AI Tools in Medical Education and Health Care: Climate Impact and Sustainable Practices

When first-year MD student Amy Edziah started at Yale School of Medicine six months ago, she found herself using artificial intelligence (AI) tools to enhance her learning process and concept understanding. However, she was concerned about how these tools might cause harm to the environment and low-income communities where data centers tend to be located.

"How do I reconcile the environmental impacts of using AI while also finding the tools useful to my studies? I don’t want to contribute to the harm these tools may cause to my future patients, but I need to learn how to take care of them and AI tools might even help me become a better doctor.”

Edziah posed her question to Jaideep Talwalkar, MD, professor of internal medicine and pediatrics and associate dean for educational technology and innovation (ETI), and learned she was one of many facing a similar conundrum.

Inspired by Edziah’s question, Talwalkar and the ETI team pulled together a panel of experts from across Yale University and Yale New Haven Health to discuss the climate impacts of AI use in health care and medical education.

The event took place on March 3 in the Cohen Auditorium and featured panelists Yuan Yao, PhD, associate professor of industrial ecology and sustainable systems at Yale School of the Environment; Connie Steel, PhD, program manager for artificial intelligence initiatives in the Yale Office of the Provost; Allen Hsiao, MD, FAAP, chief health information officer at Yale School of Medicine and Yale New Haven Health; and Jennifer Frederick, PhD, associate provost for academic initiatives at Yale.

Moderated by Talwalkar, the panelists provided insight into the conversations happening in their professional spaces about the climate footprint of AI.

For Yao and her colleagues, the focus has been trying to assess the life cycle environmental impacts of AI systems.

“How can we quantify the environmental impact of AI? It’s much more than electricity and water consumption. We must try to understand the full life cycle of the product, from the critical minerals and materials used to manufacture the hardware, to the environmental impacts associated with building and operating data centers, to the electronic waste generated at the end of the product’s life cycle. There’s an immense challenge in fully understanding AI’s environmental impact as the supply chain is distributed across many regions and companies do not publicly share detailed manufacturing and supply chain information.”

Yao continued that being a responsible citizen means being aware of how our choices and patterns of consumption can impact communities beyond where technologies are used.

Electronic waste, for instance, can be exported from higher-income countries to developing countries where recycling infrastructure and environmental protections may be limited. In some cases, discarded electronics are processed through informal recycling practices, such as burning, to recover valuable materials.

These activities can release toxic substances into the air, soil, and water, creating health and environmental risks for local communities.

She encourages people to be thoughtful in their use of AI, similar to how they would when considering a new purchase. “Ask yourself, do I need AI to complete this task or am I using it just because it’s there?”

Hsiao shared that clinicians are continually seeking tools to help them manage their workflow, improve care, and boost efficiency. For many, that means using AI tools such as Ambient AI, an AI-powered virtual scribe which records and synthesizes a patient-doctor conversation, allowing the doctor to focus their attention more fully on the patient during an exam instead of writing notes.

Some clinicians choose not to utilize these tools due to concern of climate impact. Hsiao noted that while it does take a lot of energy to train a large language model (LLM), once trained the electricity needed to generate one ambient AI note is equivalent to powering a 60-watt lightbulb for 18 seconds.

If there’s an opportunity for physicians to provide better care because of AI, he said, that could mean fewer diagnostic tests, fewer materials used, less discomfort for the patient, and faster, more accurate diagnoses.

Eventually, those efficiencies could lead to large-scale positive impacts, like reducing the size of a hospital’s HVAC system (which can total 40–50% of a hospital’s energy consumption) because fewer beds are needed.

Frederick’s conversations on AI have focused on how Yale can minimize the harm done to the environment while still remaining a leader in the field. To accomplish this, Frederick and her colleagues have been assessing usage on Yale’s AI tool, Clarity.

One of the biggest energy-saving changes they made, says Frederick, was switching Yale’s energy consumption subscription from a PTU (provisioned throughput unit) model to one that more accurately reflects actual usage. With this change alone, they were able to reduce the energy needed to power Clarity by a factor of 10.

So far, Clarity’s annual energy usage has equaled about one year of electricity for three households, and going forward it will be considerably less.

Yale is also making more climate-responsible choices, such as moving the university’s data centers into one consortia space shared with five other universities.

The space, known as the Massachusetts Green High Performance Computing Center, is in Holyoke, Mass. and is the first university research data center to achieve LEED Platinum Certification. The center uses innovative methods for significantly reducing cooling and energy consumption.

Frederick also remarked that Yale can and has been putting pressure on vendors to disclose where and how AI is being used in their products and its associated energy consumption.

For Steel, she has seen a notable increase in vibe coding during the last year. Vibe coding is a development method that uses natural language prompts to guide AI assistants in building and iterating applications.

This means application development is no longer limited to a small number of coders with highly technical skills. With vibe coding, anyone who can prompt an AI assistant can build an application.

While this opens the door to projects that might otherwise have never come to fruition, she points out that the upward trend in vibe coding may indicate increasing numbers of coding projects which each have their own energy demands.

She encourages vibe coders to explore the conversations happening in computer science spaces relating to computational efficiency and green coding. Users can also ask AI assistants to improve their code’s efficiency by prompting “Can you help green my code?”

To lessen the environmental impact, she recommends considering a pre-trained LLM that is fit for purpose. The right pre-trained model can significantly lower energy use, speed up project delivery, and be more cost-effective than building a model from the ground up.

At Yale, Clarity offers six LLM models for users to choose from: GPT-4o, GPT-4o mini, Claude Sonnet 4.5, Gemini 2.5 Flash, Gemini 2.5 Pro, and o3. Among these, Steel noted that Gemini 2.5 Flash and GPT-4o mini are more energy efficient and are under evaluation to replace GPT 4o as the default model.

For users interested in their energy usage in Clarity, Steel noted that tokens are a proxy for the computational time and energy needed to generate an answer.

Going forward, Steel encouraged attendees to consider how AI assistants can be prioritized to improve patient care. “Think about how you can take advantage of an AI teammate’s special skills and talents while also recognizing its weaknesses.”

Like Hsiao, Steel sees the potential of AI in reducing a health system's carbon footprint through more accurate and efficient care.

“A health care system is a large ecosystem, and diagnostic errors or delays happen in 10–20% of patient encounters. If you can use AI to lower that number, there’s potential not only to improve the lives of your patients but to reduce low value testing, medical and material waste, and power consumption.”