Who Owns the Value of Intelligence?

Open an AI contract at the clauses on data use, retention and exit. Somewhere in that administrative language is an answer to a question the procurement process may never have asked: who owns the value of the intelligence produced here?

The vendor supplies a model. Your institution supplies claims, case notes, transactions and scans: records made in the course of serving people, carrying the accumulated judgment of the people doing the work. What happens when these materials enter the service depends on its design and terms. They may be processed for a task; they may also be authorized for further use. Wherever that use creates an asset or improves a system, the contract helps decide who retains the benefit. An arrangement presented as a purchase can also distribute the means of future production.

That distribution is what I want to examine. Data is a raw material of intelligence, and sovereignty over it must reach beyond the moment of access to the value created afterward. Who may contribute, on what terms, and with what continuing claim? I regard this as a central economic question of the next decade. We keep approaching it as an administrative detail.

The raw material of a life

Consider the inputs. Computing power arrives with a price; engineering appears on a payroll. Data has a less uniform history. Some is bought or licensed, but much originates in an activity whose purpose lay elsewhere: a message to a friend, a claim submitted for payment, a consultation recorded so that care can continue. Before it becomes an input, the record belongs to a relationship. Its availability to a machine does not settle what obligations travel with it.

Martin Heidegger's account of enframing offers a way to see this conversion. Technology, in that account, organizes how the world becomes available for use. A river appears as a supply of power, a forest as timber: things become standing-reserve, held ready for a purpose assigned to them. Applied to data, the concept directs attention to the moment when an experience becomes a resource. The person remains in the record, even as the system learns to treat the record as material.

A patient's history can serve care, research or model development. A customer's behavior can become a prediction offered to someone else. The consequential move lies in the authorization that connects one use to another. It may occupy a few lines in the terms and conditions, passed over because the next useful service waits on the other side. Convenience makes the transition easy to complete. It does not make the transition economically neutral.

I have raised this concern since 2016. I have been part of numerous technology startups whose business models were oriented toward mining user data for predictions and business outcomes. That experience has kept my attention on the arrangement beneath the service: whose activity supplied the material, where its value accumulated, and where the consequences remained. The attention economy turned time and behavior into something saleable. Strains on attention, trust and public argument have many causes; I do not attribute them to a single system. I do see a recurring asymmetry between the capacity to capture value and the obligation to carry its costs.

The pressure to extract

Moloch is the name I use for the pressure that keeps this asymmetry in place.

The myth gives a figure to sacrifice made under the conviction that refusal is impossible. In its modern use, the trap is competitive: each participant gives up something worth protecting because restraint seems to leave the advantage to someone else. No common intention to destroy the thing is required. The damage can emerge from decisions that look defensible when examined one at a time.

Imagine a lab deciding whether to secure permission for training material while its competitor proceeds without it. A firm considering payment to contributors faces a rival that treats the same input as free; a buyer seeking better terms worries about delaying a deployment. Permission, compensation and scrutiny become costs borne by the party that insists on them. The pressure works through ordinary decisions about price and time. Each concession makes the next concession easier to describe as necessary.

This is why naming an extractor is insufficient. The arrangement rewards extraction even when the people administering it recognize its costs. A response has to alter that reward: make authorization consequential, give contributors bargaining power, and require buyers to account for what their purchases enable. The contract on the desk is one place where those pressures can be changed.

When the contribution outlives the exchange

The attention economy made behavior useful as a continuing stream of predictions. Leaving a service can interrupt the supply of new activity, though it does not establish what happens to the records already collected. AI sharpens the question of what survives the exchange. We need to ask what has been retained, what has been transformed, and which permissions continue after the immediate task is over.

Where records contribute to training, ending access to them is a different undertaking from undoing their influence on a model. A deletion promise does not, by itself, explain how either undertaking will be handled. The first task is to establish what actually happens to the material. A system processing a file for one request and a system using it to improve a model create different obligations. Procurement language should make that distinction legible.

Several features make the distinction difficult to enforce. Frontier-scale training places formidable demands on data and computing power. Concentration can then reinforce itself as organizations with resources attract further material and users, although more data is no guarantee of a better model. Closed systems limit what contributors can inspect. Behind the apparent autonomy sits human work: people labeling, sorting and cleaning material, often far from where its value accumulates. Karen Hao's reporting uses the language of empire to examine this distribution of power.

The economic difficulty goes further. The systems made with people's records can perform tasks for which people have been paid. I offer no forecast of the speed or extent of that change. The question is what happens to participation if a larger share of value is produced through systems while wages remain the principal route by which many people receive a share. Ownership of the inputs then bears on more than privacy. It bears on the terms under which people can continue to take part in the economy they help make possible.

Intelligence is abundant; responsibility is scarce. A contribution can be indispensable in the aggregate while leaving its contributors with no agreed claim on the result.

The infrastructure of sovereignty

Territorial borders remain consequential. Alongside them, computing establishes practical boundaries of access, dependence and control. Sovereignty has to be examined at both scales.

