Debate: AI and Climate Change
Reasons AI development and usage will increase climate change
Reasons AI development will decrease climate change
Key Questions for tackling AI and climate change ethics
Introduction
A lot is written about the energy that is required to power today’s AI systems and how that significantly contributes to climate change. In this essay, I explore the arguments as to how expanded AI usage can increase climate change, reduce climate change, and help us adapt to climate change.
As with all the essays in this series, it concludes with a related moral claim and series of questions. Are the environmental costs accrued with increased energy production worth the benefits? In a broader context, are there other practices and technologies we should abandon to ameliorate climate change, even if that means reducing our quality of life and longevity?
AI Usage Is Increasing Climate Change
Let’s begin by looking at each way expanded AI usage can increase climate change.
Training AI Models
Training AI models requires significant computational power, leading to high energy consumption. For example, training GPT-3, a large language model with 175 billion parameters, consumed an estimated 1,287 MWh, equivalent to the annual energy consumption of 126 average U.S. homes. Training the even larger GPT-4 model likely produced 15 metric tons of CO2, and a hypothetical GPT-5 model would have an even larger footprint. ChatGPT4 was approximately 10X the size of 3.5 (2 trillion parameters). We don’t know how large ChatGPT5 will be, but if it is in the neighborhood of 20 trillion parameters, a substantial amount of energy will be consumed in training even if it is trained more efficiently.
Much of this energy comes from non-renewable sources, contributing to greenhouse gas emissions.
Operating AI Models
Operating the AI models also requires a lot of energy. According to one new study, “shipping 1.5 million AI server units per year by 2027 and operating them at capacity (expected based on current trends), would consume at least 85.4 terawatt-hours of electricity annually.” Image generation has an especially large impact.
Currently, the world consumes approximately 25,000 terawatts of energy, so this would represent approximately a 1.5% increase in global energy consumption, most of which would probably come from fossil fuels. While it is sizable, it’s probably a smaller total percentage than many imagine, though this is just factoring in use/prompting of the models and not the initial training.
Producing the Hardware (and Renewables)
Producing AI hardware like GPUs relies on extracting rare earth elements, which has severe environmental impacts including habitat destruction, water pollution, and increased carbon emissions. The demand for these resources is expected to rise with the proliferation of AI technologies and the hardware demands they create.
And if the energy used to support AI was provided by renewable energy, rare earth elements would also have to be mined.
Responding to Answers
In this section, I will cover the reasons why AI may reduce energy consumption/CO2 emissions. It is worth noting that these are largely offsetting reasons, as no one denies that there will be a net increase in energy consumption.
Energy efficiency. Some claim AI technologies will increase energy efficiency (see below). While AI can lead to efficiency gains, this can paradoxically result in higher overall consumption. This is known as the Jevons Paradox: When technological improvements make the use of a resource more efficient, the cost of using that resource effectively decreases. This cost reduction often leads to an increase in demand for the resource because it becomes cheaper to use. Consequently, the total consumption of the resource may increase, even though each use is more efficient.
Nuclear power. Some claim that nuclear will power the AI data centers, but nuclear reactors take at least 6-8 years to build. And smaller modular reactors (SMRS), which Bill Gates has invested in, have not met expectations. It would likely be decades before nuclear could absorb any more of the energy demand.
Renewables. Renewables could be expanded to at least meet some of the demand, but there are significant regulatory and policy hurdles to expansion. Renewables also have the environmental downsides that are discussed above. From a national security perspective, the US still depends on importing rare earth elements from China, which creates national security challenges.
Fusion. Fusion energy, which does not produce any emissions and avoids most of the mining externalities, is widely considered to be necessary to support a significant expansion of AI, especially at levels close to artificial general intelligence (AGI). Although there continue to be advances, this is a ways off.
How Might AI Developments Mitigate the Amount of Climate Change?
The rapid advancement of artificial intelligence (AI) has led to the development of increasingly sophisticated language models, such as GPT-4 and ChatGPT. While these large language models (LLMs) have demonstrated remarkable capabilities, their immense size and computational requirements have raised concerns about their energy consumption and environmental impact. As a result, there is a growing focus on developing smaller, more efficient AI models that can reduce energy demands while still delivering powerful performance.
