Environmentally sustainable development and use of artificial intelligence in health care

Abstract Artificial intelligence (AI) can transform health care by delivering medical services to underserved areas, while also filling gaps in health care provider availability. However, AI may also lead to patient harm due to fatal glitches in robotic surgery, bias in diagnosis, or dangerous recommendations. Despite concerns ethicists have identified in the use of AI in health care, the most significant consideration ought not be vulnerabilities in the software, but the environmental impact of AI. Health care emits a significant amount of carbon in many countries. As AI becomes an essential part of health care, ethical reflection must include the potential to negatively impact the environment. As such, this article will first overview the carbon emissions in health care. It will, second, offer five reasons why carbon calculations are insufficient to address sustainability in health care. Third, the article will derive normative concepts from the goals of medicine, the principles of biomedical ethics, and green bioethics—the very locus in which AI in health care sits—to propose health, justice, and resource conservation as criteria for sustainable AI in health care. In the fourth and final part of the article, examples of sustainable and unsustainable development and use of AI in health care will be evaluated through the three‐fold lens of health, justice, and resource conservation. With various ethical approaches to AI in health care, the imperative for environmental sustainability must be underscored, lest carbon emissions continue to increase, harming people and planet alike.

of AI in a variety of ways, with varying degrees of ease, safety, and ethics. For instance, AI has the capacity to transform health care by delivering medical services to underserved areas, thus improving health outcomes for the poor and vulnerable, 2 while also filling gaps in health care provider availability, thus better meeting the medical needs of low and middle income countries, as well as aging and rural populations. 3 However, AI may also lead to patient harm due to fatal glitches in robotic surgery, 4 bias in diagnosis, 5 or dangerous recommendations, 6 among others.
Despite concerns ethicists have identified in the use of AI in health care, the most significant consideration ought not be vulnerabilities in the software like data manipulation, privacy breaches, or potential for exploitation of biodata, but the environmental impact of AI. Health care emits a significant amount of carbon in many countries, but the environmental impact of health care has been underconsidered, in part, because of the assumption that all available health care technologies are medically necessary and therefore carbon emissions are morally irrelevant. As such, when the carbon impact of health care is evaluated, it is primarily at the institutional level-that is, the carbon of hospital buildings. 7 This paradigm circumvents accountability for the environmental impact of health care delivery, even though hospital care and physician and clinical services are the two largest carbon contributors to health care-exceeding health care buildings. 8 As AI further becomes an integrated and essential part of health care delivery, ethical reflection must include the potential to negatively impact the environment. Carbon emissions from AI use appear throughout the lifecycle of programming, development, and use due to the high energy and resource demands of AI. For instance, 40 days of training Google's AlphaGo Zero game was comparable to the carbon impact of 1,000 hr of air travel. 9 All AI systems must go through programming, running, and training. Moreover, the extraction of minerals, metals, and plastics necessary for AI capable hardware has tremendous environmental implications. 10 By way of illustration, Kate Crawford and Vladan Joler traced the environmental impacts of an Amazon Echo, illuminating the vast web of material resources used for extraction and production. 11 More robust research on the environmental impacts of AI were published by Crawford in 2021. 12 The carbon impact of health care and biotechnologies is a rapidly developing bank of knowledge in biomedicine, but it is insufficient to address comprehensive sustainability in health care, which requires an ethical framing that goes beyond carbon emissions and is sensible within technological ethics, as well as biomedical and environmental ethics.
