AI - Good or Bad for the Planet? #24
Source of GHG emissions or enabler of a green economy?
I hope you enjoyed last week’s interview with Eleonore Eisath, Founder of Beworm!
I am Matteo, and this is my first deep dive. This week we will discuss a topic that is close to my academic background: AI’s role in the climate crises. We will reflect on how AI impacts climate change and whether it should be regarded as a source of GHG emissions or as a way to transition towards a greener economy.
🤖 AI - what is it?
Artificial intelligence (AI) has been on the lips of researchers, entrepreneurs, journalists, and tech enthusiasts over the past years. But what is AI?
AI is a novel branch of computer science. It develops computer systems that can simulate human intelligence and perform tasks. AI allows humanity to automate tedious and time consuming tasks and it can even outperform human capabilities in highly complex problems.
When talking about AI, buzzwords like machine learning, deep learning and, computer vision are used ambiguously.
Let’s try to do some clarity with these terms. I invite you to watch next this 5 min video on AI.
Most innovation in AI has taken place in the subfields of AI called machine learning and deep learning. For example, DeepMind developed AlphaFold in 2020 that solved the protein folding problem, a 50-year-old grand challenge in biology. Previously, DeepMind had developed AlphaGo that beat the world champion in Go, an ancient complex strategy board game originating from China.
The rapid development of deep learning has produced successful applications also in medical imaging. Recent research has developed deep learning-enabled computer vision that can improve the detection of breast cancer on mammography images.
Likewise, AI has lots of potential to help us solve climate change. Let’s start exploring those applications next.
🌍 AI for climate
The greenhouse gas (GHG) emissions total about 53 gigatons of carbon dioxide equivalent (CO2e) globally. To meet the goal of limiting the increase in average global temperatures to 1.5°C (Paris Agreement), we must reduce emissions by 50% by the end of this decade. AI can help us in this colossal task at hand.
According to a study conducted by Boston Consulting Group, AI can reduce GHG emissions between 2.6 and 5.3 gigatons of CO2e. This corresponds to 5%-10% of the global emissions. A similar study conducted by PricewaterhouseCoopers UK claims that AI could reduce GHG emissions by 4% between now and 2030. This 4% would be equivalent to the 2030 annual emissions of Australia, Canada, and Japan combined. So, the scale of emission reductions is there.
How is this scale of emission reductions possible with AI? AI has the superpower of learning from massive amounts of data and drawing smart recommendations for emission reductions that our human intuition may fail to see.
We summarize AI’s capabilities for the climate in the following four themes:
Distilling raw data into actionable information. For example, AI can analyze satellite images to identify deforestation or scan corporate databases and estimate companies’ climate impact.
Predict supply, consumption & emissions. For example, AI can provide minute-level forecasts of solar power generation to balance the electrical grid or predict future GHG emissions of companies.
Reduce emissions by optimizing complex systems. For example, AI can reduce the energy needed to heat a building or optimize freight transportation schedules.
Accelerating scientific breakthroughs. For example, AI can suggest promising materials for batteries and catalysts to speed up experimentation.
As we can see, the capabilities of AI for climate are wide, and so is the range of industries where it can be applied.
This paper, written by Global Partnership on AI, does an excellent job summarizing the potential applications of AI across relevant sectors. Below, we can see a condensed view of that summary with exciting examples of startups.
1. Electricity systems
AI can help balance power grids efficiently and smoothly integrate large amounts of renewables. It can also be used to precisely forecast electricity supply and demand, thereby reducing electricity waste.
Examples of startups:
Invenia increases the reliability and efficiency of the electrical grid by optimizing it through AI.
Myst AI uses the power of AI to precisely forecast the electricity and heating demand of businesses, thereby reducing energy waste.
2. Buildings and cities
AI can be used to assess the sustainability of infrastructure from satellite imagery. In smart buildings, AI can optimize functions such as heating and lighting to conserve energy.
Examples of startups:
BrainBox AI optimizes buildings’ energy consumption with AI. It can reduce carbon footprint by 20-40%.
75Fprovides a fully integrated smart building solution that reduces energy consumption by up to 50%.
