Massive benefits of AI come with environmental and human costs. Can AI itself be part of the solution?
Generative AI has a large carbon footprint and other drawbacks. But AI can help mitigate its own harms—by plowing through mountains of data on extreme weather and human displacement.
The recent explosion of generative artificial intelligence tools like ChatGPT and Dall-E enabled anyone with internet access to harness AI’s power for enhanced productivity, creativity, and problem-solving. With their ever-improving capabilities and expanding user base, these tools proved useful across disciplines, from the creative to the scientific.
But beneath the technological wonders of human-like conversation and creative expression lies a dirty secret—an alarming environmental and human cost. AI has an immense carbon footprint. Systems like ChatGPT take months to train in high-powered data centers, which demand huge amounts of electricity, much of which is still generated with fossil fuels, as well as water for cooling. “One of the reasons why Open AI needs investments [to the tune of] $10 billion from Microsoft is because they need to pay for all of that computation,” says Kentaro Toyama, a computer scientist at the University of Michigan. There’s also an ecological toll from mining rare minerals required for hardware and infrastructure. This environmental exploitation pollutes land, triggers natural disasters and causes large-scale human displacement. Finally, for data labeling needed to train and correct AI algorithms, the Big Data industry employs cheap and exploitative labor, often from the Global South.
Generative AI tools are based on large language models (LLMs), with most well-known being various versions of GPT. LLMs can perform natural language processing, including translating, summarizing and answering questions. They use artificial neural networks, called deep learning or machine learning. Inspired by the human brain, neural networks are made of millions of artificial neurons. “The basic principles of neural networks were known even in the 1950s and 1960s,” Toyama says, “but it’s only now, with the tremendous amount of compute power that we have, as well as huge amounts of data, that it’s become possible to train generative AI models.”
Though there aren’t any official figures about the power consumption or emissions from data centers, experts estimate that they use one percent of global electricity—more than entire countries.
In recent months, much attention has gone to the transformative benefits of these technologies. But it’s important to consider that these remarkable advances may come at a price.
AI’s carbon footprint
In their latest annual report, 2023 Landscape: Confronting Tech Power, the AI Now Institute, an independent policy research entity focusing on the concentration of power in the tech industry, says: “The constant push for scale in artificial intelligence has led Big Tech firms to develop hugely energy-intensive computational models that optimize for ‘accuracy’—through increasingly large datasets and computationally intensive model training—over more efficient and sustainable alternatives.”
Though there aren’t any official figures about the power consumption or emissions from data centers, experts estimate that they use one percent of global electricity—more than entire countries. In 2019, Emma Strubell, then a graduate researcher at the University of Massachusetts Amherst, estimated that training a single LLM resulted in over 280,000 kg in CO2 emissions—an equivalent of driving almost 1.2 million km in a gas-powered car. A couple of years later, David Patterson, a computer scientist from the University of California Berkeley, and colleagues, estimated GPT-3’s carbon footprint at over 550,000 kg of CO2 In 2022, the tech company Hugging Face, estimated the carbon footprint of its own language model, BLOOM, as 25,000 kg in CO2 emissions. (BLOOM’s footprint is lower because Hugging Face uses renewable energy, but it doubled when other life-cycle processes like hardware manufacturing and use were added.)
Luckily, despite the growing size and numbers of data centers, their increasing energy demands and emissions have not kept pace proportionately—thanks to renewable energy sources and energy-efficient hardware.
But emissions don’t tell the full story.
AI’s hidden human cost
“If historical colonialism annexed territories, their resources, and the bodies that worked on them, data colonialism’s power grab is both simpler and deeper: the capture and control of human life itself through appropriating the data that can be extracted from it for profit.” So write Nick Couldry and Ulises Mejias, authors of the book The Costs of Connection.
The energy requirements, hardware manufacture and the cheap human labor behind AI systems disproportionately affect marginalized communities.
Technologies we use daily inexorably gather our data. “Human experience, potentially every layer and aspect of it, is becoming the target of profitable extraction,” Couldry and Meijas say. This feeds data capitalism, the economic model built on the extraction and commodification of data. While we are being dispossessed of our data, Big Tech commodifies it for their own benefit. This results in consolidation of power structures that reinforce existing race, gender, class and other inequalities.
