Another Overstated AI Concern

Artificial intelligence has sparked a new technopanic that has endured for well over a year. From doomsday stories about how AI development could theoretically lead to the extinction of the human race to worries that it could cause mass unemployment, there has been no shortage of fear surrounding society’s premiere area of innovation.

Many have also raised concerns over the possible environmental consequences of AI development. Refreshingly, these concerns are more pragmatic, but just like the others, they are based on risks that have been overstated.

The idea that AI will somehow turn on humanity and kill us all sure grabs people’s attention, but no one has been able to credibly explain how this might happen. Last year, renowned venture capitalist Marc Andreesen wrote succinctly that “[AI] is math — code — computers, built by people, owned by people, used by people, controlled by people. The idea that it will at some point develop a mind of its own and decide that it has motivations that lead it to try to kill us is a superstitious handwave.”

And the concern that innovation will lead to mass unemployment has been around as long as development has. Roman emperors and Queens of England quashed innovations on numerous occasions over unemployment fears. Despite all of the uproar throughout history about innovation possibly leading to mass unemployment, technological development has proven to be a potent job creator.

The unemployment rate in the U.S. is just 3.7% today, as technology has increased productivity and led to greater job growth.

But what about the concerns over the environmental impact of all the computing power required to train and run AI systems? Should we be worried about the emissions associated with various models?

The short answer is no. Why? Because these concerns are based on wildly inaccurate measurements and fail to take into account important factors of technological development.

In recent years, researchers have sought to study the CO2 emissions linked to AI systems. But a recent report by the Center for Data Innovation highlights the shortcomings of the measurements that have been cited to justify the environmental concerns.

For example, the report points to widely cited 2019 research at the University of Massachusetts Amherst as an example of the inaccuracy of environmental impact estimates. The researchers estimated that training a specific neural architecture search AI model used to offer an improved model for English-German translation generated 626,155 pounds of CO2, which is comparable to 300 roundtrip flights from the East to West Coast. But the assumptions that went into the research were incredibly flawed and after the individuals working on the model released a summary of the energy use and carbon emissions from their work, it was discovered that the actual emissions were about 88x smaller than reported.

It is odd that so much attention is placed on the emissions associated with training models when actually using models over time will cause more emissions than the original training phase. But estimating the emissions associated with running AI models is just as hard as estimating the impacts of training, and these estimates are similarly hyperbolic.

One researcher that the Center for Data Innovation pointed to estimated that, in a worst-case scenario, Google’s AI alone in one year could consume as much electricity as Ireland. That would require Google’s AI model to consume more energy on its own than the entire company currently does this year. That is ludacris. Even the researcher admitted that to consume that much electricity Google would have to invest $100 billion in chips and incur additional billions in operating costs. That would be uneconomical, and a profit-seeking organization would not sink that much capital into a technology that does not justify it with adequate returns.

Doomsday predictions about emissions related to AI models also often do not take into account future innovations and efficiency gains of the technology that goes into producing these models. Technological efficiency tends to improve over time — something that ought to be considered when estimating the environmental impact of AI.

And just as important, hyperbolic predictions about the possible negative consequences of AI readily ignore the potential for AI to help us solve environmental problems. AI models enhance our ability to analyze data collected about the environment and can be used to track and predict climate patterns and mitigate the effects of climate change. AI can also help us conserve water , combat environmental disasters such as wildfires, and optimize energy usage.

It is easy — normal even — to be fearful of new technology. People have been worrying over the impact of technological innovations throughout history. But suppressing AI development comes with risks as well. AI can be used to improve efficiency and make huge advancements in medicine, clean energy, and more. It can help us obtain greater economic growth and higher living standards. Assessing the negative impacts of AI is justified, but faulty, hyperbolic predictions and concerns often dominate the discussion around this revolutionary technology. The latest environmental worries too are overstated.

Source: Real Clear Energy