Introduction by Croakey: South Australian Premier Peter Malinauskas this week announced a Royal Commission into Artificial Intelligence (AI) “to ensure South Australians maximise benefits and minimise unwanted impacts from the AI boom”, including in healthcare.
The terms of reference also refers to AI infrastructure “opportunities” including energy transformation, water usage and associated impacts (such as electricity grid implications), but the words environment and sustainability are noticeably absent from this scoping.
The Royal Commission, due to report within a year, will hopefully consider the United Nation’s latest report on the environmental costs of artificial intelligence, which maps a complex and expanding physical infrastructure of data centres, cooling systems, electricity grids, water and land resources, and critical mineral supply chains.
The report estimates that by 2030, global data centres’ electricity consumption could exceed 945 TWh, almost three percent of projected global electricity use – enough to supply residential electricity to all 1.3 billion people in Sub-Saharan Africa for about 5.5 years. By that time, the UN predicts AI infrastructure could also generate up to 2.5 million metric tons of e-waste annually, equivalent to discarding nearly 250 Eiffel Towers every year.
In the latest of our Croakey Conference News Service stories from the recent Royal Australian and New Zealand College of Radiologists’ #Intelligence26 conference, United States researcher Dr Florence Doo challenges health administrators and medical professionals to carefully consider the environmental impacts of the systems and devices they are proposing to adopt.
Talking with Croakey reporter Marie McInerney, Doo argues the health sector needs to be able to measure the planetary costs of its AI use and make judicious decisions about when and how it might be justified.
Marie McInerney writes:
Like many of her colleagues in radiology, United States researcher and academic Dr Florence Doo has hope that the healthcare innovation offered by artificial intelligence (AI) includes that it may be able to reduce the specialty’s big carbon footprint.
But she warns that AI is also a “double edged sword” with a huge environmental cost on its own, so radiology and radiation oncology need to think about what types of innovations are being invested in, how they are produced and run, and how that’s “impacting our planet”.
Doo, who spoke at the recent RANZCR AI conference in Sydney, recommends radiology departments audit their use of energy, including through AI, and that Australian researchers look into the potential equity costs of AI use, including who can access it and who is affected by rising energy or water costs.
She said healthcare professionals need to consider: “Are we making those right choices for our communities and not just burning up the environment in what we’re doing?”
Measuring sustainability
Doo is an abdominal radiologist and clinical informaticist, who is Director of Innovation at the University of Maryland Medical Intelligent Imaging (UM2ii) Center and co-lead of the AI-enabled Medical Imaging research team in the University’s Center for Applied AI.
She has written about the need for transparent efficiency metrics and reporting standards for radiology AI models, “akin to the Energy Star rating for appliances and imaging equipment”, which could inform scientists, software developers, and radiologists about the scope of AI-related greenhouse gas emissions.
Doo and her colleagues published a paper in 2024, titled ‘Environmental Sustainability and AI in Radiology: A Double-Edged Sword’, which considered AI’s potential to improve environmental sustainability in medical imaging, alongside its huge need for energy and water.
They found that the training phase of generative AI tends to be very energy-intensive — for example, just one AI model’s training was estimated to emit more than 626,000 kilograms of carbon dioxide equivalents, nearly five times the lifetime emissions of an average passenger car, including the manufacture of the car itself. That phase makes up 10 percent of AI energy use, she said.
The remaining 90 percent comes from “the inference phase”, the everyday ChatGPT-type requests by people and professions across the globe and those instances where AI models are used to make predictions.
Doo referred to statistics from the EarthDay.org movement during her presentation. Croakey’s searches of this organisation’s site found statements that ChatGPT gets roughly 2.5 billion prompts per day, with each requiring computation inside the data centres, where servers run continuously to process, store, and generate responses.
While a single AI query may feel insignificant to us, the environmental organisation says that “everyday use adds up quickly at scale”.
Each prompt uses an estimated 0.34 watt-hours of electricity, which over a day of average use can add up to 6.8 watt-hours per person, it says.
Again, that measure may seem minimal on an individual level, but scaled to one million daily users, that becomes 6,800 kilowatt-hours – enough to power roughly 225 US homes for a day.
At 100 million users, it jumps to 680,000 kilowatt-hours daily, comparable to the electricity use of around 22,000 households. “Small individual actions, multiplied globally, begin to resemble the energy footprint of the entire community,” it said.
Of course it’s not just our use of ChatGPT that is fuelling this energy problem.
A paradox or two
As is widely known, healthcare systems are major contributors to the global climate crisis. In both the US and Australia, they account for around 8-10 percent of total national greenhouse gas emissions.
Medical imaging has a heavy carbon footprint on its own, estimated to account for up to one percent of global carbon emissions, Doo said.
This footprint means that the “climate-driven diseases increasingly filling our scanners (ie, heat-related illness, the respiratory burden of wildfire smoke, the expanding geography of vector-borne infections) are amplified by the very emissions that health care produces”, she has written.
In her RANZCR presentation, titled ‘Sustainable Intelligence – Is Your AI Worth Its Watts?’, Doo said AI has the potential in many ways to improve that environmental cost.
It may be able to shorten MRI scan times, improve the scheduling efficiency of scanners, and optimise the use of decision-support tools to reduce low-value imaging, she said.
Doo and colleagues have written that about two-thirds of CT energy use occurs in a nonproductive idle state, while a third of MRI energy use happens during the system-off state.
