AI and climate change have a complicated relationship

Learn about the importance of environmental AI and its carbon impact in this comprehensive review.



By Lewis Lovejoy, CTO at Smarter.ai.

Since the concept of ‘thinking machines’ was introduced in Alan Turing’s 1950 paper "Computing Machinery and Intelligence," conversations surrounding Artificial Intelligence and its potential uses in modern technology have been widespread.

It’s entirely possible that the general population will have only been exposed to AI through Star Wars’ charming droids and The Jetsons’ robotic maid. However, while acting as backup for Luke Skywalker and Jane Jetson may be impactful enough for the average Joe, others may be wondering how AI models are tackling more complex issues like climate change and widespread pollution.

 

Climate Change and AI: What is Machine Learning’s Carbon Footprint?

 

Despite being praised for changing the world of work, the AI industry has an undeniably large carbon footprint that poses a climate conundrum.

An impactful study by the University of Massachusetts has revealed that the process of training a large AI model via deep learning released more than 626,000 pounds of carbon dioxide, [1] which is almost five times the emissions of the average American car over its lifetime.[2] Not only does this include its running emissions, but those released during manufacturing too.

When exploring this emission issue in more depth, language modeling proves an excellent case study. When various natural language processing models (NLP) were assessed, it was found that the energy consumption required to train a single model released 300,000kg of carbon dioxide.[3] OpenAI’s GPT-3 language model is a more specific example of AI’s staggering carbon footprint, releasing 552 metric tons of CO2 during its training.[4]

The training process for models is already incredibly energy-intensive, but any fine-tuning that businesses request to algorithms takes emissions to the next level. This is because fine-tuning involves reams of data being fed to a model, resulting in a massive energy drain through increased computing requirements.

 

What Can We Do to Combat Machine Learning’s Impact on Climate Change?

 

Although advancements in modeling are certainly exciting, the larger and more specific these models get, the more computation is required to train them. Using renewable energy grids for complex neural network training is something that big tech can invest in, but more efficient GPUs are crucial for reducing AI’s carbon impact on a global scale.[5]

What’s more, getting large global companies to share their data and quantify their emissions is crucial. After all, how can emissions be tackled in a meaningful way if we can’t get an accurate picture of where changes need to be made?

However, although this seems straightforward in principle, even when data is made available, pledges towards carbon neutrality and carbon-zero operations are being flouted by businesses. For example, Amazon’s emissions rose by a staggering 15% in 2019 alone despite the company’s “zero-carbon by 2025” pledge.[6]

Even though these figures are alarming, we must not lose hope.

The design of machine learning algorithms, the training hardware used, and electricity generation can all make a massive difference to AI’s carbon footprint. This infrastructure may be less cost-effective for tech firms, but a research study from the University of Berkley has shown that even shifting datacenters from coal-intensive power areas to electrically led areas could reduce emissions by a factor of 10 to 100 times.[7] If businesses migrate to the cloud for their algorithm training, the factor reduction could prove even more significant.

Furthermore, using neural network-specific chips can significantly reduce the energy required to train large models.[8] Upgrading models to require the latest AI chips available will involve initial cost investment, but the carbon savings over time are impressive.

 

How the World and Business Can Benefit from Environmental AI

 

With the climate crisis on our hands and the carbon impact of AI more visible than ever before, adopting environmentally sustainable practices should be a major consideration for every industry. Realistically, no matter how valuable a product or service is, foregoing sustainability in 2021 could lead to alienating valuable consumers.

If that wasn’t convincing enough, reducing a business’s carbon footprint can raise its profile and save money over time through improved efficiency. Mass recycling drives and solar paneling are effective stopgaps for businesses but stopping environmental degradation in its tracks requires a more radical approach with smart grid design, low-emission infrastructure, and transparent systems.[9]

Several large companies have already embraced environmental AI, with Google’s DeepMind division being one of the most prolific examples of climate change-related AI in the wider world. Although Google’s emissions are still higher than is optimal, DeepMind was instrumental in reducing Google’s data center energy requirements by a massive 40%,[10] showing just how effective environmental AI can be in making commercial firms more sustainable.

In time, companies will reduce their energy requirements through similar AI prediction models and greater resource management.