LEDCity saves 90% of the energy consumption of lights in buildings using an AI-based switch on/off mechanism.
Kapacity.io provides a demand response service for building owners and operators. Its software uses machine learning to control the HVAC (Heating, Ventilation, and Air Conditioning) system.
AI can optimize freight routing and scheduling and increase the utilization of low-carbon options, such as trains. To increase the adoption of EVs, AI can optimize charging protocols and locations, and inform the design of batteries and next-generation fuels.
Examples of startups:
Aiconics uses AI to provide a 10x speedup in designing batteries.
Transmetrics optimizes logistic planning by leveraging AI with the potential to reduce supply chain emissions.
4. Heavy industry and manufacturing
AI can be used in adaptive control and process optimization to reduce the energy consumed by industrial processes. AI is being used to discover materials such as catalysts, which may decrease the energy needs of specific chemical processes.
Examples of startups:
Carbon Re provides a state-of-the-art AI to achieve operational efficiencies, reducing carbon emissions in industries like cement and steel.
CarbonChain allows heavy industries, such as mining, oil, and gas, to track and account for their emissions throughout the entire value chain.
Precision agriculture involves using AI in automated tools, which offers the potential for increased efficiency and reduced greenhouse gas emissions associated with agricultural chemicals and land use.
Examples of startups:
Iron Ox ensures that each plant receives the optimal levels of sunshine, water, and nutrients with AI and robotics.
Phenoinspec develops AI and computer vision for cloud-based plant inspection.
FarmWise uses AI and robotics to perform plant-level interventions to increase food production efficiency.
Phytoform Labs applies AI to genome editing to explore, predict, and develop effective novel crops.
Lima Labs helps farmers monitor their crops with IoT devices and AI.
6. Forestry and other land use
AI tools are being used together with satellite imagery in carbon stock estimation to inform land management decisions and calculate carbon offsets. AI is also being used to help track deforestation and other land-use changes.
Examples of startups:
GainForest is a decentralized crypto fund that uses artificial intelligence to measure and reward reforestation.
Farm-Trace by Taking Root is a software platform that uses AI to drive and verify reforestation across the tropics with smallholder farmers.
Dendra Systems provides ecosystem monitoring and management through AI.
7. Ecosystems and biodiversity
AI can support biodiversity preservation in the face of a changing climate. AI methods are increasingly being used to monitor wildlife and assess ecosystem change.
Examples of startups:
Spoor: Leverages AI to monitor biodiversity to understand and report environmental impact.
RhionsLab solves human-wildlife conflict, illegal wildlife trafficking, poaching using IoT and AI.
It is encouraging to see that AI can help tackle climate change in numerous ways. Furthermore, it is fantastic that so many ventures work on reducing GHG emissions by leveraging AI’s power.
While AI has a clear potential in reducing GHG emissions, it has its dark side. AI’s complex algorithms and computation systems utilize significant amounts of electricity. Therefore, AI has a carbon footprint that we don’t have to forget.
🔥 AI’s carbon footprint
The best performing algorithms are computationally intensive and have considerable GHG emissions.
Quantifying AI’s GHG emissions is a challenging task. The field is moving extremely fast and the corporations (Google, Meta, Amazon, Apple, etc.) that train and deploy the largest AI models are not transparent of their systems’ computational needs. However, we can try to get some proxies to understand this negative impact.
To get these proxies, we will: (1) estimate the emissions of data centers and (2) analyze the carbon footprint of specific AI algorithms.
(1) Data center emissions
Data centers process and store data. Data centers are also the hardware on which most AI algorithms are trained and deployed. The world’s data centers consume approximately 200 terawatt-hours of electricity, 1% of the global electricity demand. This contributes to 0.3% of all carbon emissions worldwide. Forecasts tell us that data centers’ electricity use will increase about 15-fold by 2030, reaching 8% of the total global electricity demand.
How much of the electricity consumption of data centers is caused by AI algorithms is unclear. The largest AI systems are owned by big corporations that do not publicly disclose the energy needs of their algorithms.
(2) The footprint of AI algorithms
To better understand what AI algorithms consume, we will look at some specific examples.
