Berenji recommends organizations treat the environmental implications of AI as a dimension of value alongside safety, quality and cost. That means asking what happens when a model runs at scale or if a similar benefit can be achieved with fewer resources, in addition to looking at a model’s clinical benefits.
That said, an AI model’s downstream impact on the environment isn’t always easy to quantify. Wadhwa uses oncology models as an example: Identifying cancer in stage 1 as opposed to stage 4 means treatment is less frequent and less intensive, which reduces its environmental footprint. The same is true for imaging models that eliminate the need for a second study — and a patient’s second car trip to the hospital.
7 Ways Healthcare Organizations Can Adopt AI Sustainably
- Use the right model for the job. “Use the smallest model that meets the clinical performance bar,” Berenji says. High-end models built for research and complex problem-solving, such as Claude Fable, ChatGPT Pro or Gemini Pro, are wasteful when they’re applied to word processing or answering easy tasks.
- Ask vendors what they’re doing. “Procurement is the biggest lever most health systems have,” Berenji notes. Many organizations have already used this leverage to reduce physical waste or exposure to toxic chemicals. Leaders should ask vendors about energy use per inference, water use for cooling, and where workloads are hosted. Health Care Without Harm’s sustainable AI procurement guide offers a good starting point, she says.
- Design for efficiency. If a patient is in the hospital for several days, continually pasting their information into an AI-powered scribe generates a lot of redundancy. Eliminating that before dropping records into a model can cut token usage up to 40%, Wadhwa says. Berenji adds, “Don’t run AI on every encounter by default when it’s only useful for some.”
- Bring discipline to AI use. Organizations can run nonurgent workloads such as model retraining or batch analytics at times when demand is low or clean power is available, Berenji says. Upgrade cycles should account for cooling efficiency, energy use and the proliferation of electronic waste in addition to hardware performance.
- Measure and evaluate. It’s imperative to keep track of where AI models are running. Organizations should decommission tools that aren’t delivering value, Berenji says. Pilots that never get off the ground are a hidden source of wasted resources and contribute to unnecessary emissions.
- Reposition AI as a cost issue. The environment isn’t always top of mind for healthcare executives, Wadhwa says. However, money is a priority. “If you frame environmentally responsible use of AI as a cost-savings measure, then it’s a different conversation. Leaders come to that realization that tokens cost money and should be used more effectively.”
- Collaborate. Many organizations use AI models to bring efficiency to waste management, energy use and supply chain optimization. That’s all well and good, Wadhwa says, but it means leaders miss opportunities to learn from each other and reduce emissions in the process: “Is every health system trying to solve the problem itself? If one hospital solves the problem, will the next use more tokens to do the same thing?”
EXPLORE: Here are six ways embracing sustainability reduces enterprise risk.
What Sustainable AI in Healthcare Looks Like for the Future
Decisions made in the first three to five years of electronic health record implementation laid the foundation for how EHRs are used, Wadhwa notes. Today, organizations can make minor changes, but the major elements are already in place.
Healthcare is in the same place today when it comes to AI, she says. The industry is uniquely positioned to get things right. Roughly 75% of patients believe there are benefits to using AI in healthcare, so public support is there. Plus, the industry represents close to 20% of U.S. gross domestic product, so its weight can put pressure on vendors.
“Hospitals are seeing the frontline impact of climate change,” says Wadhwa, who was at Kaiser when the health system felt the effects of Los Angeles-area wildfires and the Pacific Northwest heat dome. “Action is the antidote to despair, and we have to be part of the conversation.”
For Berenji, two important inaugural steps are tying AI governance to an organization’s sustainability commitments and elevating sustainability to a senior or C-level role. From there, she hopes to see the environmental implications of using any healthcare product, not just an AI model, “become as routine a question as accuracy or bias.”
Berenji says AI can be a net contributor to the planet’s health — if models run using clean energy, only where they add value, and focus outputs on ways to “decarbonize” care delivery through greater efficiency and less waste. “Healthcare's first principle is to do no harm, and that has to include harm to the environment our patients live in,” she says.