 

What Exactly Can Environmental AI Do on a Global Scale?

 

After all that negativity, you may be wondering what AI can do for the environment.

AI naturally plays a dual role in society, and despite its shortcomings when it comes to CO2 emissions, AI modeling can still make transport more sustainable, monitor the weather, assist with adaptive urban management, and promote sustainable fishing. This is certainly impressive, but it’s only a fraction of what environmental AI hopes to accomplish on a global scale.

To drive our point forward, let’s turn to the Texan energy company Xcel Energy as a specific example of Environmental AI in practice. To tackle climate change and move away from unreliable wind energy, this company invested in machine learning to better understand weather patterns. Using NCAR data from AI sensors on turbines, they can now predict wind levels more effectively.

As a result, they have managed to reduce their prices and use less coal and gas to generate their energy.[11] Deploying sustainable machine learning models on a global scale should allow businesses worldwide to mirror this success. Over time, this should make production more efficient and reduce reliance on unsustainable energy sources.

Due to their accuracy and large reach, AI models are also used successfully in urban planning. With the correct data, machine learning models can reduce congestion, automatically control heating levels, and increase the number of green spaces available to citizens.[12] As air pollution is responsible for approximately 4.2 million deaths per year, the introduction of effective environmental AI in these sectors could have a marked impact on global mortality.[13]

With the correct algorithms and intelligent systems in place, accurate measurements can be taken to monitor any environment over time, enabling urban planners to take effective steps towards sustainability before it’s too late.

Seeing as 97% of climate scientists believe that human activity is the main driver of climate change, using AI to reduce city emissions and manage pollution is an endeavor worthy of significant time and cost investment.

 

Diving into Deep Learning: The Complexity vs. Simplicity Argument

 

As we’ve discussed, AI technology almost always plays a dual role when it comes to the environment. In essence, it’s able to equally hamper and help the climate change crisis.

The complexity vs. simplicity argument is an eternal argument within the AI industry, as, in theory, the more complex businesses can make their AI models, the more likely they will be to revolutionize the future of production. It’s undeniable that this will massively benefit the average business’s bottom line, as complex models engage hyperparameters that make them more productive over time.

However, that’s not to say that complex neural networks are required for all businesses when simple Bayesian modeling would provide the data and projections required to fix common issues. In basic terms, why should companies purchase the equivalent of a Formula 1 car when a Honda Civic has enough computing power to get the job done?

Although this analogy refers to automobiles and not data centers, opting for simple models that don’t cost the earth will democratize AI, making the benefits of this technology more accessible to all. Over time, this will lead to fewer models being trained, and a reduction in carbon emissions as smaller businesses follow in the footsteps of big tech, using AI to improve efficiency. It may bolster a CEO’s ego to invest in the most complex tech around, but in practice, it is neither necessary nor particularly business-savvy to opt for bespoke AI solutions.

To slow environmental degradation, AI scientist Deepika Sandeep from Bharat Light & Power has urged companies to find alternative solutions to new modeling. This isn’t to say that AI can’t be used to solve problems, but the goal should be to minimize training cycles and avoid turning to deep neural networks when issues could be solved with “less compute-intensive AI.”[14]

 

Why Environmental AI is the Future

 

Regardless of where it is used, the future will feature AI in one way or another. As machine learning technology becomes more pervasive and necessary for businesses looking to gain a competitive edge, those not engaging with it will simply fall behind.

Although machine learning has relevance in the business world when it comes to ramping up production and assessing customer behavior, by reallocating time-consuming administration tasks and being where humans can’t, artificial intelligence can also help combat climate change by reducing the carbon impact that comes with uncertainty.

According to a study on AI and Sustainable Development Goals (SDGs) by Nature Communications in 2020, mass use of AI technology could assist with achieving 79% of all SDGs.[15] We certainly need to keep an eye on our use of energy-draining machine learning models and minimize the need to train new ones unnecessarily to reach these goals, but reusing AI models and opting for less complex solutions where possible should keep machine learning-related climate change at a manageable level.

 

A final word on AI and the environment

 

If you’re on the fence about environmental AI and whether investing in quality, reusable solutions will have a significant impact on climate change, then let me leave you with these parting words.

