Public health emergencies, such as the COVID-19 pandemic, the 2014‐2016 West Africa Ebola epidemic, and natural disasters such as floods or hurricanes, pose serious and pressing individual and population health threats. Depending on the size and scope of these emergencies, they can require urgent and harmonized responses from governmental decision makers and public health systems at the local/tribal, state/territorial, and even national levels [1,2]. Sometimes, when an emergency is of significant size and scope, or is highly visible or consequential, decisions that are typically made by public health or emergency management experts shift to, or at least involve, elected officials, such as state governors [3]. This shift might be made because elected officials are politically responsible for the outcome of a response, or the shift might happen because only elected officials can exercise the legal powers necessary to respond. For example, in all states and territories, governors can issue emergency declarations to enact measures such as accessing and reallocating funds or suspending laws and regulations that may impede response efforts [4-6].

Regardless of the inciting event, when elected officials or other leaders are thrust into a public health decision-making role, often without formal public health or crisis response training, they may need additional support to ensure that decisions are well-considered. While public health and emergency management experts can, and should, play a significant supporting role in decision-making, leaders without good decision-making processes in place can easily fall victim to so-called decision-derailers–common biases, traps, and heuristics that can make decisions less-than sound [7].

A hallmark of public health emergencies is an initial severe information deficit, while many competing voices clamor for urgent action. There is also inevitably a tug-of-war between health and safety priorities and other societal imperatives such as economic stability. Further, the broad categories of “health,” “economic,” and “societal” impacts must themselves be unpacked; choices about how to respond, what measures and recommendations should be put in place, must consider and involve diverse populations who are affected in very different ways and to different degrees.

So, what can leaders do to ensure that their decision-making process has systematically identified and considered a full range of actions that could effectively address fundamental goals; carefully considered and balanced all the values at stake; incorporated the information available while hedging against the remaining uncertainties; and has built public support through its transparency and thoroughness?

First, policymakers should clearly define, justify, and articulate the strategic goals that they aim to achieve and then share this information with the public. Next, it is critical to identify what principles leaders want to uphold in an emergency response. Protecting the health of the public is a general principle that most leaders identify with. Are there other important principles that should not be forgotten, like ensuring the consequences of interventions are fair across populations? As decisions are not made in a vacuum, the voices of diverse experts help to consider a variety of angles to a problem.

Once the goals, principles, and structure for decision-making are in place, an assessment process for identifying and evaluating options for action is needed. Evaluating the possible downstream impacts of options (both positive and negative), comparing options against one another, and assessing the logistical feasibility of the range of options. Finally, in addition to a continual updating as new evidence and information emerges, well-documented communication practices highlight the importance of telling the public that information may change and foreshadowing how new information can change plans.

In a crisis, these structures and processes are very difficult to implement, especially under time pressure and without prior planning. Developing plans ahead of a crisis is optimal, but limited time and resources make this rare, and leaders will still face an information deficit difficult to overcome with legacy approaches to data gathering.

Artificial intelligence holds considerable promise to help decision-makers gather data to inform decisions and implement decision-making processes to systematically consider options and undertake tradeoff analyses. Ultimately, AI has the potential to greatly enhance the quality of public health decision-making in these emergencies. However, while the potential for decision support is significant, numerous challenges persist in developing and operationalizing AI for public health crisis decision making. While digital health technologies and information and communication technology (ICT) infrastructure have expanded rapidly, empirical evidence regarding their impact on population health outcomes remains mixed [8-12]. Studies show that ICT adoption does not automatically translate into measurable health improvements, particularly when institutional capacity, governance quality, workforce skills, and implementation fidelity are insufficient. Previous digital health initiatives often showed weak or context-dependent effects due to inadequate organizational readiness, data quality issues, and a lack of complementary investments in regulatory frameworks and human capital.

The following viewpoint considers technical, logistical, and ethical challenges in developing AI tools to support elected officials in public health decision making and describes important considerations when addressing AI applications to optimize system use in the public health domain.

Machine learning (ML) models that support public health decision-making are subject to technical challenges in their development pipeline prior to being scaled for public health response efforts. Underfitting and overfitting are two of the most common challenges in the model development process, which limit a model’s utility. Models are sensitive to quality and representativeness in the data used to train them.

Overfitting occurs when a model becomes too specialized in its training data and fails to perform well on new, unseen data. Overfitting limits a model’s generalizability on test or real-world data and leads to high variance in model performance based on changes to input data. For example, a forecasting model trained on historical influenza data that captures every nuance of past flu seasons may perform poorly during a novel pandemic like COVID-19, where transmission dynamics, public behavior, and health care utilization patterns differ substantially from historical precedent. Overfitting may be addressed by collecting more data to create a more robust training set, reducing the model parameters or undertaking feature selection, and using cross-validation. Underfitting occurs when a model is too simplistic and fails to capture underlying patterns in data. This leads to missed patterns and inaccurate predictions, even for training data. In emergency contexts, this might manifest as a basic model that predicts hospital admissions using only a single variable (eg, previous d’s case count) while ignoring critical factors like vaccination rates, demographic composition, or seasonal patterns. Such models provide unreliable guidance and may lead decision-makers to underestimate resource needs. Increasing the number of model parameters or features and reducing regularization may help to reduce underfitting.

