The AIM-AHEAD EEWG was created in 2021 to guide the ethical and equitable development and implementation of AI and ML tools and processes broadly within the AIM-AHEAD. Simultaneously, an Equitable Policy Development Workgroup was developed within the AIM-AHEAD Infrastructure Core. To ensure rapid and coordinated progress with respect to embedding ethics and equity into AIM-AHEAD activities, both within and outside of the Infrastructure Core, the EEWG’s efforts were harmonized and merged with the Infrastructure Core’s Equitable Policy Development Workgroup upon recommendation by the EEWG cochair and multiple principal investigators for the AIM-AHEAD Infrastructure Core. The newly reconfigured EEWG began by defining its scope of activities (Figure 1).

At the start of the program in year 1, the EEWG was comprised of 51 members (AIM-AHEAD principal investigators and coinvestigators) and 3 cochairs. AIM-AHEAD participants either requested to join or were selected to join by their project leaders within the program. During year 2, the EEWG’s membership was consolidated into 2 cochairs and approximately 40 AIM-AHEAD principal investigators, coinvestigators, leadership fellows, and research fellows. This reduction in EEWG cochairs and members occurred for two main reasons: (1) time and effort among members were reallocated to other activities within the AIM-AHEAD (administrative planning for regional hubs, research, etc), and (2) given the evolution of the program over time, the year 1 members were provided an opportunity to recommit to the EEWG for year 2. In both years, EEWG cochairs and members represented a variety of academic disciplines and focus areas, including but not limited to medicine, computational science, population health, health science, data science, bioethics, law, community engagement, human-centered design, health disparities research, biological science, social science, and engineering.

The initial effort of the EEWG during year 1 was to produce a set of principles and a glossary to inform the practice of ethics and equity in AI and ML development and implementation in health research. During year 1, members convened in weekly meetings that led to consensus on the development of specific workgroup deliverables. EEWG members reviewed the literature to identify relevant sources with perspectives on ethics, equity, and social determinants of health, especially those that were community driven, and lessons that could inform the development and use of AI and ML in health disparity and disease prevention research [12-27].

To develop the principles, the EEWG used a modified Delphi approach to facilitate discussions around tangible steps that the Consortium should take to ensure that ethics and fairness are at the forefront of AI and ML applications to build equity in biomedical research, education, and health care [28]. Specifically, the EEWG engaged in weekly (year 1) and biweekly (year 2) meetings to suggest, review, and deliberate a corpus of published content and literature considered useful toward integrating ethics and equity into AI and ML development and contributed original thought leadership and content in reaction to the content and literature reviewed to devise actionable principles. The EEWG approached the development of the principles with optimism about the potential of AI and ML to address health disparities by empowering communities, yet with recognition of complex societal challenges: inadequate or misrepresentation in data sets, algorithmic bias, imbalances in communities’ access to data and information about themselves, misuses of AI and ML tools, and threats to the civil and human rights of individuals and communities who are or may be subject to illegal or pervasive AI and ML surveillance, to name just a few.

To develop the glossary, during year 1, the EEWG began by defining ways in which outputs of AI and ML can (1) fail to be informative or useful for individuals and groups; (2) distinguish among individuals in inappropriate ways as a result of bias, failure of inclusion, or misuse; or (3) be poorly vetted by individuals and groups who are or may be subject to potentially harmful actions and decisions made by key or authoritative stakeholders that rely on AI and ML for decision support as a result of insufficient engagement with key stakeholders, including data participants.

Using a modified Delphi approach that likewise involved polling, ranking, and other exercises, consensus was reached on terms to define [29]. During its meetings, the EEWG discussed all possible terms that would be key to define to inform the ethical and equitable development and application of AI and ML, followed by 2 rounds of ranking and polling exercises to narrow their suggestions to 12 sentinel terms. Sentinel terms discussed during meetings, for example, included demographical terms such as self-defined or assigned race, ethnicity, sex, ability, and gender that can lead to errors in the development of AI and ML, which can in turn lead to potentially irreversible, intergenerational, and multigenerational harm to individuals and groups subjected to decisions informed by or based on AI and ML outputs. During year 2, remote meetings were held on a biweekly basis to further deliberate and refine the principles and glossary. Refinements were based on expert stakeholder feedback gathered through a survey among participants in the AIM-AHEAD pilot project and during remote convenings.

The EEWG initially sought to conduct a quantitative survey to assess how AIM-AHEAD researchers would implement the principles in practice. A draft survey was developed by 2 volunteers within the workgroup, who later shared the draft survey with the broader workgroup for iterative feedback and edits during weekly (year 1) and biweekly (year 2) meetings. The draft survey was also shared with awardees of AIM-AHEAD pilot projects for feedback. As the EEWG deliberated on the feedback, it ultimately determined that a qualitative interview (vs a quantitative survey) would be a more useful approach to garnering AIM-AHEAD researchers’ perspectives on implementing the principles in practice. Thereafter, the EEWG met regularly to convert the quantitative survey into an interview guide with the intent of learning the interviewees perspectives and natural reactions to the AIM-AHEAD ethics and equity principles and glossary.

