Digital technologies are now embedded within routine health care delivery across high-income and middle-income settings. AI-enabled diagnostic systems, predictive risk models, and remote monitoring platforms are integrated into primary care, hospital medicine, and community services [1]. These are frequently justified as necessary responses to workforce shortages, demographic change, and fiscal constraints. However, accumulating empirical evidence indicates that digital health technologies do not perform uniformly across patient populations and that differential performance has direct clinical consequences [2]. The central issue is not solely unequal access to innovation but also the lack of consideration of equity in product design. Clinical tools need to demonstrate reliability across the populations in which they are deployed, and generalizability remains a core requirement of safe medical practice. When performance varies systematically by ethnicity, socioeconomic position, or geography, this amplifies inequity that is already present in both the health system and wider society [3].

While concerns about algorithm bias have received increasing attention within the digital health care literature, the role of commercial developers, regulators, and procurers within the current health system as drivers of inequity has been largely overlooked. Throughout this Viewpoint, we use the term “commercial digital health” to refer to digital technologies developed and marketed by private sector organizations for deployment within health care systems. These include AI-enabled clinical decision support systems, predictive risk algorithms, digital triage platforms, remote observation technologies, and software as a medical device. In contrast to digital systems developed internally by health care organizations for local use, where governance, monitoring, and oversight structures are often embedded within existing organizational frameworks, commercial products are frequently deployed across multiple organizations and populations. However, the underlying training data, validation processes, and assessment of metrics may remain proprietary. This creates unique accountability challenges, particularly when procurement decisions are made without access to detailed evidence on subgroup performance, the representativeness of development datasets, or ongoing monitoring arrangements.

Recent frameworks by Liu et al [4,5] and Chin et al [5,5] have illustrated the importance of representative datasets, subgroup performance assessment, fairness monitoring, and fair outcomes across the AI life cycle. However, these approaches have primarily focused on algorithm development and performance. We argue that inequity in digital health is also shaped by commercial incentives, procurement decisions, and patterns of digital exclusion that determine who is represented in datasets and who ultimately benefits from innovation. This Viewpoint therefore broadens existing AI fairness frameworks by proposing an accountability model that spans the full commercial digital health life cycle, from product development and procurement through implementation and postmarket monitoring. We put forth the case that evidence of equitable performance should become a prerequisite for the procurement, implementation, and continued use of commercial digital health technologies. Just as safety and effectiveness are routinely assessed before adoption, demonstrable impartiality across relevant population groups should be considered a core requirement of trustworthy health innovation.

Inequities associated with digital health technologies arise through 4 key mechanisms that should not be considered interchangeable. First, measurement bias can occur when digital tools perform differently across populations. In a cohort of 10,789 patients, Sjoding et al [6] reported that occult hypoxemia occurred in 11.7% of Black patients compared with 3.6% of White patients when assessed by pulse oximetry, a device ubiquitous in acute and community care. Pfohl et al [7] highlight that predictive algorithms often experience performance degradation when deployed in populations that differ from those used during model development. This phenomenon, known as distribution shift, can disproportionately affect underrepresented patient groups and increase health care inequities.

Second, dataset bias can occur when training data used for machine learning software do not represent the population in which technologies are to be deployed. Obermeyer et al [8] examined a commercial risk prediction algorithm affecting more than 200 million individuals and found that it systematically underestimated illness severity among Black patients because it relied on health care expenditure as a proxy for need. As expenditure reflects access to services rather than biological burden, the model encoded structural inequity into resource allocation decisions; correcting the bias increased the proportion of Black patients eligible for additional care from 17.7% to 46.5%. This can be influenced by the third mechanism: algorithmic bias may emerge when models learn patterns that reflect existing inequities within health care systems. If underserved populations do not have equitable access to health care services, less data reflecting their actual health care needs will be available to train models. Seyyed-Kalantari et al [9] demonstrated underdiagnosis bias in deep learning systems applied to chest radiographs among underserved populations despite comparable aggregate performance. In dermatology, limited representation of darker skin tones in training datasets has been associated with reduced diagnostic reliability in underserved populations [10]. Fourth, digital exclusion can occur when certain populations face barriers to accessing or engaging with digital technologies. This presents fewer opportunities for subgroup analysis to be performed to assess model accuracy across populations—we may only know that the digital tool works for those who are already competent with health technology.

Although these mechanisms can interact and reinforce one another, they originate from distinct causes and require different approaches to mitigation. These findings differ in mechanisms but converge in implication: differential performance in widely deployed technologies alters clinical classification, escalation, and access to care. Figure 1 visually depicts these 4 mechanisms.

These disparities expose a structural epistemic vulnerability, whereby systemic inequity shapes the data from which AI systems learn. Machine learning systems are trained on historical datasets generated within health systems characterized by socially patterned access, investigation, and documentation. Electronic health records reflect who presents, who is investigated, and whose symptoms are recorded; however, these patterns do not represent true clinical need. Rather, they reflect the wider socioeconomic and cultural barriers to help-seeking behaviors or geographical disparities in service provision. For example, people with disabilities, individuals from ethnic minority groups, rural populations, people who are experiencing homelessness, and people from sexual and gender minority groups are frequently documented as having less access to health care [11]. Missingness and sampling bias can propagate through predictive models [12], and aggregate discrimination statistics may obscure clinically significant subgroup variability; a model with high overall discrimination can nonetheless systematically misclassify individuals within specific populations.

The implications are immediate—social injustice becomes encoded as algorithmic bias, and “majority” phenotypes and populations are prioritized in technological development. Algorithms influence triage thresholds, referral decisions, and allocation of scarce resources. Elevated false-negative rates in certain groups may delay diagnosis and treatment; differential calibration may misrepresent disease burden. In pharmacological and device evaluation, limited representation is recognized as a constraint on external validity. Digital systems, by contrast, may be deployed across regional or national infrastructures, affecting hundreds of thousands of patients without routine disclosure of stratified performance metrics. Rajkomar et al [13] argue that fairness assessment must be integral to clinical machine learning; however, consistent standards for subgroup reporting remain underdeveloped.

