Some students and educators ask whether artificial intelligence (AI) in medical education is bad for the planet. It's a fair question. AI runs in data centres that use a lot of energy, as well as water to cool them, and overall use is growing. Data centres, which run much more than AI, used about 1.5% of the world's electricity in 2024 (International Energy Agency, 2025).
The answer depends on the tool and what it is used for. Three tests help to judge it. Is it necessary, giving students practice they would otherwise go without? Is it proportionate, using no more energy than that practice needs? And how does its footprint compare with other ways of practising, with one person's wider footprint, and with health care as a whole?
Is it necessary?
Countries are training more doctors to meet demand for care. Across the OECD, new medical graduates per 100,000 people rose by more than 70% between 2000 and 2023 (OECD, 2025).
Clinical teaching hasn't kept pace with that growth. In the United Kingdom, the share of doctors involved in education stayed at about 16% from 2017 to 2022, and that group is ageing. Over the same period, clinical workload increased, medical student numbers rose by 6% and the number of doctors in training by 20% (Macdougall et al., 2024, BMJ). In the United States, medical school enrolment rose 52% between 2002 and 2019, and some schools have struggled to find enough placements and supervisors (Howley, 2020, AAMCNews).
Consultation skills are learned by doing consultations and getting feedback. Students get that practice on placements and with simulated patients, and both depend on clinicians' time. When that time is shared among more students, each student is likely to get fewer observed consultations than they need.
AI practice can fill part of that gap. A student can work through a consultation with an AI patient at any hour and get feedback straight away. In a randomised trial of 251 fourth-year medical students, those who practised 16 cases with AI patients and automated feedback scored higher in their clinical skills exam (OSCE) than those who didn't (Lavigne et al., 2026, NEJM AI). We set out the wider case in deliberate practice in medical education.
Nobody is proposing that AI replace in-person clinical teaching. Where AI gives students practice they would otherwise go without, it meets the first test.
Is it proportionate?
AI uses energy in two stages. Training a model is a large, one-off cost. Inference, which means running the model each time someone uses it, is an ongoing cost. At Google, inference accounted for about three-fifths of machine-learning energy between 2019 and 2021, and training for two-fifths (Patterson et al., 2022, Computer). A practice tool usually runs a model that someone else has trained, so most of the energy it adds comes from inference. A proportionate tool keeps that energy to what the job needs.
How much energy each request uses varies widely. Google reports that the median text prompt to its own AI assistant uses 0.24 watt-hours (Wh), less energy than watching television for nine seconds (Elsworth et al., 2025, arXiv). An independent benchmark of 30 models found that, for the same long prompt, the most energy-intensive models used over 65 times as much energy as the most efficient (Jegham et al., 2025, arXiv).
Much of that spread comes from how models are designed. In a study comparing models across many tasks, researchers found that "multi-purpose, generative architectures are orders of magnitude more expensive than task-specific systems" (Luccioni et al., 2024, ACM FAccT). Reasoning models use on average 30 times more energy than models that don't reason, mostly because they generate far more text (Hugging Face, 2025).
In education, the knowledge a learner needs can be set out in advance, so a teaching tool can use a model suited to its task. On the evidence above, that can take far less energy than a large general-purpose or reasoning model.
How does it compare?
In-person teaching has a footprint too. A study in a hospital in England counted the equipment and consumables used in teaching. Simulation teaching produced a median of 42 grams of carbon dioxide equivalent (CO2e) per learner per hour, and bedside teaching 118 grams (Jack et al., 2025, The Clinical Teacher). Those figures leave out travel, which can add far more.
The clearest travel figure for consultations with standardised patients comes from an exam rather than practice. When the United States ran its clinical skills licensing exam in test centres, travel produced about 130 kilograms of carbon dioxide per student (Sherpa et al., 2023, Yale Journal of Biology and Medicine). Local practice sessions involve shorter trips, but they still bring students, actors and tutors to one place. Travel is likely to be where in-person and AI practice differ most.
For an individual, everyday AI use is a small part of their footprint. The data scientist Hannah Ritchie uses a deliberately high estimate of 2 to 3 grams of carbon dioxide per AI query. On that estimate, ten queries a day for a year add about 11 kilograms. That is 0.16% of the average UK footprint of about 7 tonnes (Ritchie, 2025). Across many users, small amounts still add up, and Ritchie also treats AI's total energy demand as a fair concern. Medical students are a small group, so their practice is a very small share of all AI use.
Health care as a whole is a large emitter. Health Care Without Harm and Arup put it at 4.4% of global emissions, and noted that "if it were a country, it would be the fifth largest emitter on the planet" (Health Care Without Harm and Arup, 2019). Most of those emissions are accepted as necessary, because people need care. Hospitals and pharmaceuticals are the largest sources. In Australia, where health care produced 7% of national emissions in 2014/15, hospitals made up 44% of that and pharmaceuticals 18% (Malik et al., 2018, Lancet Planetary Health).
Doctors' decisions drive much of that footprint: who is admitted, how long they stay, which tests are ordered and what is prescribed. How doctors practise is shaped partly in training. In the United States, family doctors and general internists who trained in high-spending regions went on to spend more on their patients. In their first seven years of practice, they spent 29% more than those who trained in low-spending regions (Chen et al., 2014, JAMA). Health care carbon estimates, including those above, are built from spending data, so care that costs more tends to produce more emissions.
Doctors who reason well may keep hospital stays shorter, order fewer investigations that won't change management, and prescribe fewer drugs that won't help. No study has measured the carbon effect of better training. But because hospitals and pharmaceuticals are the largest sources, a small change in how a doctor practises over a career could significantly outweigh the emissions from their AI practice during training.
Where Gestalt fits
Gestalt builds consultation skills through repeated practice. Like any software that uses data centres, it has a carbon cost.
Its architecture is knowledge-constrained: the models work from the Gestalt knowledge graph, a curated body of clinical knowledge. We built it this way because institutions need accuracy, clear provenance and localisation before they can trust AI in teaching. The same design also makes the system more efficient.
We describe our approach as pro-human AI: AI built to extend what people can do and leave them more skilled. Being pro-human also requires awareness of AI's environmental cost.