“The political economy around tech and tech companies, and the development in advances in AI contribute to massive displacement and pollution, and significantly changes the built environment,” says technologist and activist Yeshi Milner, who founded Data For Black Lives (D4BL) to create measurable change in Black people’s lives using data. The energy requirements, hardware manufacture and the cheap human labor behind AI systems disproportionately affect marginalized communities.
AI’s recent explosive growth spiked the demand for manual, behind-the-scenes tasks, creating an industry described by Mary Gray and Siddharth Suri as “ghost work” in their book. This invisible human workforce that lies behind the “magic” of AI, is overworked and underpaid, and very often based in the Global South. For example, workers in Kenya who made less than $2 an hour, were the behind the mechanism that trained ChatGPT to properly talk about violence, hate speech and sexual abuse. And, according to an article in Analytics India Magazine, in some cases these workers may not have been paid at all, a case for wage theft. An exposé by the Washington Post describes “digital sweatshops” in the Philippines, where thousands of workers experience low wages, delays in payment, and wage theft by Remotasks, a platform owned by Scale AI, a $7 billion dollar American startup. Rights groups and labor researchers have flagged Scale AI as one company that flouts basic labor standards for workers abroad.
It is possible to draw a parallel with chattel slavery—the most significant economic event that continues to shape the modern world—to see the business structures that allow for the massive exploitation of people, Milner says. Back then, people got chocolate, sugar, cotton; today, they get generative AI tools. “What’s invisible through distance—because [tech companies] also control what we see—is the massive exploitation,” Milner says.
“At Data for Black Lives, we are less concerned with whether AI will become human…[W]e’re more concerned with the growing power of AI to decide who’s human and who’s not,” Milner says. As a decision-making force, AI becomes a “justifying factor for policies, practices, rules that not just reinforce, but are currently turning the clock back generations years on people’s civil and human rights.”
Ironically, AI plays an important role in mitigating its own harms—by plowing through mountains of data about weather changes, extreme weather events and human displacement.
Nuria Oliver, a computer scientist, and co-founder and vice-president of the European Laboratory of Learning and Intelligent Systems (ELLIS), says that instead of focusing on the hypothetical existential risks of today’s AI, we should talk about its real, tangible risks.
“Because AI is a transverse discipline that you can apply to any field [from education, journalism, medicine, to transportation and energy], it has a transformative power…and an exponential impact,” she says.
AI's accountability
“At the core of what we were arguing about data capitalism [is] a call to action to abolish Big Data,” says Milner. “Not to abolish data itself, but the power structures that concentrate [its] power in the hands of very few actors.”
A comprehensive AI Act currently negotiated in the European Parliament aims to rein Big Tech in. It plans to introduce a rating of AI tools based on the harms caused to humans, while being as technology-neutral as possible. That sets standards for safe, transparent, traceable, non-discriminatory, and environmentally friendly AI systems, overseen by people, not automation. The regulations also ask for transparency in the content used to train generative AIs, particularly with copyrighted data, and also disclosing that the content is AI-generated. “This European regulation is setting the example for other regions and countries in the world,” Oliver says. But, she adds, such transparencies are hard to achieve.
Google, for example, recently updated its privacy policy to say that anything on the public internet will be used as training data. “Obviously, technology companies have to respond to their economic interests, so their decisions are not necessarily going to be the best for society and for the environment,” Oliver says. “And that’s why we need strong research institutions and civil society institutions to push for actions.” ELLIS also advocates for data centers to be built in locations where the energy can be produced sustainably.
Ironically, AI plays an important role in mitigating its own harms—by plowing through mountains of data about weather changes, extreme weather events and human displacement. “The only way to make sense of this data is using machine learning methods,” Oliver says.
Milner believes that the best way to expose AI-caused systemic inequalities is through people's stories. “In these last five years, so much of our work [at D4BL] has been creating new datasets, new data tools, bringing the data to life. To show the harms but also to continue to reclaim it as a tool for social change and for political change.” This change, she adds, will depend on whose hands it is in.
Your phone could show if a bridge is about to collapse
Researchers have tested a new smartphone app that could gather data about whether bridges are "structurally deficient" —a low-cost citizen-scientist alternative to current methods that are expensive and complex.
In summer 2017, Thomas Matarazzo, then a postdoctoral researcher at the Massachusetts Institute of Technology, landed in San Francisco with a colleague. They rented two cars, drove up to the Golden Gate bridge, timing it to the city’s rush hour, and rode over to the other side in heavy traffic. Once they reached the other end, they turned around and did it again. And again. And again.