Rajkomar et al [13] argue that achieving fairness in health care AI extends beyond improving algorithm performance, requiring consideration of how data quality, model development, and implementation influence health equity. This highlights that reducing algorithmic bias requires organizational and clinical governance in addition to technical solutions. Liu et al [4,5] and Chin et al [5,5] have highlighted measurement bias, dataset bias, and algorithmic bias as important drivers of inequitable AI performance. We build on this work by arguing that commercial dynamics compound these methodological concerns, bringing to the surface additional mechanisms through which inequities may be reproduced and amplified during implementation, including digital exclusion and commercialization.

The development and refinement of digital health technologies are frequently driven by commercial incentives that prioritize rapid deployment and scalability, potentially at the expense of equitable design and long-term monitoring. For example, AI-enabled self-management tools for serious mental illnesses (SMIs), such as addiction and schizophrenia, lag substantially behind those for mild to moderate mental health conditions in development. Market shapes incentive: in the example of people with mild to moderate mental health conditions, there is less need for clinical governance and products can be sold directly to consumers. In comparison, people with SMI often also experience digital exclusion, fragmented service contact, and poverty. Consequently, the commercial return is smaller, while development and governance requirements are greater. Smith et al [14] argue that commercial AI development is fundamentally influenced by market incentives, meaning that populations generating the largest and most profitable datasets are often prioritized during product development. This raises questions regarding whether commercial development pathways preferentially prioritize populations who are easier to reach, retain, and monetize, potentially reinforcing existing patterns of exclusion and further driving algorithmic bias.

Aside from the example from SMI, the adoption of digital health technologies is socially patterned across diverse settings. A systematic review of 41 studies identified persistent socioeconomic gradients in digital health uptake, with lower engagement among individuals with reduced income, education, and digital literacy [15]. In the United Kingdom, approximately 6% of households lack internet access, with substantially higher exclusion among older and lower-income groups [16]. Iterative model refinement frequently relies on user-generated data; populations most engaged with digital systems therefore disproportionately shape subsequent optimization, rendering representation self-reinforcing and creating a feedback loop in which existing patterns of representation become increasingly entrenched. These dynamics highlight that algorithmic bias is not solely a technical phenomenon but is also shaped by the commercial, social, and structural contexts within which digital technologies are developed and deployed.

Available frameworks addressing AI fairness in health care have primarily focused on algorithm-level bias, including representativeness of training data, model effectiveness across demographic groups, and ongoing fairness monitoring throughout the AI life cycle. While these considerations remain essential, they largely assume that patients can access and engage with AI-enabled services. We argue that this assumption overlooks a critical upstream determinant of equity: market exclusion. For many populations, including older adults, people living with multiple long-term conditions, individuals with lower socioeconomic status, and those with limited digital literacy, inequity may arise before algorithmic interaction occurs. This could lead to an intervention-generated inequity, where only users who have the resources to engage with digital health tools do so, benefiting individuals with greater technological, educational, and socioeconomic resources [17]. This cycle is perpetuated by market needs. Consequently, we propose shifting from a narrow focus on AI fairness to a broader model of AI accessibility and inclusion that considers barriers across the entire commercialization digital care pathway, from access and engagement through to algorithm accuracy and long-term outcomes.

The European Union Artificial Intelligence Act [18] and World Health Organization [19] guidance have strengthened expectations for transparency, risk management, and postdeployment monitoring of AI-enabled health care technologies. These frameworks recognize that digital tools are increasingly embedded within routine care and that even modest differences in performance may generate substantial inequities when deployed at scale. However, existing governance approaches focus primarily on the safety, effectiveness, and fairness of algorithms. Less attention has been given to how commercial ownership, proprietary development processes, and procurement practices affect the ability of health care organizations to independently evaluate, monitor, and challenge these technologies once deployed.

The nature of commercial ownership limits health care organizations’ practical ability to verify that these principles have been achieved. National Health Service procurement frameworks, including Digital Technology Assessment Criteria [20] and NHS England’s AI Buyer’s Guide [21], require evidence of safety, effectiveness, and statutory compliance. However, they do not require comprehensive disclosure of proprietary datasets, model training procedures, or algorithm-based architecture, thereby rendering health care organizations largely dependent on evidence generated by manufacturers themselves. This constrains opportunities for independent scrutiny of likely biases, representativeness, and real-world validity, creating an asymmetry of knowledge between developers, health care organizations, regulators, and patients. Narrowing this implementation gap between development and transparent deployment necessitates a shift from regulation frameworks toward stronger mechanisms for developer accountability, independent evaluation, and representative postdeployment observation.

Within the high-income health care settings predominantly represented in the current evidence base, health care organizations should prioritize measures to ensure that AI systems are fair, transparent, and equitable. Current discussions often frame algorithmic fairness as a technical challenge for data scientists. However, responsibility for equitable performance should extend across the digital health ecosystem. Commercial developers are responsible for generating representative evidence, regulators for scrutinizing subgroup performance, health care organizations for considering equity during procurement, and health systems for monitoring real-world impacts after implementation. Without clearly assigned responsibilities, concerns regarding fairness risk remain aspirational rather than operational. Where innovation leaves particular groups behind, it undermines not only health inequity but also the evidentiary foundations of clinical practice. Ensuring that commercial digital health technologies meet standards of generalizability commensurate with the diversity of high-income patient populations is therefore not aspirational but essential to safe, effective, and ethically defensible health care in the digital era.