There is no denying that humans have excellent processing capacity. Nevertheless, neural networks and AI data processing techniques remain far more efficient than even the most diligent worker. As a result, choosing the “correct” route when it comes to AI is never going to be easy for businesses that are looking to prioritize profit over sustainability.

However, to slow climate change, reduce the need for energy-draining models, and boost overall response times to business issues, allowing reusable and less computing-intensive AI to take center stage over specific, fine-tuned models is, by all accounts, an excellent idea.

References

[1] Emma Strubell, Ananya Ganesh, Andrew McCallum, “Energy and Policy Considerations for Deep Learning in NLP”, Published 5 Jun 2019,

[2] Karen Hao, “Training a single AI model can emit as much carbon as five cars in their lifetimes”, last modified June 6, 2019, https://www.technologyreview.com/2019/06/06/239031/training-a-single-ai-model-can-emit-as-much-carbon-as-five-cars-in-their-lifetimes/

[3] Emma Strubell, Ananya Ganesh, Andrew McCallum, “Energy and Policy Considerations for Deep Learning in NLP”, Published 5 Jun 2019,

[4] Jeremy Kahn, “A.I’s carbon footprint is big, but easy to reduce, Google researchers say”, last modified April 22, 2021, https://fortune.com/2021/04/21/ai-carbon-footprint-reduce-environmental-impact-of-tech-google-research-study/

[5] Payal Dhar, “The Carbon Impact of Artificial Intelligence, Nat Mach Intell 2, 423 -425 (2020), Published 12 August 2020, https://doi.org/10.1038/s42256-020-0219-9

[6] Lisa Stiffler, “Amazon’s carbon footprint grew 15% in 2019, highlighting the challenge of net zero emissions”, last modified June 23, 2020, https://www.geekwire.com/2020/amazons-carbon-footprint-grew-15-2019-highlighting-challenge-net-zero-emissions/

[7] David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, Jeff Dean, “Carbon Emissions and Large Neural Network Training, Published 21 April 2021, https://arxiv.org/abs/2104.10350

[8] Jeremy Kahn, “A.I’s carbon footprint is big, but easy to reduce, Google researchers say”, last modified April 22, 2021, https://fortune.com/2021/04/21/ai-carbon-footprint-reduce-environmental-impact-of-tech-google-research-study/

[9] Payal Dhar, “The Carbon Impact of Artificial Intelligence, Nat Mach Intell 2, 423 -425 (2020), Published 12 August 2020, https://doi.org/10.1038/s42256-020-0219-9

[10] Richard Evans, Jim Gao, “DeepMind AI Reduces Google Data Centre Cooling Bill by 40%”, 20 Jul 2016, https://deepmind.com/blog/article/deepmind-ai-reduces-google-data-centre-cooling-bill-40

[11] Schiermeir, Quirin, “Germany’s Renewable Revolution Awaits Energy Forecast.”, Nature, vol. 535, no. 7611, 13 July 2016, pp. 212-213

[12] Eirini Malliaraki, “AI and climate change: the promise, the perils and pillars for action”, last modified 05 November 2020, https://www.climate-kic.org/opinion/ai-and-climate-change-the-promise-the-perils-and-pillars-for-action/

[13] World Health Organization (WHO), “Air Pollution”, WHO, 6 Nov 2017, https://www.who.int/health-topics/air-pollution

[14] Payal Dhar, “The Carbon Impact of Artificial Intelligence, Nat Mach Intell 2, 423 -425 (2020), Published 12 August 2020, https://doi.org/10.1038/s42256-020-0219-9

[15] Vinuesa, R., Azizpour, H., Leite, I. et al. The role of artificial intelligence in achieving the Sustainable Development Goals. Nat Commun 11233 (2020). https://doi.org/10.1038/s41467-019-14108-y

Original. Reposted with permission.

 

Bio: Lewis Lovejoy is a highly commercial technical lead with a proven track record in the financial technology arena, who has worked from a junior technologist to senior executive level by consistently delivering above expectations and within tight resource constraints. Lewis has strong self-motivation and keen leadership skills, coupled with a persuasive technical background and creative flair, which have helped a number of organisations realise their full potential in the marketplace.