It is undeniable that Artificial Intelligence (AI) has already begun to revolutionize our world, and its transformative potential only continues to expand. Yet, the power for positive change that AI brings holds the possibility for negative impacts on society. What are the sustainability challenges arising from AI and how could they be confronted to ensure that AI is truly used to enable a more sustainable future?
The three waves of AI
As AI rapidly evolves, propelled by an influx of data and computational power, the discourse surrounding AI ethics has emerged as a crucial arena of study. The first wave of AI ethics centered around speculative scenarios of superintelligent machines and robot uprisings, exploring what AI might do. The second wave of AI ethics grappled with the practical concerns of machine learning techniques: the black-box algorithm and the problem of explainability, the lack of equal representation in training data and the resulting biases in AI models, and the increase in facial and emotion recognition systems infringing on citizen’s rights (e.g. privacy). It is time to usher in the third wave of AI ethics, one that tackles the pressing environmental crisis head-on. This pivotal wave must actively engage academics, policymakers, AI developers, and the public at large, with a core focus on the environmental impact of AI.
AI vs Sustainability
In this pursuit, sustainable development must be at the heart of our endeavors. While a commendable movement to utilize AI for "good" causes (e.g. AI4Good) and align it with sustainable development goals has already gained traction, we must transcend this narrow approach. Let’s start with addressing the sustainability of developing and utilizing AI systems in and of themselves.
According to researchers, in a middle-ground scenario, it is estimated that by 2027, AI servers could consume between 85 to 134 terawatt hours (Twh) of electricity annually. To put this into perspective, this energy usage is on par with what entire countries like Argentina, the Netherlands, and Sweden each consume in a year. This is roughly equivalent to 0.5 percent of the world's current electricity consumption. The implications are clear: AI could have a substantial carbon footprint, depending on whether the power sources for data centers are fossil fuels or renewable energy.
To put the above into perspective, we need to recognize that in 2022, data centers, which support not only AI but all computers, accounted for roughly 1 to 1.3 percent of global electricity consumption. Cryptocurrency mining, responsible for another 0.4 percent, has partly transitioned to supporting AI operations. However, it's difficult to precisely quantify AI's energy usage since companies like OpenAI divulge limited details, including the number of specialized chips they require for their AI operations.
AI's impact on the environment goes beyond its day-to-day energy consumption. A significant portion of the environmental cost is linked to the manufacturing of the specialized chips needed for AI computation, says Peter Henderson, Computer Science PhD Candidate at Stanford. These chipsets may be efficient in terms of computation but are energy-intensive during production. This raises the question of how AI can tread a more sustainable path.
The legal path
From October 2023, all large companies that do business in California — including A.I.-intensive companies like OpenAI and Google — will have to become a lot more transparent about their climate risks and impacts under two major climate disclosure laws. The first requires all private companies with yearly global revenue above $1 billion to disclose how much carbon they produce in their operations by 2026, and in their supply chains by 2027. The second requires companies with revenue above $500 million to publish their climate-related financial risks by 2026.
The rules are the first of their kind in the United States, and more than 10,000 companies may be affected.
It's worth noting that the European Union took an early step in a similar direction in 2021. The EU issued regulations that compelled companies to disclose their carbon emissions. However, these rules faced some watering down in their implementation.
Addressing Challenges to a greener AI
While we may not be AI experts ourselves, the following solutions put forward by Bernard Marr in his article in Forbes did catch our attention:
Estimate carbon footprints of AI models. The Machine Learning Emissions Calculator can assist practitioners run estimations based on factors like cloud provider, geographic region, and hardware.
Go small. Developers can opt for smaller AI models when feasible, reducing computational demands and energy consumption.
Leverage Clean Energy. Running AI operations on grids powered by clean and renewable energy sources, such as hydroelectric power, can significantly reduce the carbon footprint.
Off-Peak Computing. Scheduling AI tasks during off-peak hours can help balance energy consumption and reduce environmental impact.
Location Matters. Relocating resource-intensive AI tasks to regions with eco-friendly energy sources can make a substantial difference.
Transparency and Measurement. AI researchers should publish not only the performance and accuracy metrics of their models but also the associated energy consumption. This transparency can drive innovation toward more energy-efficient AI.
Follow Google's 4M Best Practices. Google has identified four practices that can substantially reduce energy and carbon emissions when using AI, including selecting efficient model architectures, optimized processors, cloud-based computing, and location optimization to access clean energy sources
The challenges of AI and sustainability are a complex and intertwined matter that requires careful consideration. While AI shows great potential in becoming a tool for environmental and climate action, its environmental impact cannot be ignored. Systemic changes, such as climate disclosure laws, as well as smaller scale solutions could help harness the power of AI to create a more sustainable and energy-efficient world.