Healthcare AI also has a body

Healthcare AI is usually presented as a promise of precision. Better diagnoses, better risk classification, better management of resources. Yet that promise arrives at a moment when health is already shot through with climate crisis, inequality, institutional precarity and disputes over the recognition of suffering.

This radar weaves together three threads. First, planetary health, drawing on the analysis by Catherine Clare Morneau and her team in The Lancet Planetary Health on the integration of health into national climate adaptation policies. Second, sustainable AI, with work on the carbon footprint, energy and resources of deep learning models applied to medical imaging. Third, medical anthropology, which has spent decades reminding us that falling ill is not only about producing clinical data, but about inhabiting a vulnerable body within a concrete social world.

The question is no longer only whether medical AI gets things right. The question is which body it imagines, which suffering it recognises, which inequalities it inherits and which material world it needs in order to function.

Health is already inside the climate crisis

The climate crisis is not a context external to medicine. It enters the consultation, the emergency department, mental health, reproductive health, food, displacement and the real capacity of public systems to care. Heatwaves, pollution, vector-borne diseases, food insecurity and extreme events reorganise everyday life and put pressure on health systems that were often already strained.

The analysis by Morneau and her team in The Lancet Planetary Health helps us avoid a narrow reading. This is not only about adding health mentions to climate plans. It is about looking at whether those plans have funding, measurement mechanisms, institutional coordination and real participation from the communities most affected.

In that context, healthcare AI does not appear as a clean solution separate from the planet. It appears as one more technology inside health systems that already have to respond to a material crisis. A digitalised medicine for an unstable world cannot forget that its own infrastructure also consumes energy, water, minerals, hardware and budget.

The material cost of predicting

The work by Raghavendra Selvan, Nikhil Bhagwat, Lasse F. Wolff Anthony, Benjamin Kanding and Erik B. Dam on the carbon footprint of selecting and training deep learning models for medical imaging reminds us of something basic. The environmental cost of a model does not begin when it is deployed. It begins earlier, in the exploration of architectures, the selection of models, the training cycles, the repeated experimentation and the infrastructure that sustains the whole process.

This idea fits the radar’s central critique. Healthcare AI tends to be presented through its visible results, such as a prediction, a segmentation, an alert or a risk classification. Yet behind them sit technical decisions that are also material decisions. Which model is trained, how many times, with what data, on what hardware, with what energy and for what real improvement in care.

Sustainability is not a green ornament added at the end. It is a question about proportionality. A health system should not invest in ever larger models simply because it can. It should ask whether the computational, economic and environmental cost matches a clinical improvement that is significant, accessible and fair.

Falling ill is not only about producing data

Medical anthropology helps to widen the discussion. Disease is not only a localised alteration in the body, something to be measured, classified and treated technically. Falling ill transforms how a person inhabits the body, organises time, sustains relationships, works, asks for help and is recognised by others.

The framework of disease, illness and sickness, developed by Kleinman, Eisenberg, Good, Young and Helman, lets us put this clearly. Biomedicine tends to privilege disease as an objective alteration. The lived experience of suffering adds another layer. The social and institutional dimension shows how an ailment comes to be named, legitimised or called into question. In an increasingly computational medicine, those layers do not disappear. They are rearranged.

Here a decisive concept comes in, the idioms of distress. Mark Nichter proposed it to think about culturally situated ways of expressing psychosocial distress that do not always fit into stable biomedical categories. This idea lets us name something that healthcare AI may struggle to capture. Pain that does not fit, intermittent fatigue, anguish, shame, fear or the sense of not being able to go on are not cultural noise around disease. They can be the concrete form in which suffering becomes speakable.

If healthcare AI learns to read the body above all as image, signal, record or probabilistic profile, it can reinforce an old reduction. The body reappears as a technical object, separated from the experience of the person who inhabits it. Whatever does not produce a stable signal is poorly represented. Whatever does not fit into a variable can fall outside care.