Benjamin Bratton's account of planetary computation supplies a useful diagram: earth, cloud, city, address, interface and user, organized as layers of a stack assembled without a single designer. Read this way, an interface is also a point at which conduct is permitted or refused; a cloud is also a jurisdictional and contractual dependency. Control of a layer can shape what the participants above and below it are able to do. Infrastructure mediates the exercise of law. It deserves examination as part of political geography.

This is the difficulty concealed by a promise such as 'AI abundance for all.' A service can be widely available while the terms of its use remain narrowly controlled. A standard can connect participants and still leave some of them unable to shape it. Availability tells us who may enter. The more demanding question is who can alter the arrangement once inside, and who was able to establish its defaults.

A stake in the value

The question I want to put to that discussion is simple: what would it take for an AI economy that depends on people's contributions to give them a stake in what those contributions create?

The familiar image of the sovereign individual offers an expensive answer: accumulate enough wealth to move money and self beyond a state's reach. I am interested in a condition a clerk in Manila or a nurse in Leeds could hold without first acquiring the means to escape. Sovereignty would be a feature of participation, available through the systems people ordinarily use. The contributor would have a part in setting the terms under which a record enters a shared model.

The proposal concerns ownership and participation at the point where value is made. A universal basic income distributes resources through a different mechanism. Here the aim is a continuing stake in the productive arrangement itself, rather than relying only on a later transfer. That aim does not yet solve how the stake should be measured or governed.

In the policy paper I wrote in 2025, I set out several working parts. They belong together because each addresses a different way a contributor can lose control of the exchange.

Under this proposal, people would own and permission the data that describes them. They would determine what is shared, with whom and for which purpose, with withdrawal made meaningful in the system's operations and terms. Ownership here is a proposed institutional baseline, not a description of a universal property right already in force.

Identity would be portable and private. A person should be able to establish who they are and what they authorized without disclosing every other part of their life. Consent needs a record that can travel without requiring the person to surrender the rest of the record.

Participation would include a return where contributed data helps create commercial value. The principle requires a distribution mechanism, a way to contest it and an honest account of what the mechanism can measure. A promise that value will flow back is only the beginning of a design.

Shared use and individual control would have to coexist. Aggregated, appropriately anonymized material can support purposes such as planning a city or studying disease; identifiable records carry obligations to the people they describe. Calling data a public good cannot settle every use in advance. The system has to specify how a shared benefit is pursued while individual rights remain enforceable.

Cooperatives and trusts could supply the bargaining capacity that isolated contributors lack. They would pool permissions and negotiate terms on members' behalf. Consider, as a proposed arrangement, a patients' cooperative setting conditions for research use of its members' records. Its power would reside in the ability to authorize a particular use, inspect compliance and withhold permission when the terms were breached. Collective participation would acquire an institutional form.

Public bodies would maintain the parts that require shared trust, while private firms could compete in the services built on them. The division would need explicit responsibilities on both sides. Public infrastructure does not govern itself, and competition does not by itself establish rights for the contributor.

A republic of intelligences

I call the wider arrangement a Republic of Intelligences. Its test would be whether participants can exercise the rights its name promises.

A republic distributes power and binds its exercise to rules that can be challenged. Human, machine and institutional intelligences now act together, but their capacities to authorize, inspect or refuse are uneven. The task is to organize their relationships so that no participant's convenience can quietly become another's condition of dependence.

The design begins with relationships and extends into the operations that sustain them.

Sovereignty is relational. My records and models exist alongside yours, within services and institutions neither of us controls alone. The relevant freedom lies in having terms for that interdependence: limits on use, means of redress, and a capacity to change partners. These are practical protections against domination. Isolation cannot supply them wherever collective activity is necessary.

Yuk Hui gives the problem a further dimension through cosmotechnics, his account of technology's relation to a culture's understanding of the world and its moral order. His argument for technodiversity challenges the assumption that one technological default should organize every society. Applied to this proposal, it suggests a common protocol that permits different rules for participation. A faith community, an indigenous nation and a city government might govern their data commons differently while retaining a means to connect. That interoperability is a design task, with differences to negotiate rather than erase.

Rights would also have to become operations. A consent record would constrain reuse; an exit provision would specify what can be transferred; an entitlement to a return would require a distribution process. Each provision makes an ethical commitment testable within the service. This is what I mean by sovereignty compiled into the stack: an obligation carried by the arrangement's behavior as well as its charter. The contract is a small, enforceable expression of the same political design.

Personal data vaults and consent wallets are possible components of that design. Civic cloud zones could bind hosted models to charters governing rights, environmental limits and value-sharing; data stewards, cooperative auditors and an ombudsperson could make those obligations answerable to people. These are proposals, not evidence that an integrated system already works. Their difficulty lies as much in durable authority, funding and enforcement as in technical construction.

The legal floor

The 2025 policy paper asked law to reach this distribution of value. The European provisions summarized below address parts of the surrounding problem. Their differences matter: access to a record, a limit on processing and an entitlement to proceeds are different kinds of claim.