“I drove over that bridge 100 times over five days, back and forth,” says Matarazzo, now an associate director of High-Performance Computing in the Center for Innovation in Engineering at the United States Military Academy, West Point. “It was surprisingly stressful, I never anticipated that. I had to maintain the speed of about 30 miles an hour when the speed limit is 45. I felt bad for everybody behind me.”
Matarazzo had to drive slowly because the quality of data they were collecting depended on it. The pair was designing and testing a new smartphone app that could gather data about the bridge’s structural integrity—a low-cost citizen-scientist alternative to the current industrial methods, which aren’t always possible, partly because they’re expensive and complex. In the era of aging infrastructure, when some bridges in the United States and other countries are structurally unsound to the point of collapsing, such an app could inform authorities about the need for urgent repairs, or at least prompt closing the most dangerous structures.
There are 619,588 bridges in the U.S., and some of them are very old. For example, the Benjamin Franklin Bridge connecting Philadelphia to Camden, N.J., is 96-years-old while the Brooklyn Bridge is 153. So it’s hardly surprising that many could use some upgrades. “In the U.S., a lot of them were built in the post-World War II period to accommodate the surge of motorization,” says Carlo Ratti, architect and engineer who directs the Senseable City Lab at Massachusetts Institute of Technology. “They are beginning to reach the end of their life.”
According to the 2022 American Road & Transportation Builders Association’s report, one in three U.S. bridges needs repair or replacement. The Department of Transportation (DOT) National Bridge Inventory (NBI) database reveals concerning numbers. Thirty-six percent of U.S. bridges need repair work and over 78,000 bridges should be replaced. More than 43,500 bridges are rated in poor condition and classified as “structurally deficient” – an alarming description. Yet, people drive over them 167.5 million times a day. The Pittsburgh bridge which collapsed in January this year—only hours before President Biden arrived to discuss the new infrastructure law—was on the “poor” rating list.
Assessing the structural integrity of a bridge is not an easy endeavor. Most of the time, these are visual inspections, Matarazzo explains. Engineers check cracks, rust and other signs of wear and tear. They also check for wildlife—birds which may build nests or even small animals that make homes inside the bridge structures, which can slowly chip at the structure. However, visual inspections may not tell the whole story. A more sophisticated and significantly more expensive inspection requires placing special sensors on the bridge that essentially listen to how the bridge vibrates.
“Some bridges can afford expensive sensors to do the job, but that comes at a very high cost—hundreds of thousands of dollars per bridge per year,” Ratti says.
We may think of bridges as immovable steel and concrete monoliths, but they naturally vibrate, oscillating slightly. That movement can be influenced by the traffic that passes over them, and even by wind. Bridges of different types vibrate differently—some have longer vibrational frequencies and others shorter ones. A good way to visualize this phenomenon is to place a ruler over the edge of a desk and flick it slightly. If the ruler protrudes far off the desk, it will vibrate slowly. But if you shorten the end that hangs off, it will vibrate much faster. It works similarly with bridges, except there are more factors at play, including not only the length, but also the design and the materials used.
The long suspension bridges such as the Golden Gate or Verrazano Narrows, which hang on a series of cables, are more flexible, and their vibration amplitudes are longer. The Golden Gate Bridge can vibrate at 0.106 Hertz, where one Hertz is one oscillation per second. “Think about standing on the bridge for about 10 seconds—that's how long it takes for it to move all the way up and all the way down in one oscillation,” Matarazzo says.
On the contrary, the concrete span bridges that rest on multiple columns like Brooklyn Bridge or Manhattan Bridge, are “stiffer” and have greater vibrational frequencies. A concrete bridge can have a frequency of 10 Hertz, moving 10 times in one second—like that shorter stretch of a ruler.
The special devices that can pick up and record these vibrations over time are called accelerometers. A network of these devices for each bridge can cost $20,000 to $50,000, and more—and require trained personnel to place them. The sensors also must stay on the bridge for some time to establish what’s a healthy vibrational baseline for a given bridge. Maintaining them adds to the cost. “Some bridges can afford expensive sensors to do the job, but that comes at a very high cost—hundreds of thousands of dollars per bridge per year,” Ratti says.