Thus, this article will first overview the current output of carbon emissions in health care. It will, second, offer five reasons why carbon calculations are insufficient to address sustainability in health care.
Third, the article will derive normative concepts from the goals of medicine, the principles of biomedical ethics, and green bioethicsthe very locus in which AI in health care sits-to propose health, justice, and resource conservation as criteria for sustainable AI in health care. In the fourth and final part of the article, examples of sustainable and unsustainable development and use of AI in health care will be evaluated through the three-fold lens of health, justice, and resource conservation.
There is international agreement that AI should be used ethically, despite competing criteria. 13 30 And, while carbon emissions are one of the most widely accepted, and in some aspects, simplest metric to assess the sustainability of a particular good or service, using CO 2 as a criterion for health care sustainability is impractical-and ethically insufficient-for several reasons, despite it remaining the standard measurement in academic discourse.
First, although the carbon emissions of some health care developments, techniques, and procedures have been calculated, the vast majority have not. Calculating carbon numbers on all aspects of health care, in every country, and each branch of medicine will take an enormous amount of time and human resources. Like all carbon calculations, numbers are highly variable based on location (i.e., carbon of the country, national environmental standards) and available data. Moreover, carbon calculations can be elusive: just as hard data appears the inputs change. For instance, a carbon calculation of a cataract surgery will vary based on available resources, human efficiency, patient medical condition, and sourcing of energy. Despite this gap in knowledge, humankind must not wait for empirical data on carbon before making changes in consumption habits. As a society we know we must reduce our carbon. Climate change is too urgent a matter to wait for carbon calculations to justify sustainable health care.
Second, the motivation for carbon calculations is to reduce carbon either through carbon capping or carbon allocation. However, this assumes that there is a sustainable amount of carbon that can be emitted on a yearly basis. This is untrue. The amount of "safe" carbon in the atmosphere-calculated to be 350 parts per million-has 17 Zimring, C. (2017). Clean and white: A history of environmental racism in the United States. Third, and related to the second point, there is a concern that carbon calculations will lead to unfair limitation in health care. 32 Indeed, bias and discrimination that would lead policymakers to deprioritize health care needs of the medically underserved, including women, 33 LGBTQ+, 34 the disabled, 35  Fifth, while carbon as a metric does provide information on the environmental impact, simply identifying a carbon number and then declaring an item "sustainable" or "not" is meaningless since we are beyond the point of finding a carbon equilibrium and must live in a carbon recession. A carbon number, much like caloric information on food packaging, is merely descriptive unless it is set within a normative context. Unlike calories, however, there is no recommended daily carbon emissions that can be sustainably produced.
Since carbon numbers are a limited, at best, way to determine sustainability, another approach must be taken. The fourth principle for green bioethics-ethical economics-argues that humanism should drive health care developments before profitability. Financial gain often determines which medical developments, techniques, and procedures proliferate and which remain dormant. As a result, elective procedures that do not cure, treat, or prevent disease are readily available for those who can pay while the increasing cost of life-saving medicine prevents the poor from receiving health care.
Ethical economics is not opposed to generating revenue, but health care must not lose its primary mission of health and healing.
If applied correctly, resource conservation is the outcome of the four principles of green bioethics and the measure of their efficacy.
Thus, it is the most appropriate representative principle.