AI models and tools can potentially reduce that idle time and monitor other energy-consuming devices in radiology departments, such as picture archiving and communication system workstations, with automated shutdown when not in use to minimise nonproductive energy consumption, they said.
But therein lies the first conundrum, known as the Jevons paradox, where making something more efficient can increase how much of it gets consumed, she said.
And there’s the other huge consumption paradox, she said, where the “explosive increase” globally in radiology’s AI use, along with the energy and water needed to power data centres, may negate or exceed any savings in clinical energy use.
She told the conference it is crucial for radiologists to grasp the “dual nature” of AI, so they can make informed decisions and develop strategies to maximise its positive contributions while mitigating its environmental drawbacks, such as its massive energy use.
Energy and water costs
As Deepcut News recently reported, public opposition to resource-hungry data centre construction is widespread worldwide and growing in Australia, where there are already at least 162 operating, with another 90 in the works.
A recent report from Greenpeace Australia Pacific dubs data centres “energy vampires” and urges a moratorium on new facilities until governments develop appropriate regulations and safeguards.
The Sydney Morning Herald earlier this year reported [paywall] that, if every planned data centre in New South Wales is built, the combined maximum power demands in western Sydney would climb to about 4.4 gigawatts in a decade — equivalent to the average electricity load of more than 10 million households.
At their peak data centres would demand almost four times as much power as the rest of the city, it said.
On the global scale, the International Energy Agency (IEA) reports that data centres consumed about 415 terawatt hours (TWh) of electricity in 2024, about 1.5 percent of global supply, growing at about 15 percent a year over the last five years. That figure is projected to nearly double to 945 TWh by 2030.
Their impact on water is just as critical. Doo said data centres are “incredibly thirsty”, as they rely on water-based cooling systems to prevent servers from overheating.
She referred to a World Economic Forum article last year which reported that accelerated AI adoption alone could result in an additional 4.2 to 6.6 billion cubic metres of water withdrawal by 2027, including onsite cooling and offsite electricity generation.
This would be equivalent to four to six times the annual water withdrawal of Denmark, it said.
At the local level, a single AI-heavy household’s daily energy use can require the equivalent of a person’s entire daily drinking water, Earthday.org says.
In effect, it says, data centres now “drink” as much water each year as a mid-sized city.
Doo warned that the consequences of this are more severe in arid regions and may put residents at greater risk of water restrictions, rising utility costs, and heightened vulnerability during drought conditions.
For example, she pointed to a 2024 Virginia legislative report which found that the state’s typical residential electricity bill could rise by US$14-37 per month by 2040 because of grid strain tied to data centre growth – a 9-24 percent increase over current average bills.
Heating the planet
Now there’s a new concern, with a recent study led by Cambridge University finding that data centres don’t just use large amounts of water and electricity, they’re heating the environment too.
The study found that land surface temperatures around AI data centres rise by an average of 2 degrees Celsius, with some areas as high as 9C (16.2F) — a rise the researchers call “the data heat island effect”.
They say this could affect more than 340 million people globally and have “a remarkable influence on communities and regional welfare in the future”.
All of this will, Doo said, open up questions about AI access and equity.
“Is intelligence a public good when everyone needs it to function theoretically, or is it something that we all need to pay for and we’ll have different scales of how much people can access, and how much is that affecting the planet?”
Greatest gain, least harm
Doo highlighted a range of initiatives in radiology that have sought to deliver better care in different ways: the ALARA (As Low As Reasonably Achievable) principle, which encourages low-dose diagnostic imaging procedures; the Image Gently Alliance which advises how to improve radiation management for children; and Image Wisely, which operates on similar principles for adults and children.
What they have in common is thinking about limiting the amount of radiation needed to get the clinical result required while not harming patients, to “do the most good for the least harm”, she said.
That’s what she’s urging from her profession now with AI, to be asking about the hidden costs.
Doo is arguing for sustainable intelligence, the need to think about AI from a system level perspective, from a workflow perspective, and to be able to innovate to do AI better.
Her message to colleagues is: “Consider how you’re using it, both from the sustainability lens, but also… what skills you want to preserve, and what you would like to teach your future generations coming up behind you.”
Doo recently led research on developing a practical, lifecycle-based framework centred on patient safety, called translational bialignment. This concept pairs regulatory science requirements (what AI systems should deliver to clinicians and patients) with implementation science capabilities (what institutions should provide for safe deployment of AI), she and her colleagues wrote.
“This framework addresses the full AI lifecycle, from data stewardship and model development to validation, deployment, and monitoring, and articulates shared responsibilities for vendors, institutions, and clinicians grounded in trustworthy AI principles,” they said.
She has also detailed many ways the profession can minimise their contribution to AI energy and water costs, including through a Machine Learning Emissions Calculator, which considers type of hardware, duration of training and geographic region. She has also called for multi-institution collaboration on centralised data sharing to lower AI costs.
This sustainability mission is why she sticks with old fashioned slide shows at conferences like #Intelligence26, despite the temptation of getting some AI help to jazz them up.
She told participants that she tries to use AI “very judiciously”, a small but deliberate measure of the amount of work she does looking at the massive environmental and social cost of AI in medicine.
“Just like how you decide to recycle, at the end of the day I would like you to think about how you’re using AI day to day, so that we’re not completely killing our planet at top speed,” she said.
Also watch
Bookmark this link to follow Croakey Conference News Service’s coverage of #Intelligence26. Our final report will be published next week.
Add Comment