As countries worldwide strive to achieve net-zero carbon emissions by 2050, a new threat has quietly emerged that could hinder their progress. The increasing popularity and accessibility of generative artificial intelligence (AI) models, such as ChatGPT, have resulted in the widespread use of AI in everyday applications, from writing social media captions to creating slides.

Although many might hail this as the next big technological advancement, too much AI might actually not be a good thing. Despite the perception that everything operates seamlessly in the cloud, running these AI models actually has an insidious impact on our environment, similar to crypto mining. In this article, we’ll be delving into the environmental effects of AI and looking at the potential solutions to address them.

The environmental impacts of AI

Generative AI models utilize neural networks to detect patterns within extensive datasets, enabling them to generate original content.

Data centers, operating in the background of these models, are responsible for processing vast datasets using numerous Graphics Processing Units (GPUs) to train and analyze the models. In 2022, it was reported that the United States alone had 2,701 data centers.

With the continuous generation and storage of increasing volumes of data, the Global Datasphere is projected to grow to 175 zettabytes (175 trillion gigabytes) by 2025. Consequently, generative AI models are required to handle larger volumes of data. While using these models allows us to create new value from existing data, it comes at a substantial environmental cost.

The training phase of AI models consumes substantial energy, particularly for Natural Language Processing (NLP) models that analyze text and speech data. A study by the University of Massachusetts Amherst showed that the amount of carbon emitted when training the BERT language model using GPUs is comparable to that of a trans-American flight. As AI adoption in businesses continues to gain momentum, as reported by IBM, the carbon footprint associated with these models is poised to expand alongside the increasing volume of queries processed and answered.

The extensive use of water in data center operations poses an additional concern. Water-based cooling systems, which rely on evaporation, are employed to alleviate the substantial heat generated by these centers. Furthermore, these cooling systems discharge blowdown as wastewater, which is required to be treated by local utilities.

Although water serves as a cost-effective alternative to traditional air conditioning for cooling servers, its utilization can have implications for the local water supply, potentially affecting water availability in the region.

A study by The University of California, Riverside, showed that training the GPT-3 model consumed 700,000 liters of clean freshwater to keep servers cool and prevent overheating and damage to the equipment. The use of freshwater in cooling data centers is necessary to prevent issues such as corrosion, clogged water pipes, bacterial growth, and humidity control.

However, this could impact water-scarce areas, where freshwater is an important resource for other activities like agriculture and manufacturing. As a result, the establishment of data centers has faced opposition in US states like Arizona and Oregon due to concerns surrounding water scarcity and allocation.

Lastly, the rapid expansion of generative AI models may necessitate the use of more sophisticated equipment to operate these models, potentially rendering existing hardware obsolete.

Considering that e-waste is projected to increase by approximately 4% annually, reaching 74 million tonnes by 2030, the additional strain posed by AI hardware could further exacerbate this pressing issue.

AI can benefit the environment too

Although AI has been associated with negative environmental impacts, it holds the potential to optimize various sectors and foster sustainable practices.

AI can play a crucial role in analyzing agricultural areas to reduce land and water waste while enhancing crop yields for farmers. By harnessing AI-driven insights, farmers can make informed decisions during the harvesting process.

Google has established two teams that leverage satellite imagery and machine learning to develop a comprehensive “landscape understanding” of agricultural land in India. This approach has yielded valuable insights that aid farmers in preparing for drought conditions and mitigating their impact.

Vietnamese startup MimosaTEK offers agricultural solutions that enable farmers to monitor essential parameters such as soil moisture, temperature, humidity, and light intensity. These AI-powered tools empower farmers to engage in precision farming practices, leading to optimized resource management and increased productivity.

Other applications of AI include waste management, where it plays a crucial role in identifying opportunities for recycling and waste reduction. Waste Labs helps to optimize waste collection systems in cities by leveraging data to reduce costs throughout the supply chain. In addition to cost reduction, their efforts contribute to the circular economy by promoting sustainable waste management practices.

Traffic congestion is a major contributor to air pollution, particularly in densely populated urban areas. Although electric vehicles may help to reduce the environmental impact of transportation, AI can further enhance fuel efficiency by analyzing road usage patterns. Israeli-based NoTraffic tackles traffic congestion by collecting data from road users and utilizing an optimization engine to manage traffic lights effectively.

What can be done to reduce AI’s environmental impact?

While AI offers potential environmental benefits, it is crucial to address and minimize its environmental impact.

Transparency regarding the environmental impact of prominent tech companies is a crucial step in this direction. Microsoft, for instance, has pledged to become water-positive — replenishing more water than they use — by 2030. Their plan involves allocating resources toward wetland restoration projects and incorporating sustainable design standards into their data centers.

Furthermore, Google is committed to releasing its annual water metrics, revealing that in 2021, the average Google data center consumed approximately 450,000 gallons of water per day, or the same amount of water used to irrigate 17 acres of turf lawn grass once. Google has been developing low-water alternatives to cool data centers, while pledging to use alternatives to freshwater wherever possible.

Increasing transparency holds tech companies accountable, and it could foster greater data-sharing to collectively reduce the environmental impact of AI.

In addition, explorations into operating underwater data centers have shown promise. Microsoft launched Project Natick which deploys data centers on the ocean floor. The researchers hypothesized that these underwater data centers could reduce emissions and save on water required for cooling. After deploying the data center underwater for two years, the project demonstrated improved hardware reliability and a reduced need for replacement parts.

Other companies like Subsea Cloud are exploring subsea data centers too, having deployed one of its centers off the coast of Washington.

Apart from reducing consumption, prioritizing computationally efficient hardware and algorithms can further reduce the carbon footprint of AI. Untether AI is a Canadian firm that develops ultra-efficient, high-performance AI chips that scale more efficiently and cater to the needs of the masses.

Hyperscale facilities — massive centralized computing facilities that are highly efficient and are operated by a single company — are another option, as they offer significant advantages in energy consumption.

Operating at scale provides numerous cost advantages and greater resource sharing, something that smaller-scale facilities may not be able to achieve. According to a 2018 report, the adoption of hyperscale facilities for the entire cloud infrastructure could potentially result in a 25% reduction in energy consumption.

Lastly, government and regulatory bodies have a role to play in promoting green technology adoption in data centers. Carbon accounting software such as Microsoft Cloud for Sustainability and Salesforce Net Zero Cloud help track companies’ carbon footprints, increasing accountability.

Establishing standards that prioritize sustainability in data centers is essential. The AI for Good Foundation has been collaborating with governments to develop green AI frameworks that could incentivize the use of renewable energy sources by tech companies.

With greater collaboration and efficient infrastructure, AI’s environmental impact can be further reduced, allowing for its benefits to be realized while prioritizing sustainability.