Systematic evaluation is essential for responsible AI deployment in public health emergencies. Agencies should conduct pre-deployment testing through prospective validation using historical emergency data, simulation exercises, and tabletop drills to assess accuracy, identify failure modes, and evaluate integration with decision workflows before operational use. Performance benchmarking should establish baseline accuracy metrics, track performance over time, and, where possible compare AI-supported versus non-AI-supported decision outcomes to assess actual impact on decision quality. Real-time monitoring requires dashboards that track model performance during deployment with automated alerts for degradation, supplemented by regular expert review to identify issues not captured by automated metrics. Agencies must establish adverse event reporting mechanisms similar to health care systems, with clear protocols for investigating suspected AI-related errors or harms and implementing corrective actions. Post-emergency after-action reviews should systematically assess whether AI tools were useful, what worked well, what failed, how tools should be modified, and whether unintended consequences or equity impacts occurred. Finally, agencies should define clear decommissioning criteria and transition plans for retiring models when they exhibit poor performance, conditions change, or better alternatives become available.

State-level legislative activity around governmental AI use has accelerated rapidly, with more than 150 bills introduced in 2024 and at least 30 states issuing formal guidance on AI use by state agencies. For example, Connecticut has implemented the AI Responsible Use Framework (Policy AI-01), which promotes ethical AI use, fairness, privacy, and transparency across all state agencies. Maryland enacted the Artificial Intelligence Governance Act of 2024, requiring each unit of state government to conduct inventories and assessments of systems that employ high-risk AI, and providing for the Department of Information Technology to develop policies and procedures concerning the development, procurement, deployment, use, and assessment of such systems [38]. Vermont established the Division of Artificial Intelligence within the Agency of Digital Services to review all aspects of AI systems developed, employed, or procured in state government [38].

At the federal level, several policies regulate the use of AI in governmental decision-making. The AI in Government Act of 2020 facilitates the adoption of AI technologies in the federal government, aiming to improve cohesion and competency in the adoption and use of AI within federal agencies. Executive Order 13960, titled “Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government,” establishes the policy of the United States to promote the innovation and use of AI to improve government operations and services in a manner that fosters public trust and confidence while protecting privacy, civil rights, civil liberties, and American values. Additionally, the Office of Management and Budget (OMB) issued Memorandum M-25‐21, “Accelerating Federal Use of AI through Innovation, Governance, and Public Trust,” which provides guidance for federal agencies to enhance their AI governance and risk management practices [39-41]. More recently, Executive Order 14,319 (July 2025), titled “Preventing Woke AI in the Federal Government,” required federal agencies to procure only AI models that prioritize truth-seeking and ideological neutrality, restricting the incorporation of diversity, equity, and inclusion frameworks into AI outputs [42]. This policy places constraints on public health AI decision-making by limiting the use of equity-informed analytic approaches that are commonly used to identify health disparities and guide targeted interventions. In December 2025, Executive Order 14365, titled “Ensuring a National Policy Framework for Artificial Intelligence,” established a uniform national AI framework intended to preempt state regulations deemed burdensome to innovation. By authorizing federal challenges to inconsistent state laws and allowing the withholding of federal funding from non-compliant states, the order reshapes public health AI governance by centralizing oversight, potentially reducing states’ ability to tailor AI safeguards, transparency standards, and equity protections to local public health needs [43].

[38]To strengthen AI regulations in government, several policies should be implemented. Governments should establish ethical AI frameworks that operationalize the fairness and bias mitigation principles discussed earlier, with continuous assessments to identify biases, ensure privacy protection, and evaluate AI system reliability. More states should implement policies like Connecticut, Maryland, and Vermont to regulate AI use within state agencies to ensure the ethical use of AI systems in government operations [38]. Additionally, mandatory impact assessments should be conducted before deploying AI systems, ensuring that risks, including bias and discrimination, are mitigated and disclosed publicly. Governments should also require AI systems to be explainable, allowing the public to understand the decision-making logic behind AI outcomes. For instance, New York has required state agencies to publicly disclose detailed information about their automated decision systems [44]. Public oversight should be ensured through independent governance structures, such as Chief AI Officers, who would be responsible for ensuring AI deployment aligns with ethical standards and transparency.

