A researcher analyzes genomic sequencing data, illustrating the growing strategic importance of genetic databases in the development of AI-driven healthcare. Photo: Getty Images.

DeepRare is hailed as a breakthrough in the diagnosis of rare diseases. But more than that, it is symptomatic of a structural shift in the global health technology order, a shift in which the center of gravity of biomedical artificial intelligence (AI) is moving eastward, and the Global South must decide, urgently, whether to engage on its own negotiated terms or terms inherited by others.

DeepRare crystallizes a tension that will define health governance for decades to come. The same architecture that has the potential to democratize access to the diagnosis of rare diseases is designed in ways that reproduce and potentially deepen the informational hierarchy that has long structured global health injustices. Understanding these tensions requires us to move beyond the clinical register and into the geopolitics of data infrastructure, the power of databases, and the ongoing competition over who governs the knowledge layer of global health. It also requires us to face the historical record honestly that the dynamics at stake in DeepRare's global deployment have clear empirical precedents, and each reveals how the same governance failures are repeated across successive waves of health technology.

The power of databases as geopolitical terrain

The global genomics database is not a neutral scientific shared property. It is a politically constructed terrain, shaped by decades of investment, regulatory frameworks, and research priorities that are largely concentrated in the Global North. Asymmetry is fundamentally structural and the vast majority of participants in whole-genome association studies (GWAS) have historically been of European descent, while populations in Africa, Southeast Asia, and Latin America are chronically underrepresented in the reference data that underpin any diagnostic AI system built on top of them.

This is not a technical gap waiting to be corrected. This is a political condition with clinical consequences. Genetic variants endemic to African or Southeast Asian populations are systematically erroneously classified as pathogenic not because of misdiagnosis, but rather because epistemic infrastructure such as reference panels, ontological databases, clinical literature is built without them. The H3Africa consortium has documented this appropriately. The absence of African variants from global reference databases is not just a gap that needs to be filled and it is a structural feature of the way the global genomic order is constructed.

DeepRare inherited this order. Its ability to operate in a data-poor environment is a meaningful architectural feature, but it does not solve upstream problems. A system can be architecturally adaptive as well as epistemically Eurocentric. The sophistication of the model cannot compensate for the political conditions that generate the data it trains. What DeepRare offers Global South patients is better diagnostic processing in a frame of reference that was never designed to represent them and that remains regulated by institutions that they do not influence.

Three historical precedents: same failures, three different registers

The structural risks embedded in DeepRare's global deployment are not speculative. They replicate, in a new technological register, patterns of governance failures that have been repeated in three different episodes of the history of biomedical data. Each case illuminates different dimensions of the problem together, they form an empirical map of how health data extraction operates when governance is absent or co-opted.

The first precedent is the Havasupai case, which remains the most instructive example of how consent frameworks governing biological data are structurally inadequate when collecting institutions control data and their subsequent use. Between 1990 and 1994, members of the Havasupai Tribe in Arizona donated blood samples to researchers at Arizona State University for a study of type 2 diabetes. A disease that the community itself is asking to be researched. A decade later, the tribe discovered that their DNA had been reused for unrelated studies. For example, about schizophrenia, population migration, and inbreeding. Research that touches on topics that are considered culturally taboo and highly stigmatized by Havasupai. The critical analytical point is not that the researchers acted with obvious malice; rather, that the governance architecture gives the data collection institution unilateral authority to determine the scope of permitted use after the data has been submitted. The community does not have the standing to question the determination. The 2010 settlement returned blood samples, but critically, it did not set a binding legal precedent that means the structural vulnerabilities it revealed remain unresolved to this day. 

