Keep AI From Doing More Climate Harm Than Good

AI models and techniques can help drive a range of environmental goals. They can:

• Monitor and predict climate and weather-change trends such as global warming

• Manage waste and optimize recycling processes and operations

• Make transportation, mobility and routes more efficient to enhance fuel efficiencies and reduce carbon footprints

But data centers where AI is trained already account for about 2% of all U.S. electricity use, consuming 10 to 50 times more energy per floor space of a typical commercial office building. One recent study argues that ChatGPT needs to “drink” a 500ml bottle of water for every simple 20-50 questions and answers (and GPT-4 is even thirstier). 

Making AI itself more environmentally friendly is a key component of any sustainable technology program. Here are five ways to develop more sustainable AI.

• Consider adopting so-called composite AI, which uses network structures to organize and learn similarly to the efficient human brain.

• Composite AI uses knowledge graphs, causal networks and other “symbolic” representations to solve a wider range of business problems in a more effective manner.

• Monitor energy consumption during machine learning, and stop training AI as soon as improvements flatten out and no longer justify the costs of continuing. 

• Keep data for model training local, but share improvements at a central level. This type of “federated machine learning” reduces electricity consumption and bolsters data privacy. 

• Reuse models that have already been trained, and contextualize them, if necessary. 

• Use more energy-efficient hardware and networking equipment. 

• Manage when and where the AI workload happens. The carbon intensity of local energy supplies varies by country, generating authority, time of day, weather conditions, transfer agreements, fuel supply and other factors. 

• Balance follow-the-sun data center workloads, which are better for clean energy production, with unfollow-the-sun measures, which are better for water efficiency.

• Use energy-aware job scheduling, along with carbon tracking and forecasting services to reduce related emissions.

• Procure power purchase agreements (PPAs) when possible, or source renewable energy certificates (RECs) that reduce or offset greenhouse gas emissions and add new renewable energy to the grid where your organization will consume electricity. 

• Prepare for future protocols. PPAs and RECs aren’t perfect or always available, so start building a detailed plan of clean power by location, time of day or both. This type of analysis can help you build a clean-power strategy, which regulators may require going forward.

• Model environmental impacts, as well as business benefits, as you build AI strategy, and move forward with use cases that create more value than they destroy.

• Reduce the risk and energy of existing AI initiatives before proceeding. Improve their energy efficiency and lessen intellectual property and proprietary data risk.

• Do not invest in AI use cases that could damage business value or the environment.