Byron Good is useful for thinking through this point. In chronic pain, the ill body is not an object external to the subject. It is the medium from which the world is inhabited. When pain becomes persistent, it changes the relationship with time, with work, with life projects and with other people. AI can detect patterns, but it should not erase the fact that suffering is also a rupture of the lived world.

Three bodies before the model

Scheper-Hughes and Lock offer a powerful way to read medical AI. Their distinction between the individual body, the social body and the body politic lets us move beyond the idea that the problem is limited to a relationship between patient and machine.

In the individual body, AI translates symptoms, images, risks and clinical trajectories into processable signals. It can help to detect, order and anticipate. Yet it can also lose whatever does not fit neatly into a variable. A person does not fall ill only in their data. They fall ill in their lived body.

In the social body, AI takes part in the production of normality. It does not only identify deviations from a biomedical pattern. It can also reinforce which bodies are considered typical, which populations appear as the exception and which clinical histories prove less legible. A model trained on unequal data can turn prior inequality into a technical criterion.

In the body politic, healthcare AI joins forms of government. It orders priorities, calculates risks, allocates resources, steers interventions and automates suspicion. At that level, sustainability is not only about energy. It is also institutional and democratic. A medicine dependent on large models can concentrate even more decision-making power in those who control data, infrastructure and compute.

Equity does not fit inside a metric

The scoping review by João Matos and his team, published as a preprint in 2025, helps to strengthen the argument. The work reviews fairness metrics in clinical predictive models and reveals a fragmented landscape. There are many ways to measure fairness, but few are designed specifically for health and even fewer are connected to real clinical usefulness.

This matters because an equity metric is not the same as health justice. Measuring performance differences between groups may be necessary, but on its own it does not resolve what counts as harm, which population is poorly described, which clinical histories are deemed incomplete or which forms of suffering are treated as noise.

Here it helps to introduce structural violence. Arachu Castro and Paul Farmer showed how certain health discourses can shift responsibility onto the victims themselves when they ignore poverty, racism, gender, stigma, unequal access to resources and relations of power. That shift can also happen in algorithmic systems. A model can present as individual risk what is in fact sedimented social inequality.

Algorithmic equity runs the risk of becoming a technical reassurance. A system can declare that it has measured bias and carry on running on poor categories, incomplete data or infrastructures that are inaccessible to many communities. In health, justice is not exhausted by calibrating a model. It demands looking at who defines the problem, who appears in the data, who is left out, who benefits and who bears the costs.

Critical Medical Anthropology helps to hold this point without turning the radar into a methodological discussion. Disease and health are social processes. Differences of class, gender, racialisation, territory or access to resources are not external variables added at the end of the analysis. They are part of the very production of falling ill and of healing.

Sustainability also means recognition

To speak of sustainability in healthcare AI is not only to speak of energy, water, emissions, hardware or waste. From medical anthropology, sustainability also means not impoverishing the clinical relationship, not depoliticising suffering and not shifting resources from basic care towards prestige infrastructure.

Here Fassin offers a decisive framework. His work on biolegitimacy lets us think about how certain damaged bodies acquire moral and institutional recognition, while other forms of injustice remain less visible. Healthcare AI can enter that moral economy. It can help to recognise suffering, but it can also impose new conditions for a life to be credible. Data are needed. Evidence is needed. Signals are needed. Categories compatible with the system are needed.

The problem is not that measuring is bad. Without measurement there is no diagnosis, no epidemiological surveillance and no public policy. The problem appears when measurement becomes the only accepted form of existence. Then the suffering that cannot be captured is left in a fragile zone. It can be lived intensely and, even so, never come to be fully recognised.

Veena Das helps to sharpen this closing point. There are pains that fracture language and bodies that hold what cannot become clear testimony. A computationally intensive medicine must be careful with that mute zone of suffering. Not everything that matters appears as clean data. Not everything that matters arrives as a variable.

Caring is not only predicting

Healthcare AI can be clinically promising and, at the same time, socially unequal, environmentally costly and politically dependent. That tension is not resolved by technological enthusiasm nor by automatic rejection. It demands a slower gaze.

