3
AI CODE OF CONDUCT FRAMEWORK

SUMMARY OF PROCESS FOR DRAFTING PRINCIPLES AND COMMITMENTS

The National Academy of Medicine (NAM) AI Code of Conduct (AICC) framework described here was developed to align the field and catalyze action to ensure that the potential of artificial intelligence (AI) in health, health care, and biomedical science is realized. The expectation is not of wholesale adoption of the AICC framework across the industry. Rather, the AICC framework is intended to serve as a touchstone for organizations and groups developing their own considerations and approaches for inclusion and alignment when assessing internal guidance for completeness in their specific context, thereby advancing trust and minimizing the likelihood of actors across the field working at cross-purposes. The drafting of the AICC framework, consisting of the Code Principles and Code Commitments, was intentionally aligned with prior efforts of academics, industry, and governmental agencies to elucidate guiding principles for responsible AI in the health care and biomedical science domains. As detailed in the draft AICC framework (Adams et al., 2024), a landscape review was conducted to identify commonalities published to solicit public and industry comment in existing frameworks as well as critical gaps. Peer-reviewed literature highlighting considerations for responsible AI published between 2018 and 2023 was examined as was guidance on responsible AI produced by medical specialty societies and federal and international policy makers. Additionally, key informant interviews were held with industry experts to solicit input on essential components for the AICC framework.

NAM Learning Health System (LHS) Shared Commitments provided the foundation for the Code Principles, seamlessly integrating the constructs. Just

as the LHS Shared Commitments establish expectations for all participants in the health system, the Code Principles provide analogous guidance specifically tailored to the development and implementation of health AI. They represent a natural evolution of the LHS framework to address emerging technologies. The Code Principles reflect the values and norms to be applied in the context of health AI governance to promote trust while ensuring the benefits and mitigating the risks associated with Al in health, health care, and biomedical science. Consistent with the complex adaptive systems theory, which posits that a small set of simple rules can result in system-level change in complex systems (IOM, 2001), the Code Commitments are a set of decision-oriented rubrics intended to support the application of the Code Principles in practice and to guide the behavior of individuals, organizations, and communities, as well as local, national, and transnational agencies operating in complex systems. Table 3-1 presents the relationship between the Code Principles and Code Commitments, reflecting the distillation of the Code Principles into broadly applicable guidance for decision making across impacted parties and across the AI lifecycle. Cells shaded in green reflect an alignment between the Code Principles and the Code Commitments. For example, Commitment 1, Protect and advance human health and human connection as the primary aims, addresses the Code Principle “Secure” in that ensuring privacy and security of health data is integral to protecting human health and human connection, as a breach of data security could result in both breakdowns in mental well-being (health) and in trust (human connection).

SUMMARY OF PUBLIC COMMENT

The AI Code of Conduct draft for public comment, including the Code Principles and Code Commitments, was published by NAM in April 2024 (Adams et al., 2024). Public comment was solicited on release, and through an online survey; additionally, presentations on the AICC framework were requested by federal agencies, industry coalitions, collaboratives, and professional association meetings, where verbal feedback was provided to the NAM. Reactions to the AICC framework were provided by a diverse group including federal agencies, researchers, clinicians, AI developers, patients and family representatives, health product manufacturers, ethics and equity experts, standards-setting bodies, and care delivery systems.

Feedback was provided to the NAM staff verbally and in writing through industry events, and, for individuals, an online survey instrument. The survey section provided a 3-point scale (Yes, Somewhat, No) response options to reflect agreement that the Code Principles and Code Commitments engendered various

TABLE 3-1 | Crosswalk of Draft AICC Code Principles and Commitments

NOTE: Cells shaded in green reflect an alignment between the Code Principles and the Code Commitments.

qualities, including clarity, relevance, completeness, conciseness, robustness, and adaptability. Elaboration was also encouraged.

The vast majority of respondents found both the Code Principles and Code Commitments to be relevant, clear, and concise. Likewise, a smaller group, though still a majority, found the Code Principles and Code Commitments to be complete, robust, and adaptable. Despite high levels of positivity about the conceptual framework, only about half of respondents felt that the set of Code Principles and Code Commitments was sufficient to focus attention and action on issues anticipated to ensure that use of AI in health and health care optimally advances the human condition. This result was expected as outlined in the publication of the draft framework, which was proposed as a “starting point for real-time decision making and detailed implementation plans to promote the responsible use of AI” (Adams et al., 2024, p. 6). Also outlined in that publication, the plan for the second phase of the project, was to take the Code Commitments to the next level of granularity with working groups describing their accountabilities and essential collaborative actions and activities mapped throughout the AI lifecycle and aligned to stakeholder perspectives and responsibilities. The output from these working groups, an online survey, post-presentation feedback, and ongoing expert recommendations yielded additional edits for clarity and emphasis of the Code Principles and Code Commitments (see Tables 3-2 and 3-3). From the project outstet, the authors underscored that the AICC framework will require ongoing review and, as necessary, revision based on experience with application of the framework as well as changes mandated by advancing AI technologies and the evolution of governance capabilities.

TABLE 3-2 | Updated AICC Code Principles

TABLE 3-3 | Updated AICC Code Commitments

This page intentionally left blank.