| SUSTAINABLE AND UNSUSTAINABLE DEVELOPMENT AND USE OF AI IN HEALTH CARE
The lexicon of the goals of medicine, biomedical ethics, and green bioethics capitalizes on foundational efforts to codify normative commitments. Health, justice, and resource conservation have been highlighted as thematic guides in this article, which map on to each of the three ethical systems named above. Health as a criterion for the sustainable development and use of AI does not necessarily place sickness as an antonym, although it may be the case that AI causes medical error. Rather, development or use beyond the purposes of health-whether enhancement, pleasure, or luxury-ought to be regarded as the unsustainable twin of health. Justice is a necessary component of non-individualistic sustainable artificial intelligence in health care, since it recognizes the claims of others. Biomedical ethics cannot pursue that which is unjust, or contributes to injustices. Thus, AI that widens health care gaps may be unjust and unsustainable. Resource conservation might occur in various steps along the health care delivery chain. AI could make a technique or a medical process more sustainable by conserving raw materials or reducing the need for medical interventions.
There are a number of possible weaknesses with this ethical framework at the conceptual level. The criterion of health is by no means a broadly agreed upon term, despite the contextualization given above. 56 Moreover, health encompasses many physical, emotional, spiritual, and psychological aspects. 57 The ordering of these aspects in proximity to the goals of medicine is debatable. Thus, even if different aspects of health were accounted for in AI, the ranking that an AI program might give to a particular aspect of health may be incongruent with patient preferences. 58 AI ordering, even if it accounted for a "standard patient" (which is a very nebulous concept, indeed) may be irrelevant if the health care facility does not have the means to support that particular aspect of health, or if the health care providers do not have the competencies in that particular area. 59 For example, an AI algorithm that has determined a tracheostomy will eventually restore health may be against a patient's do not intubate order, or unavailable in that facility, or not part of the health care provider's skill set.
These conceptual problems of health are applicable to the terms justice and sustainability as well. Justice is particularly elusive as a concept 60 and sustainability a less recognized ethical value in health care. 61 Although there are numerous reasons to include sustainability as an essential criterion for AI ethics, enumerated above, many health care systems are reluctant to include it when evaluating the ethics of medical decisions. One notable exception is the United Kingdom's National Health Service (NHS), which adheres to legally binding carbon reduction measures, 62 indicating that sustainable health care is a cornerstone of modern medicine.
For the purposes of this application, "health care" will include: (a) health care systems, such as hospitals and clinics; (b) medical procedures used within health care systems, such as robotic surgery and implantation of monitoring devices; and (c) health care insurance organizations, such as Blue Cross Blue Shield and the NHS, that collect and store health data.
To be sure, this excludes a number of places where health care is delivered in less formal settings, for instance home care and holistic doctors' offices. 63 This also excludes health care delivery outside of a medical setting, such as emergency medical services, which may nonetheless have access to AI. 64 Furthermore, many forms of technology collect and store health data-from Apple watches to Google searches for medical questions. 65 However, systems, procedures, and insurance define the parameters of health care in many countries where AI is being developed and used, thus providing a first place for inquiry, based on the potential for mass utilization of AI biotechnologies and environmental impact.
"Sustainable development" refers to theoretical AI. These are forms of AI that are being considered for development or expansion.
"Sustainable use" refers to applied AI. These are the ways that AIonce ready for deployment-is actually used in health care. The line here is discreet. The time between development and use is not fixed, as there is a period of trial and error in any new technology. 66 56 Huber, M., Knottnerus, J. A., Green, L., van der Horst, H., Jadad, A. R., Kromhout, D., Moreover, differing rates of development across interdependent sectors in health care biotech, for example bioengineering and computational mathematics, place development and use of AI as an ever moving field of ethical inquiry. 67 There are also differences of opinion among stakeholders about which aspects of AI should be developed, which will set the tone for health care innovation. 68 These complications are compounded given the various objectives that global partnerships and country-specific research teams must meet in AI development. 69 AI technologies are rapidly progressing. As they advance, a measure of agility in ethical application is required.

