Sustainable AI in Healthcare: Cutting Its Carbon Cost

Healthcare produces about 4.4% of the world’s net greenhouse gas emissions, according to a 2019 estimate from Health Care Without Harm and Arup. The AI tools now entering hospitals run on data centres whose electricity use the International Energy Agency expects to more than double by 2030. That overlap is why sustainable AI in healthcare has become a practical question for hospital IT and procurement teams, not just researchers.

Sustainable AI in healthcare means using AI where it improves care or cuts waste, while measuring and limiting its own energy, water and hardware footprint. The evidence is thinner than the marketing. A January 2026 review found no standard way to measure it, and the research on AI sustainability in healthcare is still young. Below, you’ll find the scale of the problem, what the studies actually measure, one place AI is already saving energy, a six-step checklist and the gaps that remain. Figures are as of November 2, 2026, and each one is dated.

Why Healthcare Sustainability and AI Now Collide

The 4.4% figure equals about 2 gigatons of CO2-equivalent, based on 2014 data. Roughly 71% of it comes from the supply chain, such as manufacturing, medicines and equipment, and 17% from facilities and vehicles. In England, the NHS accounts for about 4% of national emissions and has set a net zero target of 2040 for its own footprint.

Data centres are the other half of the picture. The IEA estimates they used about 415 terawatt hours in 2024, roughly 1.5% of global electricity. Its base case reaches about 945 terawatt hours by 2030, just under 3%. Electricity demand from AI-optimised data centres is expected to more than quadruple over that period.

Globally, that is a modest share of electricity growth, less than 10% by 2030. The pressure is local, though. In the United States, data centres account for nearly half of expected demand growth.

What Does Sustainable AI in Healthcare Actually Measure?

A scoping review in the European Journal of Radiology (January 2026) found 13 studies on AI’s environmental impact in radiology. They clustered around four themes: energy use, computing resources, carbon footprint and water. The review found no standardised method or comparison across studies. Headline figures often leave out the repeated development work before the final training run, and the energy used when a model runs on real patients is rarely reported. A January 2026 commentary in Nature Reviews Nephrology adds electronic waste and ethics to the list.

One CT study shows how AI fits into the larger picture. It put a CT workflow with cloud-hosted AI at 11.32 kg of CO2e per exam, a figure that covers the scanner, consumables and AI together, so AI is only part of it. The same study estimated that running the AI on a hyperscale cloud rather than on-premises servers avoided 200 kg of CO2e a year. Its power usage effectiveness was 1.22 for hyperscale data centres against 1.84 for typical enterprise ones. That study is a 2024 summary and covers one workflow, so treat the numbers as an example, not a benchmark.

AI Sustainability in Healthcare: Where It Saves Energy

Imaging gives the clearest case for AI saving energy. Researchers from Sapienza University of Rome presented work at RSNA in December 2025 comparing conventional and AI-assisted MRI sequences. For upper abdomen exams, AI shortened scan time by 58% and saved about 1.39 kWh and 0.57 kg of CO2e per exam. Presenter Tiziano Polidori noted that medical imaging accounts for about 10% of healthcare’s CO2 emissions and that MRI uses the most energy. His group translated its total savings across exam types into 19.4 million kilometres of car travel.

Two cautions apply. These are conference results, not yet published in a peer-reviewed journal. And savings per exam don’t guarantee lower emissions overall, since shorter scans could lead to more of them. Researchers call this a rebound effect, and it is one reason sustainable AI in healthcare needs measurement over time, not one-off claims.

How to Use AI Sustainably in Healthcare: A Six-Step Checklist

A sustainable AI in healthcare programme starts with simple habits, not new technology.

1. Start with the problem, not the model. If a simpler method meets the clinical need, use it.
2. Ask vendors for numbers. Request energy per inference, how often the model is retrained, and where it is hosted. No standard exists, so ask how each figure was calculated.
3. Choose hosting on purpose. The 1.22 versus 1.84 efficiency gap above matters, but so does vendor lock-in, which the same study flagged.
4. Measure the whole workflow. Track exam time, energy use and retraining before and after rollout, as the Rome team did for MRI.
5. Be sceptical of “carbon-neutral” claims that rely on certificates or offsets. A 2025 position paper argues those claims hide real growth in data-centre demand.
6. Plan for hardware and data. Servers and storage become electronic waste, so keep only the data and equipment the service needs.

What Sustainable AI in Healthcare Still Lacks

Standards come first. Without agreed metrics, hospitals cannot compare tools fairly, and vendors cannot be held to a number. Reporting on inference energy is thin, and water use appears in only a handful of studies. A systematic review on green AI in health (Artificial Intelligence in Medicine, August 2026) identifies three themes for action: green computing practices, energy-efficient models and renewable energy in healthcare settings.

Rebound effects are the hardest problem. The same position paper notes that efficiency improvements haven’t stopped total AI energy use from rising. For healthcare sustainability teams, that means a hospital can run a leaner AI tool and still see its footprint grow if usage expands faster than efficiency improves.

Conclusion

Sustainable AI in healthcare is less about a single green technology than about measuring the right things: whole workflows, retraining, hosting and hardware. Some tools, like AI-assisted MRI, show real promise, but the studies are early, and the numbers are not standardised. Hospitals can start now by asking vendors for energy data and tracking results before and after deployment.