The EEWG’s efforts in developing the interview guide and conducting the interviews were focused exclusively on program-specific planning for the AIM-AHEAD and were not intended as human subjects research. AIM-AHEAD investigators’ responses to the interviews were wholly voluntary, and their comments were used exclusively to develop the program’s principles and were subject to further assessment for generalizable knowledge.

Developers of AI and ML platforms and tools must contemplate, anticipate, mitigate, and address potential issues with downstream data aggregation, interpretation, and use. Meeting these goals requires a shared understanding of the terms used in these policies and processes. The EEWG determined that, in many cases, sensitive demographic characteristics (eg, race, ethnicity, sex, ability, and gender) are particularly problematic as variables used in AI and ML because they are often inappropriately understood as being rooted solely or primarily in genetic or phenotypic differences rather than strongly influenced by discriminatory sociohistorical and sociocultural practices.

To capture and promote a shared understanding of key terms, the EEWG developed a glossary of 12 words (Table 1) out of 28 considered that follow or build upon existing understandings of these concepts, highlighting their particular importance for the optimal development, refinement, and implementation of AI and ML.

In addition, the EEWG developed a concept relationship diagram that describes the logical flow of and relationship between the definitional concepts described in Table 1 and Figure 2. The center of this diagram is equity, which requires AI developers and implementers to enforce fairness and avoid bias in a population with sufficient diversity by being inclusive. To implement diversity, representatives that are characterized by a minimal set of aspects—ethnicity, race, gender, and sexual orientation—need to be collected. They will form a representative sample if they can reflect the characteristics of a population. A representative sample can mitigate algorithmic bias, which is one specific type of bias.

As mentioned, extensive and iterative feedback received during the development of the quantitative survey led the EEWG cochairs and members to determine that a qualitative engagement approach is warranted to facilitate meaningful and diverse stakeholder engagement to disseminate and facilitate implementation of the principles and glossary. Therefore, the EEWG developed an interview guide that can be used or adapted to garner and understand AIM-AHEAD members’ and other community perspectives on the principles and glossary. The interview guide is provided in Multimedia Appendix 1.

The role of those who will be affected by the findings of the research enterprise has evolved from their initial role as objects, as illustrated in the iconic painting of Edward Jenner administering the life-saving inoculation of the English boy with cowpox in 1796, the multiepisode television documentary “Microbes and Men,” and the abuses of Black men in the US Public Health Service Study of the natural history of untreated syphilis at Tuskegee [30-33]. Over time, more attention has been devoted to assessing the potential harms and benefits of research to the people who are studied, albeit primarily as viewed by investigators, typically White men, and institutional review boards, typically comprised of researchers with minimal or latent community involvement. Incentivizing representation of nonscientific, nonaffiliate community members on institutional review boards, engaging members of historically underrepresented groups in more visible roles as investigators, and engaging minority-serving institutions as partners in AI and ML research is necessary to promote equitable access to opportunities and careers in AI and ML. Such an intentional approach also, importantly, demonstrates an appreciation for local knowledge and facilitates the design of more culturally informed interventions that consider how research will affect heterogeneous populations being studied in AI and ML research. This form of appreciation is necessary for tailoring engagement to the needs of diverse groups and understanding how to overcome barriers to AI and ML research and use [34].

Beyond promoting diverse and equitable opportunities for participation in AI and ML research, it is necessary to recognize the need to translate that work into actual practice, which historically has also been a barrier to health equity. For example, the association of the lower-quality data measured by pulse oximetry with dark skin tones has long been known, and there have been versions of the technology designed to account for this discrepancy, but versions of pulse oximeters with biased tendencies remain in wide use [35]. There is a real risk that AI and ML technology will follow a similar pathway if there is not sufficient action to build ethics and equity into the research.

Overall, our effort reported here achieves 2 goals. The first is to describe what is needed procedurally and substantively to achieve equity. This is a complex process that must take place and evolve over time. It cannot be addressed as a 1-time event or by filling out a checklist. Achieving equity requires rebalancing the interests at stake in research, which, at a minimum, means truly considering and addressing the interests of the people who will be affected by the results. Ideally, research participants can become cocreators as ethics in AI and ML and related ethical principles evolve into more commonly accepted policies and practices. The second goal of this reported effort is to emphasize that addressing equity requires an inclusive, ongoing process with a shared understanding of salient terms that will evolve over time. Recent engagements within the AIM-AHEAD program have noted this to be true even for terms like AI and ML, as today very few stakeholders have been able to clearly articulate how AI and ML can be or is used in the real world [34]. New and ongoing national initiatives, such as the National Academy of Medicine’s AI Code of Conduct project, which intends to develop a “code of conduct for the development and use of AI in health, medical care, and health research,” are encouraged to learn from the EEWG’s efforts [36].