Small language models (SLMs) have emerged as a promising solution to address the energy intensity of LLMs. These compact models feature streamlined architectures with fewer parameters and smaller training datasets, enabling them to operate efficiently on less powerful hardware platforms. Companies like Microsoft and Apple have recently introduced their own SLMs, such as Microsoft's Phi-3 and Apple's OpenELM, which offer comparable performance to larger models while consuming a fraction of the energy.
Small models can be as capable in specific instances.
Another key trend in AI development is the shift towards on-device AI, where models are designed to run directly on smartphones, IoT devices, and edge computing platforms. By processing data locally, on-device AI eliminates the need for constant communication with cloud servers, reducing latency and enhancing privacy. Small language models are particularly well-suited for on-device AI, as their compact size allows for efficient deployment on resource-constrained devices.
The development of SLMs and on-device AI has significant implications for reducing the energy demands driven by AI. By optimizing computational efficiency and minimizing reliance on power-hungry data centers, these approaches can help mitigate the environmental impact of AI technologies. Researchers and industry experts are actively exploring techniques such as model compression, knowledge distillation, and hardware optimization to further enhance the energy efficiency of AI systems.
New experimental approaches such as using optical networks to transfer between graphical processing units (GPUs) have the potential to reduce energy consumption.
However, it is important to recognize that the widespread adoption of AI, even with smaller models, on-device AI, and more efficient processing, will still likely lead to an overall increase in energy consumption. As AI becomes more ubiquitous in various industries and applications, the cumulative energy demands could be substantial.
How Might AI Reduce Climate Change?
Energy efficiency. By leveraging AI's capabilities, we can optimize energy consumption, reduce waste, and promote sustainable practices. Here's a summary of how AI can contribute to increased energy efficiency:
Smart Buildings and Energy Management: AI-powered building management systems can analyze data from sensors to efficiently regulate lighting, heating, and cooling. By monitoring occupancy, weather, and usage patterns, AI can optimize energy consumption in buildings, leading to significant energy savings.
Optimizing Industrial Processes: In manufacturing and logistics, AI can streamline processes and minimize energy usage. Intelligent robotics can automate tasks, while machine vision can perform quality control and identify defects early. This eliminates unnecessary manual labor and reduces wasted materials and emissions.
Smart Grid Management: AI can efficiently monitor and control the flow of electricity in real-time. By analyzing data from smart meters and other sources, AI can predict energy demand, optimize power distribution, and integrate renewable energy sources more effectively. This leads to a more stable and efficient power grid.
Energy Audits and Recommendations: AI applications can automate energy audits for buildings and suggest tailored improvements. By considering factors such as insulation needs, equipment upgrades, occupancy patterns, and local weather conditions, AI can provide faster and more accurate insights for energy efficiency.
Sustainable Product Design: AI can assist in designing sustainable products by identifying materials and manufacturing processes with minimal environmental impact. By analyzing vast amounts of data, AI can optimize product design for energy efficiency and sustainability.
Renewable Energy Integration: AI can help integrate renewable energy sources into the power grid more effectively. By forecasting energy supply and demand, AI can match variable renewable energy output with consumption patterns, maximizing the value of clean energy and promoting its adoption.
Fusion. Artificial intelligence (AI) is playing an increasingly important role in the development of fusion power, a promising source of clean, safe, and virtually limitless energy. By harnessing the power of AI, researchers are making significant strides in overcoming the challenges associated with achieving sustained fusion reactions.
One of the key ways AI is contributing to fusion research is by predicting and preventing plasma instabilities. In February 2024, a team from Princeton University and the Princeton Plasma Physics Laboratory (PPPL) developed an AI model capable of forecasting potential plasma instabilities up to 300 milliseconds in advance. This breakthrough allows researchers to adjust reactor parameters in real time, maintaining the stability required for successful fusion reactions.
AI is also being used to enhance the overall safety and efficiency of fusion reactors. By identifying and diagnosing issues, AI contributes to predictive maintenance and improved safety protocols. Furthermore, AI technologies optimize reactor design, material selection, and real-time monitoring, increasing operational efficiency and reducing maintenance costs.
