Ownership debates in genetics have usually been framed in physical terms. A sample is taken, a sequence is generated, a repository stores it, and a company or laboratory seeks permission to use it. Law and ethics followed that chain: consent, custody, privacy, patentability, benefit-sharing. Yet by mid-2026 the more consequential contest is migrating away from the vial and towards the model. Genomic data are increasingly ingested into large predictive systems that learn relationships among variants, expression patterns, ancestry signals, disease risks and molecular function. Once that happens, the original sample may disappear from view while much of its economic and scientific value persists in statistical form.
This shift matters because existing rights regimes were built to govern identifiable materials and datasets, not weights, embeddings and generative capabilities. A rights architecture designed for biobanks can struggle when a model no longer stores a genome in any ordinary sense, yet still allows its characteristics to be inferred, simulated or monetised. The strategic question is no longer only who has your DNA, but who can regenerate its value from a model.
From specimen logic to model logic
The modern legal treatment of human genetic material has long been marked by an awkward compromise. In many jurisdictions, people do not straightforwardly own excised tissue as property in the ordinary commercial sense, but institutions that process, sequence and curate it may exercise strong practical control. At the same time, patents on naturally occurring DNA sequences were narrowed sharply in the United States by Association for Molecular Pathology v Myriad Genetics, which held that naturally occurring DNA is not patent eligible merely because it has been isolated. That settled one argument without resolving the larger one. If genes are not ownable as discoveries, and bodies are not reducible to chattel, what exactly is the asset?
For a time, the answer was: the dataset. Access-controlled genomic databases, cloud repositories and linked health records became the defensible resource. But model-centric biology changes the unit of value. A well-trained model can capture regularities from millions of genomes and related phenotypes, then support downstream drug discovery, clinical prediction or synthetic design without redistributing the underlying records. Genomic property is becoming less about exclusion at the point of collection than about leverage at the point of inference.
Why genomic models complicate privacy
Genetic data are already treated as unusually sensitive. The GDPR classifies genetic data as a special category of personal data, and European health-data policy has tightened around secondary use, interoperability and safeguards. The OECD’s recommendation on health data governance likewise emphasises trust, stewardship and public interest conditions. Those frameworks assume, correctly, that genomes are both identifying and enduring. A person cannot rotate their DNA as they would a password.
Yet models trained on genomic material create a more ambiguous object. They may not contain an intact individual sequence in a directly retrievable form, but they can still preserve information about populations, kinship structures and variant-disease relationships. In machine learning, privacy risk does not vanish when data are abstracted; it changes shape. Membership inference, model inversion and leakage from embeddings have already become familiar concerns in other domains. In genomics, the stakes are higher because even partial reconstruction can expose familial facts, ancestry clues or predispositions that affect more than one person.
The strategic question is no longer only who has your DNA, but who can regenerate its value from a model.
The strategic question is no longer only who has your DNA, but who can regenerate its value from a model.
The problem of derivative genetic knowledge
Most consent mechanisms still revolve around primary and secondary use of data. Participants are informed that their samples may support future research, perhaps including commercial work, often under broad consent. What these instruments generally do not capture well is the production of derivative genetic knowledge: a model that learns from thousands or millions of contributors and then behaves as a reusable knowledge asset. Once trained, it may enable variant interpretation, biomarker discovery, target identification or sequence design in settings far removed from the original research context.
This derivative layer is awkward for both privacy and property law. Privacy law asks whether personal data remain personal when embedded in a model. Intellectual-property law asks whether the model itself, or its outputs, merit protection through copyright, trade secrets, database rights or patents. Neither inquiry tracks the moral intuition many participants hold, namely that value extracted from a collective genomic resource should remain subject to some continuing social obligation.
Why existing IP categories fit poorly
Patent law offers only limited traction. Naturally occurring sequences are not generally patentable as such, while many model outputs will look more like predictions than inventions. Copyright is a poor fit for facts of biology, though software code and some model artefacts may be protected. Trade secrecy is therefore likely to become the preferred enclosure mechanism: not ownership of the genome itself, but secrecy around model architectures, parameters, tuning methods and proprietary labelled corpora. A sequence may be hard to own outright, yet the system that predicts from it can become a durable private asset.
This is one reason the centre of gravity in genetic rights is shifting. Earlier fights focused on whether someone could claim too much over the raw materials of life. The next fights concern whether control over trained systems can function as a substitute for control over life’s informational patterns. In practice, model governance may matter more than sequence patent doctrine.
A public-goods dilemma for biobanks
Public and university biobanks were often justified as civic infrastructure. They pooled samples under promises of scientific openness, public benefit and controlled access. The NIH Genomic Data Sharing Policy, for example, rests on the idea that broad sharing accelerates knowledge while protecting participants through governance. But when external actors can train powerful models on access-controlled data and then retain the resulting systems as quasi-private assets, the bargain changes. The public may provide the substrate; the durable advantage may sit elsewhere.
This is not an argument against private participation. It is a reminder that the location of value extraction has moved. If biobanks become training grounds for exclusive models, access committees are no longer merely deciding whether a project is meritorious. They are deciding whether a collectively assembled biological resource may be converted into an enduring computational moat. Traditional material transfer agreements and publication rules were not designed for that choice.
