A Call to Address the Generative AI-Environment Paradox in Graduate Medical Education

Abstract

Visual Abstract.

A Paradox

Generative AI has been hailed as a positive catalyst for revolution in medical education and health care.^1,2 At the same time, systems vital for human life and health are faltering due to compounding environmental crises.^3,4 While AI has the potential to support sustainability initiatives, including in health care, rapid uptake of AI is accelerating drivers of climate change and climate-driven health risks.^3,5-7 Large language models (LLMs; eg, ChatGPT, Copilot, Gemini), multimodal LLMs, and other generative AI systems require data centers that consume massive amounts of water and energy.^8-11 Demand for this AI is fueling rapid expansion of data centers—and their associated resource consumption and greenhouse gas (GHG) emissions.^6,12

Confronted with both realities, graduate medical education (GME) must simultaneously prepare trainees for an ongoing climate crisis and widespread use of AI. How can GME reconcile this innovative technology’s potential to improve health with its deleterious effects on the natural systems on which human life depends?

Currently, AI discussions in the medical literature center on privacy, confidentiality, efficiency, and accuracy. GME concerns also include mis-skilling, de-skilling, and never-skilling.^13,14 As yet, environmental and patient health concerns do not appear to be key factors in selecting or employing these tools.

In a 2024 American Medical Association survey, two-thirds of physicians—a 78% increase from the prior year—used AI.¹⁵ The most common uses of AI among those surveyed were note documentation, other charting, billing and coding, and the creation of discharge instructions.¹⁵ Many medical students, GME trainees, and faculty query LLMs daily. Some uses are intentional, driven by individual curiosity or pursuit of efficiency. Other uses are more insidious, as AI tools are integrated into electronic medical record systems.^16,17 Most trainees and faculty likely use LLMs without active consideration of their environmental—and ultimately human health—costs.

Studies show that GME training is a powerful force in shaping future practice habits, ethics, and professional identity. GME could provide guidance and guardrails for residents and fellows balancing AI’s potential to improve versus harm human health.¹⁸ This moment of rapid adoption of AI calls for GME’s collective creativity to challenge assumptions and engage meaningfully with the AI-environment paradox.

Environmental Health Considerations

To date, AI expansion has far outpaced efficiency gains, ability to rely on renewable resources, and resource conservation strategies, which affects the health of patients and communities.^19-21 In 2023, data centers consumed 4.4% of US electricity, and they are expected to consume nearly 12% of US electricity by 2028.²² Scientists agree that rapid decrease in GHG emissions is essential for avoiding catastrophic climate-related consequences, yet AI data centers undermine global efforts to curb GHG emissions.²³ Unfortunately, precise data around LLM energy use and GHG emissions is not readily available,^24,25 which impedes productive conversations about tradeoffs of AI use in health care.²⁶

Data centers also require immense amounts of water for cooling and electricity generation. In 2021, data centers were among the top 10 water-consuming industrial or commercial industries in the United States, with a medium-sized data center requiring the same amount of water as 3 average-sized hospitals.^27,28 Data centers are compounding water scarcity problems in drought-prone areas such as the US Southwest. Scientists forecast that by the end of the decade US data centers will imbibe as much water as 10 million Americans.²⁹

From rural farmlands to cities, data centers generate concerns around environmental justice, energy affordability, grid failures, water shortages, and local pollution.^23,30 The Electric Reliability Council of Texas forecasts that the state’s power grid could fail in the summer of 2026, in its most extreme scenario.³¹ In response, lawmakers created legislation authorizing the grid to disconnect data centers at times of energy crisis.³² In Newton, Georgia, just 1000 feet from a Meta data center, an elderly couple’s water supply ran dry months after the data center construction began, and the entire community’s well water is facing future shortages.³³ Data centers also create noise and air pollution. Many data centers utilize backup generators fueled by diesel which produce particulate matter and nitrous oxide pollution. These emissions can trigger health symptoms even with short-term exposure and have well-documented negative implications for human health.³⁴ In addition, power blackouts threaten access to critical medical equipment required by patients. Or, in hotter summers and with increasing energy costs, those with lower incomes may risk choosing between food or medicines and air conditioning.

Globally, data centers are shaping foreign policy and contributing to an aggressive pursuit of energy and natural resources. While data centers promise economic development, they raise local and immediate environmental and health concerns.³⁵ In Querétaro, Mexico, Microsoft’s data center region uses direct outdoor air for cooling throughout 95% of the year. However, cooling in the remaining hottest days of the year requires water, which led to scarcity during the city’s worst drought in a century.³⁶ In Ireland, 20% of the country’s electricity is consumed by data centers. In early 2025, storms caused power outages across western Ireland, which exposed stressors on the country’s electricity infrastructure.³⁷ Populous areas in over a dozen countries, in all continents, bear a disproportionate weight of the negative implications of data centers.³⁷ The resource extraction by affluent countries and regions, from those with lower wealth and poorer health outcomes, mimics colonial legacies³⁸: a kind of modern-day colonization by the cloud.

