Artificial intelligence (AI), as covered in other portions of this volume, refers to a promising set of tools that confer opportunity for advancing knowledge, access, and the patient experience across areas of dental medicine. As AI spreads, experts across health care have heralded the possibility for this technology to create broader spread of services, expand time for clinicians to spend with patients, improve pattern recognition across diseases, discover new and advanced treatments, and uncover deeper pattern recognition and analyses than can be reasonably expected with human cognition alone. Models that improve predictions of disease development and response may support improved individual and population health. Further, using these algorithms may create estimates in ways that are faster, less expensive, and more applicable to a broad population than what can currently be achieved when acting without these technological adjuncts. In doing so, AI may improve among others diversity of understanding, fairness in access to care, diagnosis of rare diseases and uncommon presentations, and comparison to and prediction of other conditions. These advances approach the European Union Artificial Intelligence Act goal of human-centeredness in AI, spreading the tools for practical and tangible gain.

Clinicians, regulators, and developers, however, must evaluate the promise of AI in the context of ethical principles that can guide appropriate use. This may include consideration of both accepted and new standards of moral development related to the use of these new technologies. Explicit consideration of specific ethical quandaries must be made, including where an AI system sources its training references, applicable legal regulation, and anticipated pitfalls in both technological capacity and availability. ^, Further, these considerations must be balanced against the impact on the patient and should combine practical frameworks with and supplement guidelines that already exist throughout health care. ^, Evaluating AI through this framework will allow the resultant guidelines to confer a balanced benefit, rather than merely respond to societal trends.

Much of health care, including dentistry, bases ethical guidelines in large part on existing and shared principles initially proposed by Beauchamp and Childress in their influential Principles of Biomedical Ethics . The framework set out in this volume, referred to as principlism , structures health care ethics around the principles of autonomy , beneficence , nonmaleficence , and justice . The American Dental Association adds to these a fifth principle of veracity in their official Code of Ethics. When applied to questions of AI in dentistry and other health care fields, these principles have been adopted to support appropriate use and have been expanded to account for novel considerations posed by this growing area. What follows begins with an evaluation of how the traditional ethical principles inform and structure appropriate use and development of AI in health care and dentistry.

Autonomy , or the right to self-governance and independent decision-making, is a principle both inherently based in humanity and with clear applicability to the world of AI. This principle applies clearly to the training and development of AI systems, as patients are generally considered to have a right to understand the use of their own information and data, including how and for what purpose it is being employed. Autonomy may be distilled as a form of privacy, given that it protects patient freedom to decide how personal data is used. This principle is particularly relevant when data is used outside of direct treatment, such as in the training of an AI system. AI systems require large amounts of high-quality data on which to be developed, much of which can be deidentified to large extents. However, the consistency with which deidentification occurs requires confirmation and careful monitoring. Additional consent terms may also be required, both given the use of data outside of direct treatment and the possible participation of third-party entities that run AI systems. What constitutes sufficient deidentification and anonymization also remains an important, open question for those both developing and employing AI systems. It remains imperative that these model systems be unable to trace data to an individual. These considerations, however, are most appropriately balanced, as always in the case of ethical principles, with benefits to the population supplied by the development of AI systems and the advancement of freedom conferred through recruitment of new experts outside of a treating institution. Balance of these tradeoffs in a fully realized AI may maximize good for both individuals and populations.

Beneficence , or the requirement to do good, is similarly employed in conversations around AI development both for individual good (related to the considerations around autonomy) and the good of the population. The good of the population may be considered to include both population-level measures of health and maintaining the relationships between society and the health care system through fostering trust, as in the question of privacy. Nonmaleficence , or the requirement to do no harm, is again employed in this setting to safeguard privacy and extends to considerations such as security of systems and sustainability. Justice , or fairness, is a broad consideration, including in the spread of technology to benefit all populations equally and the promotion of social justice through access and benefit without discrimination. Specifically, the development of and access to AI systems must balance burdens and benefits across populations. For example, to the extent possible, training data must accurately represent all populations and training results must be accessible and employable by representative portions of the population.

These classic principles permeate most conversations around health care but have limits in their application to new technology. Such broad concepts may not directly correspond to specific concerns in developing areas. This is particularly relevant to AI, where the overlap between technology, engineering, and medicine bring together expertise from different backgrounds and require explicit accountability and clarity in understanding. Therefore, general knowledge and relevance have to be balanced with considerations unique to these specific areas by independent governments or oversight bodies. While patients are considered more willing to share data and trust in systems developed under government oversight or involvement of other trusted parties, appropriate development and dissemination also has to include honest sharing of information around both strengths and weaknesses. Regulation must also keep pace with the development of new systems.