Collective rights, not just individual rights
Genomic property is becoming less about exclusion at the point of collection than about leverage at the point of inference.
Genetic governance has always had a collective dimension because DNA is relational. One person’s sequence reveals information about relatives, communities and, in some circumstances, Indigenous or geographically distinct populations. International law has already grappled with analogous issues in biodiversity through the Nagoya Protocol’s access and benefit-sharing logic, though human genetic resources sit in a different legal and ethical category. The relevant lesson is structural: where biological material has shared provenance and uneven bargaining power, individualised consent does not exhaust the justice question.
In the model era, collective claims become stronger, not weaker. Population-specific models can generate high value precisely because they capture patterns that belong to a group history rather than an isolated individual fact. If a model trained on a rare founder population yields superior disease prediction or therapeutic insight, conventional privacy notices seem thin compensation. The issue is not simply whether participants agreed to data use, but whether communities retain any say over how their inherited statistical distinctiveness is capitalised.
What the EU AI regime does and does not solve
Europe’s AI Act introduces a risk-based framework that will affect many medical and biometric applications, while the proposed European Health Data Space seeks to structure secondary use of health data at scale. Together, they create a more serious governance environment than the laissez-faire posture often seen elsewhere. They require documentation, risk management and, in some contexts, heightened scrutiny for high-risk systems. That is useful, especially where genomic models feed clinical decisions.
But these instruments are not ownership regimes. They regulate deployment, safety, accountability and access conditions more than they allocate downstream claims over value. A compliant system may still convert shared genomic resources into concentrated private advantage. Nor do these frameworks fully settle when trained models should be treated as carrying residual personal-data risk. The law is improving at supervising use, but it remains less capable of governing appropriation through abstraction.
A sequence may be hard to own outright, yet the system that predicts from it can become a durable private asset.
Synthetic biology raises the temperature
The ownership question becomes sharper when genomic models are used not merely to predict disease but to design biological functions. WHO guidance on human genome editing has emphasised governance, oversight and public legitimacy. In parallel, AI-assisted design tools are making it easier to propose proteins, regulatory elements and other biological constructs that need not mirror any one natural sequence. Here the legal debate veers from privacy towards inventorship, dual use and the boundaries of permissible commodification.
If a model has internalised functional lessons from human and non-human genomic corpora, who owns the resulting designs? The answer will often be whoever owns the laboratory output or patentable invention, not those whose data helped teach the model biology. The distance between contributor and product grows. So does the plausibility of saying that nothing identifiable was used in the final artefact. That may be formally true and politically insufficient.
A sequence may be hard to own outright, yet the system that predicts from it can become a durable private asset.
The rise of non-rival extraction
One reason model-based appropriation is so difficult to contest is that it is non-rival. A biological sample can be depleted, lost or withheld; a model can be copied, fine-tuned and redeployed at low marginal cost. This weakens the intuitive link between possession and control on which many rights narratives depend. People understand a tissue sample in a freezer. They struggle to picture a latent space that has absorbed disease associations from hundreds of thousands of genomes.
That opacity alters bargaining power. Participants may be assured that their records remain secure within a trusted environment, while the real commercial and strategic asset is generated just beyond the perimeter. The older politics of data localisation and repository access still matter, but they no longer mark the whole battlefield. Ownership by containment is less effective when extraction occurs through training.
What a more serious rights framework would recognise
A credible framework for genetic rights in the model era would start by distinguishing among custody of samples, access to datasets, and control over derivative models. These are different powers and should not be collapsed into one consent checkbox. It would also recognise that some benefits from genomic AI are best treated as obligations of stewardship rather than as discretionary corporate generosity. The language of fiduciary duty, trusted research environments and public-interest licensing is more promising than the fantasy that individuals can negotiate fair terms one by one.
Such a framework would probably avoid declaring personal ownership of DNA in any absolute sense; that route has always been conceptually unstable and ethically crude. Instead it would attach governance duties to the institutions that convert human biological information into reusable intelligence. Those duties could include transparency about model training, limits on downstream exclusivity where public biobanks are involved, stronger community representation, and technical auditing for privacy leakage and group harms. The point would be not to freeze science, but to ensure that abstraction does not dissolve responsibility.
The politics ahead
Courts and regulators are unlikely to settle this quickly. They will face a familiar mismatch between legal categories and technical reality. Personal-data law can reach too far or not far enough; IP law can reward engineering while ignoring provenance; research ethics can protect the moment of collection while overlooking the life of the model. Meanwhile, health systems want innovation, states want domestic capability, and researchers want access at scale. Those imperatives will not disappear.
The likely result is a period of institutional improvisation: revised biobank contracts, stricter data-access conditions, new expectations for model reporting, and growing pressure for benefit-sharing arrangements tied to computational derivatives rather than only to samples. Much of this will occur below the level of headline legislation. Yet that is where the next settlement on genetic ownership may actually be written.
For two decades, the emblematic question was whether anyone could own a gene. By the second half of the 2020s, that is no longer the most revealing question. The more important one is whether societies will allow the informational value of shared human biology to be enclosed once it has been transformed into predictive machinery. Ownership, in other words, is migrating from the molecule to the model. Law has noticed the data. It is only beginning to notice the derivative power.