Making sense of the readouts they gather is another challenge, which requires a high level of technical expertise. “You generally need somebody, some type of expert capable of doing the analysis to translate that data into information,” says Matarazzo, which ticks up the price, so doing visual inspections often proves to be a more economical choice for state-level DOTs with tight budgets. “The existing systems work well, but have downsides,” Ratti says. The team thought the old method could use some modernizing.
Smartphones, which are carried by millions of people, contain dozens of sensors, including the accelerometers capable of picking up the bridges’ vibrations. That’s why Matarazzo and his colleague drove over the bridge 100 times—they were trying to pick up enough data. Timing it to rush hour supported that goal because traffic caused more “excitation,” Matarazzo explains. “Excitation is a big word we use when we talk about what drives the vibration,” he says. “When there's a lot of traffic, there's more excitation and more vibration.” They also collaborated with Uber, whose drivers made 72 trips across the bridge to gather data in different cars.
The next step was to clean the data from “noise”—various vibrations that weren’t relevant to the bridge but came from the cars themselves. “It could be jumps in speed, it could be potholes, it could be a bunch of other things," Matarazzo says. But as the team gathered more data, it became easier to tell the bridge vibrational frequencies from all others because the noises generated by cars, traffic and other things tend to “cancel out.”
The team specifically picked the Golden Gate bridge because the civil structural engineering community had studied it extensively over the years and collected a host of vibrational data, using traditional sensors. When the researchers compared their app-collected frequencies with those gathered by 240 accelerometers formerly placed on the Golden Gate, the results were the same—the data from the phones converged with that from the bridge’s sensors. The smartphone-collected data were just as good as those from industry devices.
The study authors estimate that officials could use crowdsourced data to make key improvements that would help new bridges to last about 14 years longer.
The team also tested their method on a different type of bridge—not a suspension one like the Golden Gate, but a concrete span bridge in Ciampino, Italy. There they compared 280 car trips over the bridge to the six sensors that had been placed on the bridge for seven months. The results were slightly less matching, but a larger volume of trips would fix the divergence, the researchers wrote in their study, titled Crowdsourcing bridge dynamic monitoring with smartphone vehicle trips, published last month in Nature Communications Engineering.
Although the smartphones proved effective, the app is not quite ready to be rolled out commercially for people to start using. “It is still a pilot version,” so there’s room for improvement, says Ratti, who co-authored the study. “But on a more optimistic note, it has really low barriers to entry—all you need is smartphones on cars—so that makes the system easy to reach a global audience.” And the study authors estimate that the use of crowdsourced data would result in a new bridge lasting about 14 years longer.
Matarazzo hopes that the app could be eventually accessible for your average citizen scientist to collect the data and supply it to their local transportation authorities. “I hope that this idea can spark a different type of relationship with infrastructure where people think about the data they're collecting as some type of contribution or investment into their communities,” he says. “So that they can help their own department of transportation, their own municipality to support that bridge and keep it maintained better, longer and safer.”
Lina Zeldovich has written about science, medicine and technology for Popular Science, Smithsonian, National Geographic, Scientific American, Reader’s Digest, the New York Times and other major national and international publications. A Columbia J-School alumna, she has won several awards for her stories, including the ASJA Crisis Coverage Award for Covid reporting, and has been a contributing editor at Nautilus Magazine. In 2021, Zeldovich released her first book, The Other Dark Matter, published by the University of Chicago Press, about the science and business of turning waste into wealth and health. You can find her on http://linazeldovich.com/ and @linazeldovich.
The Friday Five: Sugar could help catch cancer early
In this week's Friday Five, catching cancer early could depend on sugar, how to boost memory in a flash, a tiny sandwich cake could help the heart, and meet the top banana in the fight against Covid.
The Friday Five covers five stories in research that you may have missed this week. There are plenty of controversies and troubling ethical issues in science – and we get into many of them in our online magazine – but this news roundup focuses on scientific creativity and progress to give you a therapeutic dose of inspiration headed into the weekend.
Listen on Apple | Listen on Spotify | Listen on Stitcher | Listen on Amazon | Listen on Google
Here are the promising studies covered in this week's Friday Five:
- Catching cancer early could depend on sugar
- How to boost memory in a flash
- This is your brain on books
- A tiny sandwich cake could help the heart
- Meet the top banana for fighting Covid variants