Will AI be helpful or harmful in the fight against climate change? 

Recent breakthroughs in artificial intelligence (AI) and the proliferation of user-friendly software libraries have led to AI applications across various sectors, including healthcare, finance and marketing. These advancements have enabled the automation of complex processes. However, the substantial computing power and vast datasets required to train these models raise both ecological and social concerns. Moreover, AI predictions can be difficult to explain and as such this poses challenges for their implementation in sensitive areas such as healthcare, public policy, and surveillance. 

Adoption of the AI Act by the European Commission 

Addressing some of these challenges, the European Commission adopted the AI Act on February 2, 2024. This regulation aims to establish “harmonized rules for artificial intelligence” and introduces a risk assessment framework specific to AI applications. It mandates a set of control and transparency measures to uphold ethical principles and protect fundamental rights. 

At EcoAct, we welcome this regulation, which aligns with our internal practices. Our Climate Data Analytics team is proactive in ensuring traceability and transparency in the use of predictive models within our applications. However, it’s disappointing to note that the ecological impact of AI model deployment is not addressed in this regulation. 

What impact will AI have on the climate? 

While cloud computing solutions have made AI models accessible to anyone with a smartphone, this convenience creates an illusion of immateriality and inconsequentiality. In reality, the recent performance gains in AI are built on increasingly large models trained on ever-expanding datasets. 

Between 2018 and 2023, the size of reference models grew by a factor of 3,000 (from BERT’s 340 million parameters to GPT-4’s estimated 1.5 trillion). This growth translates to larger data centres, requiring significant electricity for powering servers and water for cooling them. 

To put this into perspective, the final training of GPT-3 is estimated to have emitted 500 tonnes of CO2. For GPT-4, based on leaked information about its training duration, estimates range from 6,000 to 12,000 tonnes of CO2. Following the trend of OpenAI’s model updates, we can project that consumption may increase tenfold for the fifth version. 

It’s worth noting that model training is just the tip of the iceberg, with numerous intermediate experiments potentially consuming as much or more energy. The lack of transparency regarding the precise impact of these processes is concerning. 

Moreover, the use of these models post-training is not without consequence. A single query to ChatGPT (GPT-3) consumes several times more energy than a Google search, a gap likely to widen with newer, larger models. 

How can AI become a positive force in the fight against climate change? 

Given this growth and impact, it’s crucial that best practices in neural model usage become widespread. Technical solutions such as model quantization (drastically reducing model size with minimal performance loss) and intelligent routing to models of varying performance based on query complexity can help mitigate these issues. 

However, it’s important to remember that the most environmentally friendly feature is the one not implemented. We must critically assess the necessity of such solutions on a case-by-case basis. 

At EcoAct, we use machine learning approaches only when simpler, less resource-intensive IT solutions are not viable. We also work to reduce the impact of these solutions when they are necessary. For instance, in developing a document database chatbot, we’ve implemented preliminary filtering of irrelevant queries and display the energy consumption of each query to the user. 

While we must consider the climate consequences of AI use, we shouldn’t overlook its potential to protect the environment. Possible applications include modelling climatic phenomena, optimizing energy production and urban systems, and leveraging neural models to process satellite imagery for monitoring ecosystems and combating illegal fishing. 

How do we use artificial intelligence at EcoAct? 

At EcoAct, we harness AI capabilities to enhance our climate, biodiversity, and carbon consulting activities. We use predictive models to estimate missing emissions data in carbon assessments and language models to improve access to documentary databases for our biodiversity missions. 

In this era of widespread AI adoption, it’s crucial to consider the significant climate impact of its development and use. While it’s regrettable that the recently adopted AI Act doesn’t address these issues, at EcoAct, we remain convinced of AI’s potential as a powerful tool in the fight against climate change. With responsible development and deployment, AI can indeed be a force for good in our collective efforts to protect the planet.