While most of the narratives to date stem from media reports and white papers, these environmental justice case examples prompt both ethical and practical patient care concerns that should be discussed among GME trainees, faculty, and sponsoring institutions.

Considerations for the GME Community

AI is rapidly becoming embedded in GME.³⁹ Programs are adopting AI for diverse tasks, such as evaluating residency applications,⁴⁰ predicting performance on essential clinical tasks,⁴¹ and assessing feedback quality.⁴² At the same time, applicants increasingly use AI to prepare residency and fellowship applications,⁴³ and trainees use LLMs to support clinical decision-making and create patient education materials.⁴⁴ Further, ambient AI scribes and AI-drafted patient portal messages may soon become commonplace in GME.⁴⁵

The life cycle of an AI model typically includes development, training, implementation, inference, and de-implementation. Currently, trainees’ and programs’ interactions with AI are most likely to occur during the inference phase, in which end users use AI outputs to perform a specific task (eg, a LLM to identify drug-drug interactions or create a patient note). While inference typically requires lower resource utilization than development and training, the frequency of inference tasks translates into substantial energy consumption and GHG emissions. For example, Ramachandran and colleagues estimated that an AI-based portal messaging tool at a large academic health center would produce approximately 48 000 kg of CO₂, equivalent to the amount of CO₂ sequestered by 2300 trees in a year.⁴⁶

However, GME isn’t a purely educational endeavor. Residents and fellows are employed within complex health systems that shape trainees’ AI exposure and use.⁴⁷ In the AI era, environmental stewardship and sustainability should be among the metrics that institutions and GME programs collaboratively address to ensure responsible and judicious use of AI-based clinical and educational tools.

The University of Michigan was the first university in the world to provide a custom suite of powerful generative AI tools (U-M GPT) to all members of its community.⁴⁸ U-M GPT enables the institution to monitor and estimate the tool’s energy use and the resulting carbon footprint.⁴⁹ Examples like U-M GPT highlight the potential to monitor energy use, at resident/faculty, GME program, and organizational levels, and to provide meaningful feedback about usage patterns and environmental impacts. These data allow the community to make informed decisions about AI use.

In terms of AI uses to enhance sustainability efforts, studying how AI-based clinical decision supports might identify non-inferior, resource-efficient treatment pathways versus more wasteful pathways could provide opportunities for scholarship as well as improved environmental stewardship. GME should have a seat at the table for institutional discussions that compare AI products—hardware, software, and data—regarding waste, efficiency, system maintenance, and reusability of resources.

Call for Articles and Potential Directions

The Journal of Graduate Medical Education (JGME) December 2024 supplement issue, on climate and GME, explored potential strategies to enhance GME trainees’ skills to address patients affected by climate changes and health system waste and sustainability efforts.⁵⁰ JGME has since recruited an editor to explore best practices in AI use, for its AI Teaching Rounds section. The critical yet challenging intersection of AI, GME, and environmental risk requires more study; JGME encourages educators and researchers to submit their work, in any JGME category. Questions for our community include the degree to which environmental and ethical concerns are considered in governance structures for AI use at academic institutions and teaching hospitals, and if and how GME programs are providing a foundation for these ethical concerns (Box). Other avenues for creativity may include consideration of AI in existing frameworks, such as the Planetary Health Report Card for GME,⁵¹ or exploring how the GME community can join calls for more AI transparency regarding effects on communities, GHG emissions, water and energy use, and sustainability efforts, to aid informed AI tool choices.^26,52

Box.

Questions for GME on the Generative AI–Environment Paradox

• How do residents, fellows, and faculty currently approach the AI-environment paradox?

• What are GME educational gaps and curricular opportunities around the AI-environment paradox?

• What tools are needed for trainees and programs to measure the environmental effects of AI deployed in GME settings?

• How can the GME community assist in seeking transparency around the resource intensity of health care–related AI tools?

• How can discussions about AI tool conservation be included in program or institutional decisions about AI use?

• How is AI used currently to improve waste reduction and sustainability efforts in GME training settings?

• How could AI be included in GME curricula on medical ethics, aside from privacy and confidentiality concerns?

• How can systems provide meaningful feedback on health care delivery–associated energy use and environmental impacts at the individual, program, and organizational levels?

Abbreviation: GME, graduate medical education.

From the paradox of potential AI benefits versus environmental and health risks, we hope the GME community will coalesce around principles for responsible AI use to support patient health today without degrading patients’ health tomorrow.