Our work builds on and can be incorporated into current AI and ML ethics and equity frameworks and policies within and outside of the United States, focused on improving population health through broad community involvement in AI and ML application development [17,36-38]. This includes, but is not limited to, the National Institutes of Health’s policies and programs on AI and ML application development in health research; policy developments undertaken by the US Senate Health, Education, Labor, and Pensions Committee; the National Academy of Medicine’s Artificial Intelligence Code of Conduct project; the European Commission’s Guidelines for Trustworthy AI; Asilomar AI Principles; and lastly and importantly, a groundbreaking and recent US White House Executive Order explicitly supporting the mission of the AIM-AHEAD [36,39-42].

Importantly, our work provides a complementary, fundamental, and basic blueprint or process, along with operational tools and building blocks, to educate stakeholders on this practice of creating safe spaces and setting culture tones for diverse stakeholder engagement and consensus around best practices and shared terminology. Also importantly, our tools enable the collection of ongoing and iterative feedback concerning the local implementation of our principles and glossary. Iterations may be further disseminated, along with public-facing endorsements of the principles and glossary in their current form, by like-minded stakeholders seeking to ensure that researcher diversity, community, and social justice concerns influence AI and ML application development processes in health research and, broadly, science and technology.

Inclusive and ongoing processes to develop a shared understanding of salient terms like AI and ML and those described in our glossary require more time, greater inclusion, and deeper incorporation of diverse community perspectives. This approach differs drastically from the typical project life cycles afforded by the gold rush mentality that has emerged with AI and ML today. Therefore, one key step, moving forward, would be to persuade leaders in the AI and ML research enterprise to broadly disseminate the lessons that may be learned in operationalizing our EEWG principles and glossary. Programs such as the AIM-AHEAD need to objectively assess their administrative processes and evaluation criteria for what constitutes ethical and equitable opportunities for an AI and ML investigation, including investigator inclusion, data governance, data sources, and data infrastructure.

There are limitations to consider in our process and recommendations. First, the EEWG has continuously revisited the principles and glossary for potential editing based on the members’ evolving experience and expert opinions, even though making these deliverables “living documents” complicates the process of achieving sustainable consensus. Nonetheless, the principles and glossary will require reflection, appreciation, and adjustments over time to account for the effects of real-world events, human choices, or interpersonal phenomena from relevant perspectives. Also, some of our proposed glossary terms may already be limited in scope with respect to real-world events and phenomena. For instance, although our definition of “representative” concerns “an individual or body chosen or appointed to act or speak for an individual, population, or subpopulation,” there are certain matters in which a representative may be self-appointed without specific authorization from those they wish to represent.

Therefore, ongoing engagement around the use of our principles and glossary in AI and ML research settings is encouraged to maximize their potential benefits and minimize any potential harm. However, ongoing engagement with institutions that have limited resources to support their full participation requires careful consideration and discussion of how to incentivize, support, and sustain meaningful engagement beyond mere compensation. One way to accomplish this is to seek institutional input through authentic connections to determine what they consider a valuable investment for their time, instead of deciding for them. For example, such connections can be made both within and outside of conferences, convenings, and events hosted by minority-serving institutions nationwide (eg, the Annual Biomedical Research Conference for Minoritized Scientists or the National Society of Black Engineers’ Annual Convention).

An overemphasis on speed or velocity works against taking the time needed to foster the inclusion of historically and presently underrepresented communities in the development of AI and ML, ultimately rewarding AI and ML “haves” over “have-nots.” In the private sector (eg, big technology companies and startups), the pace of AI and ML development is extremely rapid and difficult to manage. Inequitable divisions in access to resources like computers, smartphones, and the internet have vastly decreased over the past decade. Yet, AI and ML technology that is used with adequate operational know-how and e-literacy, cost of use, human resources and staffing needs to maintain cyberinfrastructure, and many other technical and nontechnical resources, is where these inequitable divisions can be addressed.

An equity-oriented public sector intervention, such as the AIM-AHEAD, can be more effective in achieving diversity and inclusion goals by emphasizing actions that do not sacrifice trust-building for the sake of rapid development of technology, especially in the initial stages. By slowing down to meet historically and presently underresourced institutions and communities where they are and where they are capable of engaging and competing, we can more effectively evaluate AI and ML implementation and results for bias over time and expand the potential to achieve the aims of ethics and equity. We envision a virtuous cycle of shared learning, building on our EEWG deliverables, that may bridge researchers and impacted communities into a new intersection of computational sciences, ethics, and health equity.