The last decade has seen an explosion of AI tools approved for clinical use, yet fewer than 4% of the studies submitted to the US Food and Drug Administration (FDA) reported any racial or geographic diversity, and almost none disclosed socioeconomic data [1]. Systematic assessments show that the publicly accessible training corpora for clinical AI are drawn disproportionately from a small cluster of wealthy nations, leaving most of the world essentially invisible in the data landscape. Among 7314 PubMed-indexed clinical AI studies analyzed in a 2022 global review, 71% of the datasets originated in just 10 high-income countries (HICs), with the United States alone contributing 41%. This underscores the pronounced geographic skew that shapes model development [2]. Predictably, the devices trained on data from HIC cohorts have shown unstable performance when deployed in low- and middle-income country (LMIC) settings, undermining confidence in AI as an enabler of universal health coverage [3]. The resulting “health-data poverty” [4] is no longer merely a research inconvenience; it is an equity crisis that threatens to widen outcome gaps precisely where disease burdens are greatest [5].

In this paper, we outline how LMIC-led, quality-graded, openly benchmarked data frameworks, exemplified by the Medical Imaging Datasets for India (MIDAS) program, can provide a practical solution for transitioning from data poverty to data sovereignty. Data sovereignty refers to institutional decision-making authority and accountability over the full data lifecycle (including collection, curation, access terms, downstream validation, and benefit-sharing) exercised by the institutions closest to the data subjects. It is distinct from data localization, which is primarily a jurisdictional requirement governing where data are stored or processed, and from data access, which concerns permissions to use data without necessarily shifting custodianship or governance authority. By enabling locally governed, demographically annotated datasets, this approach also supports equity for historically marginalized populations within LMICs, including rural communities, ethnic minority groups, and socioeconomically disadvantaged groups whose clinical representation is systematically absent from global AI benchmarks. We also propose a multilateral South-South Data Commons built on 3 pillars: a harmonized dataset-grading rubric, distributed custodial governance, and outcome-linked incentives for AI-enabled products.

Commercial enthusiasm has yielded more than 1000 FDA-listed AI-enabled devices, yet transparency gaps remain striking. Only 3.6% of regulatory dossiers disclose race or ethnicity, and fewer than 1% report socioeconomic indicators. Among the regulatory dossiers, 46.1% provided detailed performance data, 1.9% provided a link to a scientific publication, and only 9% included a prospective study for postmarket surveillance [1]. Similar deficits plague public benchmark datasets: a systematic review in The Lancet Digital Health identified major documentation shortfalls in more than 75% of the COVID-19 pandemic datasets most frequently reused for AI development [6].

Although radiology societies advocate for radically open corpora, the public availability of population-level imaging datasets remains scarce, and nearly all such datasets originate from high-income settings. Many studies have highlighted this geographical skew in datasets. Using the MEDLINE database, Google’s search engine, and Google Dataset Search, a set of 94 open access datasets containing 507,724 images and 125 videos from 122,364 patients were identified. Most datasets originated from Asia, North America, and Europe [7]. A 2023 review of 110 magnetic resonance imaging (MRI) datasets found no datasets from LMICs, underscoring how this geographic skew hampers the development of robust, generalizable AI models for global use [8]. The resulting geographic skew is now well cataloged: models trained on US and EU chest radiograph data lose up to 25% points in the area under the curve when validated in African populations, even after standard transfer learning [9]. The WHO 2024 guidance on AI governance warns that nonrepresentative datasets can entrench structural inequities at scale and urges member states to cultivate AI learning ecosystems that advance equity and safety [10].

The current solutions for providing LMICs with access to datasets from HICs treat the problem as a simple data shortage rather than as an issue of unequal power. The concept of data sovereignty reframes the debate: the origin of the data determines both their scientific reliability and the authority over their use. A BMJ Global Health scoping review shows that local stewardship improves adherence to context-specific ethical norms and accelerates regulatory approvals [11]. Sovereignty, however, demands more than national firewalls; it requires interoperable standards that enable cross-border collaboration without ceding custodial control. In this manuscript, data sovereignty is not treated as synonymous with data localization or access restriction, but as institutional agency across the AI development, validation, and deployment lifecycle. Table 1 summarizes the differences between the prevailing HIC-centric AI development pipeline and the proposed LMIC data sovereignty model across key stages of the AI lifecycle. One concrete step in this direction is the MIDAS platform, which is a collaborative effort between the Indian Council of Medical Research (ICMR), the Indian Institute of Science (IISc), and AI and Robotics Technology Park (ARTPARK) to develop high-quality, standardized, and diverse datasets for developing contextually relevant, AI-driven health care solutions in India. Using a hub-and-spoke model, MIDAS enables the systematic collection, annotation, and harmonization of multimodal clinical data.

To operationalize these principles, MIDAS uses a 4-domain Dataset Quality Matrix covering representativeness, documentation, technical fidelity, and governance. Scores (0‐100) map onto Bronze, Silver, Gold, Platinum, and Diamond grades, with an external audit conducted prior to public release (Multimedia Appendix 1). The first gold-graded corpus, comprising 1000 oral-lesion images drawn from 6 Indian cancer centers, was released in October 2024 under a Creative Commons Attribution–NonCommercial (CC-BY-NC) license [12]. Recently, the dataset was onboarded onto AIKosh [13], India’s national AI repository. Within 1 month, the dataset ranked third among trending datasets on the platform (out of 5577 datasets) and was downloaded 738 times, indicating early uptake by the AI research and developer community. A Gold-graded dural lesion dataset followed in March 2025, codeveloped with the Department of Neurosurgery, All India Institute of Medical Sciences (AIIMS), New Delhi. Crucially, both datasets are anonymized; however, they include demographic fields such as age, gender, and ethnicity, which are required for developing robust AI applications, as well as versioned consent artifacts, thereby addressing the transparency gaps currently evident in FDA filings.