A study by Strubell et al. (2019) performed a life cycle assessment of training several large AI algorithms. They proved that these processes emit more than 284 tons of carbon dioxide equivalent. This corresponds to nearly five times the lifetime emissions of the average American car. The figure below shows how these emissions compare to other sources of CO2 emissions.
A growing body of evidence suggests that AI emissions are rising. The current state-of-art AI models are growing in complexity, implying that they also consume more energy. The amount of computing power used in the largest AI models doubles approximately every 3.4 months. (Compare this to Moore’s law, which predicts that the number of transistors in a circuit doubles every two years!)
Do these large and complex algorithms bring the added value required to justify such computational increase? Hard to say. It definitely depends on the problem that has to be solved. However, it can be shown that often marginal improvements come at very high carbon costs. For example, Parcollet and Revanelli show that a state-of-the-art deep learning model emits 50% of its total training released CO2 to solely achieve a 0.3 decrease in the error rate.
We’ve now seen both sides of AI. On the one hand, it is a powerful enabling technology to tackle climate change across many industries. On the other hand, the increasingly complex and computationally intensive AI models require significant energy to run.
What can we do to minimize the footprint of AI on the environment?
Firstly, it is crucial that researchers, tech professionals and entrepreneurs are aware of the technology’s carbon footprint. Awareness triggers reflection that can lead to change.
Paying attention to the choice of algorithms, location of data centers, and other parameters can alleviate the carbon emissions of these processes while providing similar performances. Luckily, there are solutions for measuring the CO2 emissions of computing. For example, CodeCarbon is an open-source software package that allows programmers to estimate the CO2 footprint of computing.
Secondly, within the climate space, we should focus on using AI in applications with a high GHG emission reduction potential. For example, if training an AI model for optimizing buildings’ heating systems initially causes tons of emissions, these emissions might be well justifiable in case the AI can reduce the building’s energy consumption by gigatons. This would result in a positive impact by orders of magnitude.
In conclusion, I believe that AI can help us overcome the climate crisis. However, I also think that we have to be aware of its negative externalities and pay attention to the choice of algorithm to make the best out of this impressive technology.
Your guide to AI - newsletter by Nathan Benaich
What do you think about AI & climate? Write it in the comments section on Substack, or send me an email!
I hope this deep dive gave you a good overview of the both the positive and the negative impact that AI can have on climate.
If you liked this article, it would mean a lot if you could share it below! Our humankind needs more people to join the mission💚
If you aren’t yet a subscriber, remember to hit that Subscribe button💚
I am very happy that you read my first deep dive. You will hear from us again on the 6th of January, after the Christmas break. See you soon Human!
Happy holidays! 🎅🏻
BCG (2021). Reduce Carbon and Costs with the Power of AI. Link.
Belkhir et al. (2017). Assessing ICT global emissions footprint: Trends to 2040 & recommendations. Link.
Ekin, A. (2019). AI can help us fight climate change. But it has an energy problem, too. Link.
EPFL (2020). Data centers need to consider their carbon footprint. Link.
GPAI (2021). Climate Change and AI, Recommendation for Government Action. Link.
Gupta, S. (2021). The imperative for Sustainable AI Systems. The Gradient. Link.
Jones, N. (2018). How to stop data centers from gobbling up the world’s electricity. Link.
Neslen, A. (2021). Here’s how AI can help fight climate change. Link.
OpenAI (2018). AI and Compute. Link.
PWC (2021). How AI can enable a sustainable future. Link.
Ravanelli et al. (2021). The Energy and Carbon Footprint of Training End-to-End Speech Recognizers. Link.
Rolnick et al. (2019). Tackling Climate Change with Machine Learning. Link.
Shen et al. (2019). Deep Learning to improve Breast Cancer Detection on Screeing Mammography. Link.
Shroff, R. (2019). Artificial Intelligence explained in simple terms. Link.
Siddik et al. (2021). The environmental footprint of data centers in the United States. Link.
Straubell et al. (2019). Energy and Policy Considerations for Deep Learning in NLP. Link.
World Economic Forum (2021). What would it take to make AI “greener”? Link.