In addition to improving existing technology, AI is accelerating the development of practical fusion energy solutions. Initiatives like the IAEA's "AI for Fusion" program provide a platform for collaboration and innovation among stakeholders. Companies such as Tokamak Energy and Commonwealth Fusion Systems are leveraging AI to build smaller, more affordable fusion reactors and drive advancements in critical components.
The potential of AI in fusion research has not gone unnoticed by industry leaders. Sam Altman, head of ChatGPT creator OpenAI, has invested heavily in fusion and believes it will eventually provide the enormous amounts of power demanded by next-gen AI. This symbiotic relationship between AI and fusion could lead to a future where clean, sustainable, and efficient energy solutions are widely available.
As the world faces the urgent need to address climate change and transition to clean energy sources, the combination of AI and fusion research offers a promising path forward. By accelerating fusion R&D with AI and fostering cross-community collaboration, we are moving closer to a future powered by fusion energy. While challenges remain, the recent breakthroughs in AI-driven plasma control and the growing investment in fusion technology suggest that this once-elusive goal may be within reach sooner than previously thought.
Natural gas leaks. Methane, the primary component of natural gas, is a potent greenhouse gas that traps over 80 times more heat in the atmosphere than carbon dioxide over a 20-year period. How secretive methane leaks are driving climate change Recent studies have found that methane leaks from natural gas production and distribution may be much higher than previously estimated, potentially erasing some of the climate benefits of switching from coal to natural gas. In fact, research suggests that leakage rates as low as 0.2% could make natural gas as damaging to the climate as coal.
To address this issue, companies are turning to AI-powered solutions that leverage satellite imagery, sensors, and advanced algorithms to detect and quantify methane leaks. For example, Duke Energy has reduced recordable methane leaks by over 85% using an AI platform that combines data from satellites, sensors, and other technologies to prioritize leak repairs.
Other companies, like Aclima, offer continuous gas leak intelligence solutions that use mobile monitoring and AI to identify leaks down to the source, enabling faster and more efficient leak repairs. Google is partnering with the Environmental Defense Fund to map methane leaks globally using AI applied to satellite imagery, making this information accessible to researchers and organizations through platforms like Google Earth Engine.
As the number of methane-detecting satellites increases, AI will play a crucial role in handling the vast amounts of data they generate.
The upcoming MethaneSAT mission, set to launch in 2024, will rely on AI to analyze the data it collects, helping to identify and mitigate methane leaks more effectively.
Supporting carbon capture and storage. AI can play a crucial role in supporting carbon capture, a technology that is essential in the fight against climate change. Carbon capture involves trapping carbon dioxide (CO2) emissions from power plants and industrial facilities and storing them underground, preventing them from entering the atmosphere and contributing to global warming. However, designing effective carbon capture systems can be challenging and time-consuming.
This is where AI comes in. Researchers are using AI to identify new materials for carbon capture, such as metal-organic frameworks (MOFs), which can selectively absorb CO2. By using generative AI techniques, machine learning, and simulations, scientists can quickly explore countless potential MOF configurations and identify the most promising candidates for carbon capture.
AI can also help optimize existing carbon capture systems. For example, a new AI-based tool developed by scientists can help lock up greenhouse gases like CO2 in porous rock formations faster and more precisely than ever before. The tool, named U-FNO, simulates pressure levels during carbon storage in a fraction of a second while doubling accuracy on certain tasks, helping scientists find optimal injection rates and sites.
Furthermore, AI can assist in carbon tracking, separation, and emission prediction. Image processing algorithms can detect carbon output from factories and power plants, while AI can predict future emissions and help policymakers plan emission reduction targets.
Carbon capture and storage (CCS) is a critical technology in reducing greenhouse gas emissions and is viewed as the only practical way to achieve deep decarbonization in industries such as cement, steel, and chemicals. The Intergovernmental Panel on Climate Change (IPCC) has highlighted that CCS is necessary to limit future temperature increases to 1.5°C and tackle global warming.
Can AI Help us Adapt to Climate Change?
As the world grapples with the increasing impacts of a changing climate, AI is being leveraged to help communities and ecosystems adapt to new realities.