In addition to the traditional principles of biomedical ethics, a range of new, more specific principles has been proposed to frame the development and use of AI in health care. These bring together the field of engineering ethics with the previously mentioned standards from health care and expand the healthcare-specific considerations that apply to the growing world of healthcare AI. From engineering, AI ethics brings in the influence of technology on society and the environment. This lens examines technology directly to optimize development and regulation through understanding how its makeup informs ethical development and how potential pitfalls inform ethical use.

The health care perspective expands these considerations to include limits on: what is entrusted to algorithms, the impact on humans, and transparency of the involvement of AI in individual decisions. This has been summarized in various forms, which may include specific principles of transparency , explainability , accountability , and fairness . ^, Explainability has been a particular focus, given the nature of AI systems and the minimal background many in healthcare, not to mention patients, have to understand the development and detailed workings of these systems. Understanding, however, must be in balance with safety and efficacy, as understanding alone is insufficient as an ethical safeguard. This requirement connects to the principles proposed by the World Health Organization, which balance autonomy , public interest , explainability , accountability , inclusiveness , and sustainable development . Other guidelines hold these same principles and add equity , trust , and solidarity as concepts essential to building and maintaining appropriate uses of AI. Across guidelines, each of these principles is proposed to guide knowledgeable entities capable of understanding and interpreting outcomes, often with a specific requirement for human control given the centrality of safety , privacy , fairness across populations, and accountability . ^,

Privacy is again raised in these conversations as an explicit principle central to the development and use of AI in health care settings. Responsible use of large amounts of high quality of data is essential for the training of new AI systems and must be balanced against protection of patient information and identity. In this way, individual concerns such as use of patient data are balanced against societal good of AI capabilities, and all systems must be designed such that identity is protected while maximizing the capabilities of the system. This may be achieved through intentional design of systems in a way that employs interdisciplinary best practices.

Taken together, these principles allow an end user to determine the source of information and communication from AI, the balance between human and AI decision making, how AI reached its conclusions and how accurate these are, how representative the data used in the algorithm was (and therefore the applicability of findings to a given case), and how these and other decisions are controlled by human oversight. Such considerations allow not only for trust in the accuracy of AI output but also in the privacy of the data being used, the reliability of results, and the application of any AI recommendation across populations. The intersection between these new proposed principles and their relationship to traditional principlism is summarized in Table 1 .

Concerns related to the development of ethical systems of AI in health care and dentistry are mirrored in considerations of best practice. General concerns include how principles may practically control new technology. Specific concerns around lack of ethical oversight include privacy breaches from insufficient deidentification or theft by third parties and systems that produce biased, inaccurate, or insufficient results. ^, The long-term output of AI systems is of particular concern given the autonomous nature of these systems. Their unsupervised evolution might allow for iterative conclusions that are made without clear moral responsibilities, in contrast to human decision-making. Unlike traditional ethics, then, which relies on individuals as autonomous moral agents, AI systems will require human oversight for fair use. Issues of agency in AI are particularly salient and encompass not only oversight but also the development and design of systems. AI systems must be created in ways that centralize the human end user and respect their autonomy. However, it remains unclear how this principle would be designed or enforced and how both individual patient and societal good would be protected through oversight from any evaluating boards or legal regulatory frameworks. ^,

A further salient challenge in the development of medical and dental AI relates to the propagation of design flaws or sub-par training data. This concern highlights the need for diverse high-quality training data that is representative of entire populations. Indeed, systems trained on insufficient or nonrepresentative data may not be generalizable in novel contexts, which may confer safety risks around the use of their conclusions. This concern is particularly salient in dentistry, given the rarity of many conditions treated within the field and the relative paucity of population-level data. The decentralized nature of dentistry is also likely to only compound questions around data-sharing, storage costs, mining, and legal protections. For this and other reasons, the concept of algorithmovigilance , or careful monitoring of development of datasets, holds central importance. This includes not only how the data is used but also the source, diversity, and funding behind each input.

Rapid development of AI systems across health care has also meant that literature is currently insufficient to fully assess or address these concerns. While some medical reviews of AI either include or focus on the ethics of AI use, mention of AI ethics in dentistry to date only exists as a component of other topics. One scoping review in medicine did identify issues discussed in the medical literature, including prudence, equity, privacy, responsibility, democratic participation, and solidarity, as most frequently raised. These were mentioned across both computer science and medical perspectives, with no increasing focus found on ethics in AI as time progressed. Of the data included in this review, though, only 7.9% of identified articles mentioned the importance of ethical guidelines and 12.4% presented individual issues that may arise. Dental literature lags even further behind, highlighting a general need for more detailed consideration of ethical principles.