The second precedent is the case of DeepMind-Royal Free, which translates the same governance asymmetry into the digital health AI register. In 2015, Google DeepMind acquired access to the medical records of 1.6 million patients from the Royal Free London National Health Service (NHS) Foundation Trust, nominally to develop an app called Streams to detect acute kidney injury. The UK's Information Commissioner's Office (ICO) later ruled that the arrangement failed to comply with data protection laws because patients were not adequately informed, stated objectives were far less specific than the actual scope of data access, and the pretext of "direct care" used to bypass consent requirements was applied to patients who had never been treated for kidney conditions including some who had died. The analytical significance is that the same architecture that characterized DeepRare's deployment model was already present in 2015: a private AI company positioned its medical device as a clinical benefit to gain access to patient data at a population scale, and the existing governance framework was structurally incapable of limiting what the company could do with that data once it was in their systems. The ICO ruling is a corrective action after the fact rather than the previous prevention. For Global South countries that adopt DeepRare without an equivalent regulatory infrastructure, an ex post corrective mechanism is completely absent.

The third precedent, and the most directly analogous to the case of DeepRare, is the global expansion of BGI Group, a Chinese state-backed genomics company. BGI builds a global data collection infrastructure through the strategic deployment of low-cost or subsidized healthcare services, especially prenatal testing kits sold in more than fifty countries that collect genomic data from populations around the world while positioning BGI as a generous health technology provider. The U.S. Department of Commerce eventually placed several of BGI's subsidiaries on its Entity List over findings that their collection and analysis of genetic data posed a risk of contributing to Chinese state surveillance and military programs. As Stanford Law School's analysis of the U.S.-China genomic data competition documented, China has treated human genetic data as a strategic resource and regulated its outflows from Chinese territory since 1998 because Beijing understands that genomic databases are a form of national power. The BGI case is a structural template for understanding DeepRare: open-access positioning as a market entry strategy, subsidized or free services as an adoption incentive, and the accumulated data flow into a strategic asset regulated by the country that controls the platform.

DeepRare in a geopolitical context: setting the standard as foreign policy

Treating DeepRare as a stand-alone scientific product is misreading its geopolitical significance. It is one node in a deliberate, state-supported program to position China as the next-generation architect of global health infrastructure and its logic is read through the BGI precedent. The difference is the register. BGI operates through genomic sequencing services while DeepRare operates through diagnostic AI.

This is the logic of the power of standard-setting applied to health technology. Rather than reflecting philanthropy, DeepRare's open-access model serves a strategic function: broader adoption expands the pool of clinical data available to the platform, enabling continuous improvements in diagnostic performance and reinforcing its role as the dominant technological standard in rare disease diagnostics.

It represents the next phase of China's medical soft power strategy that is categorically more enduring than its predecessor. Physical health infrastructure such as donated hospitals, medical teams sent projects influence through invisible, limited, and politically reversible presence. Diagnostic AI platforms project influence through dependencies, are invisible, distributed throughout clinical workflows, and are structurally difficult to abandon once embedded. When a hospital in Phnom Penh or Nairobi integrates DeepRare into its diagnostic pipeline, it enters a governance relationship where patient data flows into an infrastructure governed by China's institutional norms including the Beijing Health Law Society's "Deployment of DeepSeek Expert Consensus" where the adoptive country has neither jurisdiction nor a voice.

The paradox of access and why denial is not enough analytically

Geopolitical critics must honestly confront structural paradoxes and for much of the Global South, the alternative to DeepRare is not a better-governed system but no system at all. This, it should be noted, is the same structural condition that has allowed BGI's global expansion i.e. the absence of affordable alternatives creates a void that subsidized Chinese providers can fill. The diagnostic results of epistemicly imperfect tools may substantially surpass the diagnostic results of no tools. This is not an argument for uncritical adoption; This is an argument against the analytical convenience of a simple rejection.

This access paradox also reveals the deeper strategic problems facing the Global South. The Global South is offered a choice between relying on Western proprietary systems and relying on Chinese open-access platforms. The BGI case shows that the consequences of this choice are not just commercial but strategic. The real stake is not which platform is adopted, but rather whether adoption can be converted into the leverage needed to restructure the terms of participation in global health AI as a whole.