Building public trust in AI use requires active engagement with stakeholders, including community leaders, ethicists, and public health experts, to ensure AI applications meet the needs and values of the community. Governments must prioritize clear and accessible communication about how AI is used in decision-making, including detailed reports on AI system performance, challenges, and successes. Regular audits should be conducted to address potential biases, ensuring AI systems are fair and inclusive. Furthermore, ethical AI governance structures should be implemented, including ethics boards and external reviews, to monitor and guide responsible AI use. Creating avenues for public feedback, such as town halls or online platforms, will allow communities to voice concerns and provide input on improving AI systems. Strong privacy protections should be enacted to safeguard citizens’ data, ensuring AI systems are compliant with privacy standards. Lastly, ongoing public education and training on AI literacy will help demystify AI and foster confidence in its use, ultimately increasing public trust in governmental AI decision-making.

However, even with emerging governance frameworks, empirical evidence suggests that formal governance structures do not automatically translate into improved outcomes [45-48]. Cross-country studies find that standard governance indicators (government effectiveness, rule of law, transparency) show weak or inconsistent relationships with population health outcomes once other factors are controlled. Similarly, AI governance arrangements (such as ethical guidelines, transparency principles, and oversight bodies) do not necessarily translate into effective implementation at the operational level. This disconnect occurs because high-level institutional indicators may fail to capture sector-specific capacities essential for public health emergency response: data quality, frontline health system integration, decision-maker competence, organizational culture, and implementation fidelity. Successful AI governance requires attention to these operational realities, not just policy documents.

Drawing on the technical, ethical, organizational, and policy considerations discussed above, we propose an AI Decision Support Lifecycle framework specifically designed for governmental public health emergency response (Figure 1).

Phase 1 consists of defining the problem at hand and conducting a needs assessment. Stakeholders must clarify the decisions that will be made, the information that is required to help decision makers, the end users of AI outputs, and the consequences of wrong decisions. This phase involves engaging elected officials, public health leaders, emergency managers, and community representatives to ensure AI development targets genuine decision needs.

Phase 2 consists of equity-driven, technically robust model development and validation. Key details and steps include using representative data, evaluating performance across demographic and geographic subgroups, and documenting model limitations and uncertainty. Additional safeguards include independent expert review prior to operational deployment.

Phase 3 includes integrating AI models into workflows and pre-deployment testing. AI tools must be integrated into existing workflows and information systems. Steps include but are not limited to developing user interfaces compatible with incident command structures, training end users on interpretation and appropriate use, and establishing clear protocols for human validation and override of AI recommendations.

In Phase 4, a model is ready for deployment during an emergency. During an emergency, AI tools provide decision support while human officials retain ultimate authority. When AI is being deployed during an emergency, model performance must be monitored in real time, validating outputs using expert evaluation before high-stakes decisions, and clear communication of uncertainty to decision makers.

Monitoring and adaptation comprise Phase 5, which includes tracking accuracy metrics, comparing predictions to observed outcomes, and adjusting models as new data becomes available.

Finally, in Phase 6, models are evaluated after emergency deployment, and lessons learned are gleaned. Evaluation includes understanding the model’s usefulness for decision making, identifying what worked well and what didn’t as well as any unintended consequences, and understanding how the tool may be used or modified for future emergencies.

A number of critical principles apply throughout the phases of the AI Decision Support Lifecycle for Public Health Emergencies. First, it is vital that there are robust privacy and confidentiality protections in place and that there is limited data use for limited and specific purposes. Next, equity principles should be positioned as a key priority in model development and implementation. Similarly, there must be clear documentation for model development and use, explainability for model outputs, and public accountability for the use of AI in public health decision-making. Finally, human authority must remain the centerpiece of public health emergency decision-making. AI can only serve as decision support, never an autonomous decision-maker.

Table 1 organizes and synthesizes technical, ethical, organizational, and policy challenges that occur throughout the lifecycle, with specific actionable recommendations for model developers, health agencies, and elected officials.

Artificial intelligence holds substantial promise for supporting governmental public health emergency response, but responsible implementation requires adherence to core ethical principles rather than simply deploying technology. First, AI must function strictly as decision-support tools while preserving human authority. Elected officials and public health leaders retain ultimate accountability for decisions, with clear protocols for validating and overriding AI recommendations. Second, equity must be designed into AI systems from inception, not treated as an afterthought, through representative data, subgroup performance evaluation, community advisory boards, and transparent equity impact reporting. Third, successful AI deployment demands comprehensive organizational readiness, including workforce capacity (data scientists working collaboratively with epidemiologists and communication specialists), robust data infrastructure, governance structures, and sustained funding beyond initial acquisition costs.

Fourth, transparency and accountability mechanisms are non-negotiable, requiring public disclosure of when AI informs decisions, comprehensive model documentation, explainable outputs, and clear processes for challenging AI-informed decisions. Finally, continuous monitoring, evaluation, and adaptation throughout the AI lifecycle ensures systems remain reliable as conditions change, with pre-deployment testing, real-time performance monitoring, adverse event reporting, and willingness to decommission underperforming models. Achieving AI’s potential requires genuine collaboration between technical developers, public health agencies, and elected officials, with each stakeholder group taking specific, complementary actions to ensure models meet legal and ethical standards while meaningfully supporting evidence-based decision-making that reduces rather than exacerbates health inequities.