To some, artificial intelligence (AI) is a silver bullet – a solution to all of our societal and environmental woes. We’re approaching the era of self-driving cars, AI-assisted medical diagnoses, AI- managed food systems and mass cloud computing; all of this amid what the United Nations calls the ‘defining issue of our time’ – the global climate crisis. Here, too, AI offers many solutions.
AI is set to improve the efficiency of technologies that power the grid, hastening the transition to low-carbon energy sources such as wind and solar. Machine-learning solutions can improve forecasting of supply and demand, enabling electricity plants to produce the right amount of power for a given area. Automated control of generators and infrastructure using AI systems can help to control energy outputs, or produce algorithms that calculate how much energy can be stored. Moreover, AI and machine learning are leading to breakthroughs in climate modelling.
However, there are signs that AI systems and the cloud computing that facilitates them need to clean up their own energy bills. In 2019, researchers at the University of Massachusetts Amherst explored the carbon emissions released when building and training natural processing language (NLP) models – AI systems that process human language. By converting the energy consumption in kilowatts to equivalent CO2 emissions, they showed that training a single NLP model emitted 300,000kg of CO2, equivalent to 125 round-trip flights between New York and Beijing.
Roel Dobbe, AI researcher at Delft University of Technology and formerly at the AI Now Institutev, explains that the carbon-intensiveness of many AI systems is driven by a belief in the power of what’s known as ‘big compute’. ‘In the AI field, there is a dominant but false belief that “bigger is better”, and that assumption drives the use of increased computation and bigger data sets in the development of AI models,’ he says. ‘As AI relies on more computational power, its carbon footprint increases.
‘AI and big computing are on an exponential trajectory,’ he adds. In 2018, OpenAI reported that ‘since 2012, the amount of compute used to train the largest AI systems has doubled every 3.4 months’. That equates to a 300,000-fold increase in the amount of computing power used in AI training runs. ‘It’s not just that you train these AI systems once: companies obtain data to keep training their AI systems, making some very carbon intensive,’ says Dobbe. What’s more, the speed of new 5G networks will drive the continued development of AI systems. 5G networks have latency figures (the amount of time it takes for a packet of data to get from one designated point to another) of three milliseconds – three times faster than the time taken for visual stimuli to travel from the human eye to the brain.
A NEED FOR TRANSPARENCY
Last year, a report identified that the broader tech sector will contribute 3–3.6 per cent of global greenhouse gas emissions by the end of 2020. A critical barrier to bringing down these emissions is transparency. Currently, cloud-computing players – such as Amazon Web Services, Google Cloud Platform and Microsoft Azure – aren’t required to disclose the energy use of data storage centres. This makes it difficult for the millions of companies and countless industries who use cloud computing to fully assess their own digital-carbon footprints. ‘Consumers need more insight into how their use of cloud computing affects their sustainability. The big players should provide this information in numerical form to their customers,’ says Dobbe. He’s perplexed as to why disclosure isn’t already mandated: ‘The hardware is already running; we know how many operations various algorithms need to run. It’s mostly a matter of political will and consumer awareness to enforce transparency.’
AI researchers at Cornell University in New York have taken matters into their own hands, developing an algorithm that shows how much equivalent CO2 is emitted as a result of computing and machine-learning models. The researchers would like to see their ‘Energy Usage Reports’ become widely used as part of a standard accountability process to improve the environmental awareness of big computing.
There is also the fact that AI and big compute are still being used to prop up fossil fuels. Last year, Amazon launched a programme called Predicting the Next Oil Field in Seconds with Machine Learning, while Microsoft held an event called Empowering Oil & Gas with AI. ‘On the one hand, tech companies are investing in renewable energy and trying to make their data centres more efficient. Their investment in oil and gas runs contradictory to this,’ says Dobbe. ‘When they’re selling their AI capabilities to oil and gas players, they’re actively trying to make non-renewables a more attractive investment prospect. AI should be used more responsibly, and for the global good.’