In the European Union, the General Data Protection Regulation already gives a person the right to receive the personal data they provided, in a machine-readable form, and to move it elsewhere (Article 20). It also requires a vendor acting as a processor to act only on the customer’s documented instructions and to delete or return the data when the service ends (Article 28). The EU Data Act, which the Commission states has applied since 12 September 2025, gives users of connected products a right to the data their use generates and sets a framework for switching between data-processing providers (European Commission). The EU AI Act requires providers of general-purpose models to publish a sufficiently detailed summary of the content used for training, using a template from the AI Office (Article 53(1)(d)). The Commission states that those obligations entered into application on 2 August 2025 and that its enforcement powers began on 2 August 2026 (European Commission).

These provisions establish forms of access and portability, constraints on processing, and disclosure about training content. They give a contributor or customer specific questions to pursue within their scope. This remains a description of the published rules, subject to their conditions, and no substitute for the advice of counsel.

The provisions listed here do not themselves establish a general right for data contributors to share in a model's earnings. That is the gap this essay addresses. A framework for participation would have to account for the value created after access has been authorized, alongside the protections already attached to the record.

The limits of the proposal

The proposal has limits that should shape its design from the outset.

An individual record may command very little value on its own. There is no promise of personal wealth here. The case for cooperatives and trusts rests partly on that imbalance: a collective can bargain over a pool of contributions where an isolated person's leverage may be slight.

Attribution remains unresolved. Measuring how much one contribution adds to a model's performance is an open research problem. Any distribution formula adopted before that problem is adequately answered will contain approximations. Those choices must be disclosed and made contestable; a precise-looking payment does not establish a precise account of contribution.

Nor would sovereignty be served by sealing data into national silos. The proposal depends on shared protocols that allow local variation. Its promise has to be judged by whether people can participate across boundaries while retaining meaningful terms, rather than by how completely a jurisdiction can close itself off.

The institution on both sides

For a leader accountable for AI in a bank, insurer, health system or public body, these questions meet in the same contract. The institution contributes material to a service while holding that material on behalf of people with claims of their own.

As a contributor, your institution brings records, prompts, corrections and staff judgments about what a useful answer looks like. Establish which of these remain confined to the task and which may be retained or used to develop other assets. Where further use is permitted, ask who owns the result and what happens when you leave. The capacity to create value should be visible in the agreement that allocates it.

As a holder, the institution answers to the customers, patients or citizens described in those records. Their purposes do not disappear when the procurement team selects a service. A vendor agreement has to carry those obligations into the next relationship. You can delegate the task. The accountability remains with you.

Begin this week with the AI contract most consequential to your institution, whether already signed or still under negotiation. Put five questions to the vendor. Require a written answer tied to a clause and, where relevant, to an observable procedure. This is where the larger proposal becomes a practice.

1. Does any of our material train or improve a model, yours or another provider's? Include prompts, files, outputs, corrections and usage logs, and follow the permission through every company upstream. Ask which uses are prohibited by contract and which depend on a configurable setting. A current default is not an enduring commitment.

2. Who controls the assets made from our material? Name the fine-tuned models, embeddings, indexes and evaluation sets where they apply. Establish rights to use and transfer each asset, including after termination. The price of leaving can reside in a dependency built during the service, long before anyone asks to leave.

3. Where our material creates value for you, what returns to us and to the people it describes? Ask for the proposed basis of distribution, even if the answer is that none exists. Make the absence explicit. The negotiation cannot reach a question it has never put on the table.

4. What can we take when we leave, and what can be deleted? Specify data, derived assets and logs; require readable export formats, defined responsibilities and evidence of completion. Identify limitations before relying on the exit provision. The practical meaning of sovereignty becomes clearest when a participant chooses to change partners.

5. Who else can access our material, and under which jurisdiction? Trace staff access, sub-processors, model providers and legally compelled access. Locate backups and logs as well as primary storage. Establish what notice is required when access is requested, including any limits on giving that notice. A diagram of the service is incomplete until these relationships appear on it.

Read the answers for evidence of an operating arrangement: a clause, a named responsibility, a log that can be inspected. Reassurance leaves the allocation of power obscure. Record where the agreement supplies a right you can exercise and where it supplies only a promise you cannot yet test.

The terms of participation

Moloch's pressure depends on treating the terms of the race as given. Yet those terms are made and renewed in standards bodies, procurement offices and agreements of the kind now on your desk. Each permission for reuse, each retained asset and each obligation at exit helps determine what the system rewards. These are places where an institution can act.

One buyer cannot redesign an economy. It can make a use visible, refuse an unjustified permission or require a transferable asset. Repeated across institutions, such demands can become an expectation that suppliers have to answer. A broader right to participation needs law and collective organization; procurement gives that ambition a present place to begin.

Return to the data-use clause. It describes more than the journey of a file: it sets terms for whose contribution can become whose asset. Your data is your sovereignty. Ask who shares in what it creates, and require an answer that can be exercised after the contract is signed.