Anthropology helps us not to separate the layers. A medical model does not work only on diseases. It works on lived bodies, institutions, categories, waiting, emotions, inequalities and forms of recognition. Nor does it work outside the planet. It works on energy, water, minerals, data centres and budgets.

The medical innovation of the future should not be measured only by the scale of the model. It should be measured by the scale of the care it makes possible. Caring is not only about predicting better. It is also about sustaining relationships, recognising suffering, distributing resources fairly and not building a medicine that consumes world while it promises to save bodies.

Sources

• Morneau, Catherine Clare et al. “Health integration in national climate adaptation policies from 198 countries: a global policy analysis”. The Lancet Planetary Health, 2026.
• Selvan, Raghavendra; Bhagwat, Nikhil; Wolff Anthony, Lasse F.; Kanding, Benjamin; Dam, Erik B. “Carbon Footprint of Selecting and Training Deep Learning Models for Medical Image Analysis”. MICCAI 2022, Springer, pp. 506-516.
• Matos, João; Van Calster, Ben; Celi, Leo Anthony; Dhiman, Paula; Gichoya, Judy Wawira; Riley, Richard D.; Russell, Chris; Khalid, Sara; Collins, Gary S. “Critical Appraisal of Fairness Metrics in Clinical Predictive AI”. Preprint, 2025.
• Nichter, Mark. “Idioms of Distress: Alternatives in the Expression of Psychosocial Distress: A Case Study from South India”. Culture, Medicine and Psychiatry, 1981, 5(4), 379-408.
• Kleinman, Arthur; Eisenberg, Leon; Good, Byron. “Culture, illness, and care: Clinical lessons from anthropologic and cross-cultural research”. Annals of Internal Medicine, 1978, 88(2), 251-258.
• Young, Allan. “The Anthropologies of Illness and Sickness”. Annual Review of Anthropology, 1982, 11, 257-285.
• Helman, Cecil G. “Disease versus illness in general practice”. Journal of the Royal College of General Practitioners, 1981, 31, 548-552.
• Good, Byron. Medicine, Rationality, and Experience. Cambridge University Press, 1994. Spanish edition: Medicina, racionalidad y experiencia. Bellaterra, 2003.
• Scheper-Hughes, Nancy; Lock, Margaret M. “The Mindful Body: A Prolegomenon to Future Work in Medical Anthropology”. Medical Anthropology Quarterly, 1987, 1(1), 6-41.
• Castro, Arachu; Farmer, Paul. “El Sida y la violencia estructural: la culpabilización de la víctima”. Cuadernos de Antropología Social, 2003, 17, 31-49.
• Farmer, Paul. “An Anthropology of Structural Violence”. Current Anthropology, 2004, 45(3), 305-325.
• Fassin, Didier. “Gobernar por los cuerpos, políticas de reconocimiento hacia los pobres y los inmigrantes en Francia”. Cuadernos de Antropología Social, 2003, 17, 49-78.
• Fassin, Didier. La raison humanitaire. Une histoire morale du temps présent. Gallimard-Seuil, 2010.
• Das, Veena. “Lenguaje y cuerpo. Transacciones en la construcción del dolor”. En Arthur Kleinman, Veena Das y Margaret Lock, eds., Social Suffering. University of California Press, 1997.
• Das, Veena. Sujetos del dolor, agentes de dignidad. Universidad Nacional de Colombia / Pontificia Universidad Javeriana, Instituto Pensar, 2008.
• Baer, Hans; Singer, Merrill; Susser, Ida. Medical Anthropology and the World System. Praeger, 1997.
• Menéndez, Eduardo L. Morir de alcohol. Saber y hegemonía médica. Alianza Editorial Mexicana, 1990.
• Singer, Merrill. “A Dose of Drugs, a Touch of Violence, a Case of AIDS: Conceptualizing the SAVA Syndemic”. Free Inquiry in Creative Sociology, 1996, 24(2), 99-110.