursue new forms of value-creation.

One component of these offerings involves better energy tracking for environmental reports and disclosures.

The business of sustainability

IBM in February conducted a survey of 5,000 global C-suite executives that found that 75% of those surveyed believe sustainability "drives better business results," though about 70% said that sustainability efforts need to be made a higher priority.

"This is a boardroom discussion that needs to happen," Shim said, referencing the results of the survey. "The business case is getting stronger and stronger as there's more data coming out around it."

"Businesses I think are recognizing that AI is a powerful tool for sustainability," Shim added. "There's ways that they can reap more benefits and minimize environmental impact while they're doing so, and we're seeing that with a lot of these partner examples."

As part of its suite of AI-enabled enterprise products, IBM offers its customers a number of sustainability-minded solutions. These products are all focused on increasing operational efficiency across organizations, and optimizing data flows to better understand (in order to reduce) energy usage.

The balance

Still, the promise of AI for climate is one that is hinged upon cost-benefit analyses, particularly because the environmental cost of using such technology is so high.

"I think that we're currently seeing a trend of a one size fits all solution," Luccioni told TheStreet last year. "Everyone's trying to plug in LLMs and see what sticks and so I think we're going to see more energy usage, more compute, just because now everything has to have an LLM behind it, just because it's trendy."

I asked Shim and Nixon-Saintil about the balance between pushing generative AI for corporate productivity while also pursuing sustainable AI use cases. Neither provided details on the energy use or carbon footprint of IBM's models, saying instead that IBM has done a lot of research and investment in foundation models, which only have to be trained once before they can be tuned and applied in a number of different areas.

"There are ways to reap the benefits of AI while also minimizing the environmental impacts," Shim said.

Contact Ian with tips and AI stories via email, [email protected], or Signal 732-804-1223.

Original Post