4 steps to using AI in an environmentally responsible way

With Artificial Intelligence (AI), a software system can process a huge amount of purpose-specific data and identify patterns and correlations. The result is an automated decision or performance of a task that would otherwise require a human to perform.

Since its inception, AI has proven to be a transformative technology that helps optimize a wide range of operations across numerous sectors. Companies are using AI to re-engineer processes in nearly every business and government function from sales, marketing and customer service to manufacturing, R&D, IT, human resources and finance. Companies are also using AI to augment their workforce with collaborative robots, or “cobots.” Cobots operate in conjunction with or in close proximity to humans to do such things as hyper-personalizing customer engagement and, ultimately, create better products.

But with such unprecedented prevalence and wide-ranging impact, it is imperative to ask whether the massive computing power required to run AI might have a negative effect on the environment. Or, could AI be that vanishingly rare phenomenon that advances economic growth – without causing environmental degradation?

AI can provide breakthroughs in water and energy conservation – but it comes with its own carbon footprint. It can improve numerous impacts of society – but it can also reinforce existing biased or unjust social status quos. AI holds the promise of reducing our collective global carbon footprint and conserving our natural resources. But to do so, we must learn to use it responsibly.

AI can help sustain the environment in a myriad of ways:

Reducing carbon footprint

A 2019 joint study by Microsoft and PwC forecasted that responsible use of AI can lead to a 4% (2.4 giga tonnes) drop in worldwide greenhouse gas (GHG) emissions by 2030. AI is already being used to optimize industrial and residential energy intake and reduce their respective carbon footprints. It is also being used to optimize carrier routes, predict and forecast supply and demand, predict and forecast maintenance, and manage autonomous transportation. All these optimizations will directly and indirectly lead to reductions in carbon footprints.

A notable example is how Google is using AI to optimize the energy consumption of its data centres. Using machine learning technology developed by DeepMind, coupled with optimization algorithms, Google has been able to improve the energy use of its data centres by 35%.

Optimizing the use of natural resources

AI is also predicting the output of energy generated by such green sources as solar, wind and hydro-based energy, thus ensuring minimal waste of these natural resources. AI helps conserve water usage in residential, manufacturing and agricultural areas. Predictive AI algorithms have developed new agricultural processes such as precision farming, ensuring that the exact amount of water required is used and only ripe crops are picked. Algorithms also assist in farmland planning, monitoring the health of crops and livestock, and developing efficient power-generation schemes and setups for power generators and power consumers alike.

Salo Sciences, which develops solutions to climate change and biodiversity loss, partnered with Vibrant Planet and Planet Labs to build a data-driven platform that maps the behavior of wildfires across California. The platform, CFO (California Forest Observatory), is an AI engine that leverages LIDAR and satellite data to provide a tree-level view of forest structure and fuel loads. By combining this information with data on wind and weather, soil and vegetation moisture, and population and infrastructure, CFO captures the complex drivers of wildfire risk. In the future, CFO will be integrated with contemporary wildfire models to provide a real-time, dynamic map of wildfire risk – one that can support both restoration planning and active fire operations.

Although AI offers dramatic process improvements, creates innovative new processes and is the main catalyst in disrupting several industrial sectors, it does not come without a price. A study developed last year by the University of Massachusetts Amherst concluded that training a typically large AI model would have 300 times the carbon footprint of flying from San Francisco to New York.

Even when used with the best of intentions, AI can have unanticipated consequences. As a result of its ability to dramatically optimize a business, AI can increase demand – and carbon footprint. Imagine a scenario in which a hauling business operates a fleet of trucks that can perform 100 deliveries a day. By developing an optimization algorithm, the company can improve truck routing and reduce carbon footprint for the 100 deliveries. This is a positive outcome – until the resulting increased demand and utilization raises deliveries to 150 per day and creates a larger carbon footprint than before.

Understanding that AI is a double-edged sword, we recommend following this four-step process:

1. Select the proper use case.

Not all optimizations lead to meaningful reductions in carbon footprints. The abstract nature of AI enables it to tackle a wide spectrum of challenges, so users must filter and prioritize only those processes that will be optimized by AI. This is also true when choosing potential sustainability use cases.

2. Select the right algorithm.

Use the right AI algorithm for the right problem or use case. The process of training an algorithm greatly affects the energy consumption level. If, for example, you have an algorithm that translates from language A to B, you can optimize the algorithm in a way that would only require less than 100 hours of training. Or, you could use a brute-force trial-and-error algorithm and increase the compute by approximately 250,000 hours.

3. Predict and track carbon footprint outcomes.

Good intentions are not enough and may lead to additional carbon footprint. Treat sustainability as a key success indicator in any AI-based project. Companies are also beginning to include carbon footprint estimates in their cost/benefit analyses for deploying AI selectively and responsibly. AI implementers should understand the collective effects of process optimization on sustainability.

4. Compensate footprint with renewable energy use.

Offset carbon footprint by utilizing green renewable energy to power AI models. Knowing that AI running on its data cents is a major contributor to energy consumption, Google committed to power its data centres by renewable energy – and has been a net-zero carbon emissions company since 2017.

CodeCarbon, a lightweight, open-source software package that integrates into a Python codebase, is one of the tools that can help organizations conduct these steps. By automatically fetching power and grid data, CodeCarbon can track the amount of carbon dioxide (CO2) produced by the cloud or by local computing resources used to execute an experiment such as training a machine-learning algorithm. It then provides developers with dashboards displaying the CO2 outcomes of the experiment or series of experiments. This visibility into the CO2 impact creates opportunities to reduce the resulting carbon footprints, by hosting the cloud infrastructure in geographical regions that use renewable energy sources, or by using more efficient hardware. CodeCarbon was jointly developed by Mila, a world-leading AI research institute in Montreal; BCG GAMMA, Boston Consulting Group’s global data science and AI team; Haverford College in Pennsylvania; and Comet, a meta machine-learning platform.

Artificial Intelligence has quickly become one of the most powerful tools humanity has ever created. But as with other breakthrough technologies that have led to both positive and negative outcomes, it is imperative to use it wisely. If used with care, AI could make all the difference in the world as humanity struggles to solve the looming climate crisis.