One key area where AI is making a difference is in weather forecasting and climate modeling. Advanced AI algorithms can process vast amounts of climate data, identifying patterns and making more accurate predictions about future weather events and climate trends. This information is crucial for helping communities prepare for and respond to climate-related hazards, such as floods, droughts, and heatwaves.
AI is also being used to monitor and assess the impacts of climate change on natural systems, such as forests, oceans, and glaciers. By analyzing satellite imagery and sensor data, AI can track changes in land use, ocean temperatures, and ice cover, providing valuable insights into the health of these ecosystems and informing conservation efforts.
In the agricultural sector, AI is helping farmers adapt to changing weather patterns and optimize crop yields. Machine learning algorithms can analyze soil conditions, weather data, and crop performance to provide farmers with personalized recommendations on planting times, irrigation schedules, and pest control measures. This not only helps farmers maintain productivity in the face of climate variability but also promotes more sustainable agricultural practices.
Urban areas, which are particularly vulnerable to the impacts of climate change, are also benefiting from AI-driven solutions. AI is being used to optimize energy use in buildings, reducing greenhouse gas emissions and improving resilience to extreme weather events. It is also being applied to urban planning and disaster response, helping cities identify areas at risk and develop targeted adaptation strategies.
In Context: Is the Increased Energy Consumption Worth it?
As humans, especially those living in the developed world, we’ve always justified the use of energy to support our lifestyle and improve longevity. While many environmentalists have argued for a shift toward more sustainable forms of energy such as solar and wind, few (see the degrowth movement) have argued for reducing energy consumption, at least at the cost of quality of life and longevity.
Access to energy has enabled improvements in healthcare, education, and living conditions. Energy is required to sustain and improve quality of life. For example, having access to electricity in the home allows people to refrigerate food, stay in contact using the internet, and forgo the use of wood and charcoal-burning stoves, which can negatively affect health. Providing people with access to energy who live in areas with lower living standards can work to reduce poverty and increase quality of life. Moreover, energy usage has contributed to increasing life spans. There is a strong positive correlation between energy use, carbon production, and life expectancy.
There is strong evidence that AI will work to revolutionize healthcare, making diagnosis and treatment more accurate and accessible. It can also enhance education, personalize learning, and make knowledge more widely available. In the developing world, AI could be particularly impactful, as providing access to energy and technology in areas with lower living standards can help reduce poverty and improve quality of life.
So while it is undeniable that absent a shift to fusion or some type of synthetic biological process that can support the processing of data with only a trivial amount of energy consumption, AI will increase the demand for energy and a good chunk of that demand will be met with CO2 emitting sources that will increase climate change. The increase may not be as strong as some critics of AI contend, as there are energy-limiting uses such as small language models and on-device AI. But some energy use will increase, and absent breakthroughs in fusion or a quick adoption of nuclear (which has its own risks), environmental harms will increase.
This question in my mind is, is it worth it?
Key Questions to Consider/Discuss
If AI lengthens the average human lifespan by 5 years, is it worth it? 10 years? 20 years?
If AI makes education accessible to the developing world, is it worth it? If not, should we hold back “third world” development in to prevent climate change? Similarly, China lifted 100 million people out of poverty over the last 30 years, but significantly increased energy usage and climate change in the process. Should China have been held back?
If many of the impacts of climate change are irreversible due to the already-existing concentration of greenhouse gases in the atmosphere, is it right to develop AI to support carbon capture solutions?
If many of the impacts of climate change are irreversible due to the already-existing concentration of greenhouse gases in the atmosphere, is it right to develop AI to support needed adaptations to climate change?
Mining rare earth minerals that are in all consumer electronics, powering our electronics (phones, computers, etc.), and daily social media usage all consume substantial amounts of energy. Should we abandon these?
What about cars and airplanes? These all provide services that improve the quality of life and probably longevity but consume a lot of energy. Should we abandon these?
Considering the environmental cost of mining rare earth minerals, should we reconsider our reliance on consumer electronics?
The key determinant of energy consumption is economic growth. If AI were to cause such a dramatic increase in unemployment that the economy crashed, would this be good?
Should societies move to reducing economic growth to reduce the harms of climate change?
What modern conveniences and essential goods that consume energy are you willing to abandon to reduce climate change?