Global data-sovereignty efforts such as the African Health Data Space and the Organisation for Economic Co-operation and Development (OECD) Health Data Governance Principles have emphasized legal interoperability, ethical safeguards, and cross-border data-sharing norms. These frameworks are critical for establishing trust and harmonization but largely stop short of operationalizing sovereignty within the AI development lifecycle itself. MIDAS complements these initiatives by introducing a quantitative, auditable dataset-grading framework that links local stewardship to model validation, benchmarking, and outcome-linked incentives. Rather than focusing solely on access governance or data flows, MIDAS embeds sovereignty at the level of dataset readiness and evaluative authority, enabling LMIC institutions to influence how AI systems are tested, certified, and rewarded. This positions MIDAS not as an alternative to existing governance frameworks, but as an implementation layer that translates global principles into measurable, AI-relevant practice.

Building on the MIDAS foundation, we propose a 3-pillar framework for establishing a South-South Data Commons that transcends bilateral collaborations to create a genuinely multilateral infrastructure (Figure 1).

Replicating MIDAS without coordination risks a balkanized landscape of incompatible locally developed metrics. We therefore endorse the adoption of a common, open-source rubric extending MIDAS to nonimaging modalities and aligning it with the emerging Medical AI Data for All (MAIDA) framework for global medical datasets [14]. These perspectives have called for a “shared quality lingua franca” to facilitate cross-registry discovery [15]; a South-South rubric jointly maintained by the ICMR, the IISc, and the African Health Data Space could satisfy that mandate while preserving local autonomy [16].

Distributed custodial governance within the proposed South-South Data Commons is operationalized through a clear separation of roles across data stewardship, technical validation, and trust certification. Primary custodianship remains with the institutions closest to the data subjects, which retain authority over consent management, ethics approvals, and long-term stewardship of datasets. Technical validation can be distributed and conducted by designated nodal evaluators such as public research institutions or accredited AI evaluation laboratories with domain expertise using predefined grading rubrics such as the MIDAS Quality Matrix. Trust and certification are established through publicly accessible, versioned artifacts, including dataset grades, audit summaries, and metadata records hosted on national platforms such as AIKosh. The Trustworthy Evaluation of Clinical AI consortium has demonstrated that external evaluators can audit diabetic retinopathy algorithms using encrypted uploads without exposing patient-level data [17]. This functional separation allows local institutions to retain sovereignty while enabling credible, repeatable, and scalable validation, thereby preventing centralization and reinforcing mutual accountability within the Commons [17].

Market pull is indispensable. England’s National Health Service (NHS) AI Award releases final-stage funds only when applicants’ algorithms meet accuracy and fairness targets on the test partition of the National COVID-19 Chest Imaging Database [18]. By hard-wiring gold-standard dataset performance into payment schedules, these programs turn data quality into a revenue driver and create a self-reinforcing loop: vendors eager for outcome bonuses invest in expanding and diversifying the very datasets that will be used to evaluate the next procurement round. Governance is operationalized through transparent evaluation criteria, third-party validation, and outcome-linked disbursement rather than discretionary funding, thereby operationalizing sovereignty as participation in value creation rather than passive data provision (Figure 2).

Transitioning from data scarcity to sovereignty is more than an LMIC agenda; it is a prerequisite for valid, generalizable AI everywhere. Commercial developers seek regulatory clarity, journal editors demand reproducibility, and patients deserve equity by design. A South-South Commons anchored in graded, openly documented datasets would convert what WHO terms “an ethical imperative” into a tangible, auditable infrastructure. HIC regulators and payers stand to benefit: models validated on Commons data will arrive already stress-tested across demographic gradients that do not exist in single-country registries, reducing the risk of catastrophic postdeployment failures.

Algorithmic equity cannot be retrofitted after deployment; it must be engineered upstream through the data themselves. By institutionalizing rigorous quality grading, distributed custodianship, and value-linked incentives, LMICs can recast themselves from data supplicants to data sovereigns producing datasets that are not only locally trustworthy but also globally indispensable. The blueprint outlined here, grounded in the MIDAS experience and amplified through a South-South Data Commons, offers a practical, scalable route toward that future.

From a policy perspective, operationalizing data sovereignty does not require a wholesale system redesign. Immediate steps for LMIC governments and research agencies include formally adopting transparent dataset quality grading standards within publicly funded research programs; designating or accrediting national institutions to perform independent, privacy-preserving AI validation; and embedding gold-standard dataset performance requirements into public procurement, regulatory evaluation, and grant funding criteria. National data platforms can be made interoperable with Commons architectures by publishing versioned metadata, audit artifacts, and validation outcomes rather than raw data alone. Taken together, these measures allow sovereignty to be exercised incrementally through evaluative authority, incentive alignment, and regulatory participation while remaining compatible with cross-border collaboration and global AI development norms.