Reconfigured bio-colonialism: the problem of governance asymmetry

The theoretical framework of bio-colonialism requires an update for the AI era. The classical formulation assumes a diad that considers the Global South as the source of data, the Global North as the producer of knowledge. DeepRare complicates this geography without dissolving it. China's platform now occupies a position historically held by Western institutions but the asymmetry of governance, the structural conditions that make extraction possible and recovery difficult, are fully reproduced. The Havasupai case confirms that even in jurisdictions with functioning legal institutions, the absence of ex ante governance requirements leaves affected communities without a path to recovery. The DeepMind-Royal Free case confirms that even with national data protection regulators, corrective mechanisms operate after the damage has occurred. For Global South countries that adopt DeepRare, none of these conditions apply which means there is no data protection authority equivalent to jurisdiction over China-regulated infrastructure, and no ex post corrective mechanism at all.

As documented by Frontiers in Genetics, the global circuit of for-profit genomic knowledge production remains structurally intact regardless of which institution occupies its center. The Lancet Digital Health further documents how LMICs systematically hand over ownership of data in research partnerships structured by high-income state institutions. What DeepRare's deployment model introduces is a more consequential variation of previous patterns. Unlike episodic research partnerships, continuously deployed diagnostic platforms generate a continuous and growing flow of patient data into governance frameworks that are not controlled by contributing countries. Governance asymmetry is a critical variable and this is a condition that all three historical precedents fail to resolve.

Restructuring participation: the health data diplomacy agenda

The appropriate policy response to DeepRare is strategic engagement, the conversion of adoption leverage into governance rights. Historical precedents are also instructive here. What the Havasupai case, the DeepMind-Royal Free case, and the BGI case have in common is the failure to establish governance conditions before dependency is created. The policy agenda must be designed to prevent such failures from recurring.

Data sovereignty must be operationalized as a non-negotiable contractual condition. Countries that adopt DeepRare or equivalent systems must negotiate explicit provisions on data localization, access controls, retraining rights, and dispute resolution prior to adoption. The European Union's GDPR and the African Union's Malabo Convention on Cybersecurity and Personal Data Protection provide an adaptable framework. The 2021 ASEAN Guidelines on AI Governance and Ethics require substantial strengthening specifically for health data, but provide regional institutional anchors. The principle is structural, i.e. the governance requirements must precede technical integration because, as the DeepMind-Royal Free case shows, attempts to impose such requirements after the fact, once contested, and practically inadequate.

Reciprocal representation should be demanded as a prerequisite for participation. The logic is both ethical and technical as research has shown that the increased representation of the Global South in genomic reference databases substantially reduces diagnostic misclassification. The H3Africa Consortium model that makes data ownership provisions a prerequisite for international genomics partnerships provides an operational template. The BGI case further demonstrates what happens without such conditions: data flows out without any mutual benefit, and corrective efforts only arrive after strategic asymmetry has been consolidated.

The multilateral governance architecture must be accelerated beyond the current pace of development of the WHO's normative development. ASEAN and the African Union should develop a binding interim framework for cross-border health AI governance. The COVAX mechanism shows that health-specific multilateral governance can be built quickly when there is political will. The BGI case in which the US BIOSECURE Act only arrived after China's genomic infrastructure had penetrated more than twenty countries and demonstrated that reactive governance was structurally inadequate. Health data sovereignty requires an architecture built before the dependency crisis was fully consolidated.

The map already exists, what is lacking is the courage to read it

DeepRare is more than just a diagnostic tool. It is a geopolitical event that reveals, in a concentrated form, the structural conditions of the global health AI order that is being formed. Historical precedents are unambiguous either through the disclosure of consent framework failures by the Havasupai case, the demonstration by the DeepMind-Royal Free case of how AI companies use health benefit claims to obtain data at the population scale, or the mapping by the BGI case of how Chinese state-backed genomics companies are using open-access strategies to build a globally consistent data infrastructure. Governance that is absent at the point of adoption cannot be restored once a dependency is formed.

The Global South does not face a binary choice between access and sovereignty but it must act on both simultaneously, and must act before the architecture of the next health order becomes fixed. The decisions being made now about which platforms are adopted, under what conditions, and under whose governance, will not only determine who can diagnose rare diseases but who controls the global health knowledge infrastructure for a generation to come. History has provided its empirical map. The question is whether the political will to read it exists.