| Sustainable development of AI in health care
Sustainable development of AI in health care may include, for instance, triage algorithms in emergency rooms. 70 71 Errors might occur if an AI programmer fails to adjust for regional or state-wide differences in the type of emergency room cases that are frequent in a locale (e.g., frostbite or heatstroke). 72 The effects of bias, particularly unconscious bias-of sexism, racism, heterosexism, and so forth-may influence the programming people develop. 73 That is, a person may not be aware of their own biases and develop programs that replicate their own insensitivities. In cases of "deep learning" where AI gathers information from itself and adjusts accordingly, biased inputs at the beginning of the algorithm can contaminate the entire sequence, which can result in biased outputs. 74 Even in a perfectly designed AI algorithm, efficacy-in terms of health, justice, and sustainability-depends on human execution. This is both a benefit and a burden. On the one hand, health care practitioners retain moral responsibility for how they interpret and apply the results of AI algorithms. 75 On the other hand, inefficient or deviant implementation of AI algorithms may cause more harm, either through medical error, exacerbating inequality, or producing excess medical waste. 76 Thus, sustainable development and sustainable use of AI in health care often go hand-in-hand.

| Sustainable use of AI in health care
Sustainable use of AI in health care may include, for instance, analyzing rich text data to detect emerging outbreaks with novel symptom patterns 77 84 Elective, resource intensive health care may be postponed due to prioritization. 85 The tension between resource use of medical care and prolonged lifespans does not indicate that using AI for early detection is in-se unsustainable, but rather a critical reflection of global resource use must be undertaken. Humankind must choose where to use resources; medical care would be less of an environmental concern if other areas of life were sustainable.

| Unsustainable development of AI in health care
Contrarywise, unsustainable development of AI in health care may include gene-editing for aesthetic characteristics. 86 Superficial characteristics, by definition, fail to address health. This cosmetic option could quickly become a commercialized service that exacerbates inequality through inhibited access, thus ignoring justice. Geneediting for aesthetic characteristics has the potential to be used primarily in high carbon countries, hence resources will be exploited rather than conserved.
There are many debates about the value of utilizing "aesthetics" as part of health care. 87 These arguments are generally put forth in locales where basic health care is available and personal desires drive the pursuit of cosmetic "medicine." 88 Another serious concern about limiting unsustainable AI developments is that of moral luck and moral responsibility. 89 Moral luck is the theory that recognizes that people are born into better or worse circumstances through no merit of their own and moral responsibility therefore attempts to balance these inequities. While those in the developed world who are accustomed to capitalistic choice in a market economy may opine that it is unfair to restrict access to elective treatments on the basis of resource use, this objection is minimized with the recognition that climate change harms everyone. Therefore, so-called "entitlements" of individuals to particular elective procedures must be balanced with the environmental effects that they-and others-will experience, while also raising the medical standard in places that are underserved.
In many ways, ethical concerns about developing AI for aesthetic characteristics cut across the goals of medicine, biomedical ethics, and green bioethics. The discussion on therapy and enhancement, 90 access and distribution of medical resources, 91

| Unsustainable use of AI in health care
Unsustainable use of AI in health care might include Care Bots for children 93 and other patient populations. Care Bots, which are robots that provide assistance (such as moving an overweight person), 94 care (such as nursing assistance), or companionship (in the form of an animal or humanoid, for example), 95  Using AI in Care Bots may cause medical harm, but like other forms of high-tech health care, this is a risk that patients might be willing to undertake once properly consented. 96 The issue of moral luck might also be invoked in support of AI in Care Bots, that if available, they should be used, with the caveat that this form of AI 83 Tham, J. (2010). Challenges to human dignity in the ecology movement. The Linacre appears to be more "value added," rather than at the core of medicine, and therefore more subjected to ethical scrutiny.
There are many different types of Care Bots and it may not be the case that all Care Bots would be unsustainable. However, while some technologies are intentionally programmed to be obsolete, all will become obsolete eventually, 97 causing unnecessary resource use. Whereas human caretakers may need to learn new skills through continuing education, this is a low-impact activity; reprograming or updating Care Bots is a more resource intensive endeavor. 98 A similar but related objection to the unsustainability of Care Bots is the availability of satisfying alternatives. Even with health care provider shortages in some countries, 99 mobility and training can supply human health care workers to those in need, thus making this form of AI in some ways redundant. Unlike genetic editing, or rapid algorithms, which have less appealing alternatives due to cost, time, or intellectual investment, the functions of many Care Bots can be performed equivalently, if not better, by humans. 100 These examples of sustainable and unsustainable development and use of AI are entry points for ethical assessment, but far from comprehensive. In the future, the most innovative and ethically complex forms of AI will need to be evaluated based on the criteria of health, justice, and resource conservation.

| CONCLUSION
Consensus about "the" best, or most relevant, ethical system for health care is an ongoing conversation. As AI becomes indispensably ubiquitous, there is societal reticence to discontinue it unless there is a compelling reason. The relationship between technology and ecology will be a defining feature of biomedicine in the 21st century.
As such, ethicists are better situated to advocate for the sustainable use of established AI rather than persuade society to abandon new developments in AI altogether. By utilizing the criteria of health, justice, and resource conservation the goals of medicine are ethically supported while the possibility-indeed, the necessity-of integrating just sustainability into health care is actualized.

CONFLICT OF INTEREST
The author declares no conflict of interest.