Where Liability Collects

Lately in a wreck of a Californian ship, one of the passengers fastened a belt about him with two hundred pounds of gold in it, with which he was found afterwards at the bottom. Now, as he was sinking—had he the gold? or the gold him?

Consider a constructed case. A diagnostic system marks an image for possible malignancy, ranks the finding, and prepares a report. A physician later examines the image under ordinary time pressure and signs. The signature changes the status of the output. A computational recommendation has become an institutional act.

If the finding is wrong, consequence does not return to the inference that produced it. That execution has ended. Logs may preserve inputs, model version, and outputs, but a record cannot explain, reconsider, compensate, or be compelled. Nor does responsibility therefore fall naturally and exclusively upon the physician whose name appears last. The hospital selected the system. A vendor designed and updated it. Insurers and regulators defined acceptable use. Someone decided how much review the workflow would permit.

Liability collects across this surviving arrangement.

A liability sink is the person or institution against which an enforceable claim can finally be made. The term is useful only if it does not conceal the chain above it. In high-stakes domains, injured parties need an address that can receive notice, disclose evidence, submit to judgment, and fund a remedy. That address may be a professional, a firm, an insurer, a statutory fund, or several of them in sequence. What matters is not biological personhood. It is reachable standing joined to assets and an answerable office.

Authorization and liability are therefore related without being identical. A signature may authorize an act while an employer bears most of the loss. A platform may control deployment while a developer supplies the model. A regulator may recognize a use without indemnifying anyone harmed by it. Each arrangement decides who may act, who must explain, and whose balance sheet lies behind the explanation. Those are political-economic allocations, not technical details.

As cognitive work migrates into computational processes, compensation may move with the functions that remain scarce. Some of it may reward judgment. Some may pay for insurance, reputation, statutory standing, or the willingness to be reached by law. The empirical question is which function a credential still performs. A license can attest competence, grant authority, route liability, restrict entry, or do all four. Treating these as one service makes the resulting rent impossible to see.

The V/C heuristic helps locate pressure without deciding the outcome. Where error can be detected cheaply and correction arrives before harm, delegated work can be supervised with relatively little residual judgment. Where consequences unfold slowly, evidence is incomplete, or the decision itself changes the facts, verification may be expensive even when prediction is cheap. In those settings, a human signature can persist after meaningful review has become impossible. The signature remains legally potent while becoming epistemically thin.

This is the competence trap. It does not arise whenever a professional uses a model. It arises when routine delegation erodes the capacity to review, yet institutional design continues to treat the reviewer as if the old competence were being exercised in full. A person may then occupy the workflow mainly because insurers and courts require someone recognizable at its edge.

The danger is not merely unfairness to the signatory. Ceremonial review can make a weak system look governed. A physician, engineer, loan officer, or public official appears to stand between a machine and the world, although the time, information, and independent capacity required for judgment have already been removed. Accountability becomes theater when the named answerer cannot inspect the basis of the act or alter its consequence.

The trap is avoidable. Institutions can preserve independent competence through selective manual practice, adversarial audit, calibrated workloads, meaningful sampling, and the power to halt deployment. They can give reviewers access to provenance and uncertainty rather than only a recommendation. Most important, they can assign responsibility to the principal that designed and benefits from the system instead of concentrating it upon the last human hand to touch the file.

That assignment changes industrial structure. Principal liability favors organizations able to govern deployments and absorb losses, but it can also entrench large firms. Platform liability places pressure upon intermediaries with visibility across many deployments, while increasing barriers for smaller providers. Developer liability reaches upstream toward model design and testing, yet may advantage well-capitalized laboratories over open development. Professional liability preserves a familiar defendant, but may leave the party with the least control carrying the clearest name.

No allocation is neutral. Each joins control, proof, and loss differently.

Diffuse causation makes the choice harder. Harm may arise from a model's training, a deployer's missing safeguard, a platform's permissions, an institution's workload, and a user's reliance. Sometimes every component behaves as specified and the failure appears only in their composition. A regime that insists on finding one bad actor will misdescribe such cases. A regime that disperses responsibility everywhere may leave no one responsible enough to provide a remedy.

The constitutional requirement is narrower and more demanding: a surviving principal must remain reachable, evidentiary continuity must make the path of the act contestable, and an independent forum must be able to change the consequence. Computation does not abolish human responsibility. It changes the temporal and organizational distance between the act and the people who authorized the conditions under which it occurred.

Other Ways to Stand Behind an Act

Legal personality is one technology for making consequence attach. It is not the only imaginable one. A treasury can hold attachable assets. A bond can be forfeited. An insurance pool can compensate losses. A protocol can require collateral before granting access. Srinivasan's account of communities coordinating before they obtain territorial recognition offers one speculative template for institutions that assemble capacity before formal sovereignty.¹

None of these devices supplies judgment by itself. Collateral answers how a remedy might be funded after a specified event. It does not decide whether harm occurred, whether causation is established, whether a rule was legitimate, or whether exceptional circumstances justify release. Where the relevant facts are machine-observable and the governing rule is accepted in advance, automated enforcement can be powerful. Where facts or purposes are disputed, a bond without an answerable adjudicator is only precommitted force.

Adjudication need not always be supplied by the territorial state. Berman's account of medieval mercantile law describes merchant institutions operating amid overlapping legal orders rather than under one sovereign source.² Modern commercial arbitration likewise allows parties to choose specialized forums, although awards still depend upon legal systems for recognition and enforcement. The ICC's contemporary caseload demonstrates the scale of such private ordering, not its independence from public law.³

This history establishes possibility, not equivalence. A forum can be private and still answerable; a process can be decentralized and still reproduce capture. The tests are practical. Can a claimant enter? Can evidence be compelled? Can an adverse decision bind the powerful party? Can the remedy leave the protocol and reach the injured world?

Domains with sharply specified events may support more automated forms of bonding and dispute resolution. A shipment arrived or it did not. Funds moved or they did not. Even there, sensors fail, identities are contested, and force outside the ledger remains public or institutional. As interpretation, negligence, foreseeability, or proportionality enter the dispute, an independent human office becomes harder to replace because the controversy concerns which rule should govern the facts, not merely whether a bit changed state.

Scarcity Made by Law

Authorization can become a source of rent when cognitive capability diffuses faster than permission to use it. This does not make every license a guild device. Scarcity may protect patients, consumers, systems, or the public from risks that cannot be repaired after deployment. The same rule may perform genuine safety work and preserve an incumbent's income. Motive alone will not distinguish them.

The better inquiry follows function. Does the authorization requirement improve verification? Does it identify a principal with real control? Does it fund insurance or remedy? Does it preserve the competence needed for intervention? Could another arrangement achieve those purposes with broader access? A requirement that survives these questions may be constitutional infrastructure. One that cannot is a toll.

Insurance complicates the boundary. Underwriters can convert novel risk into priced exposure only when losses, controls, and responsible parties are sufficiently legible. Human review may help create that legibility, but a signature is not automatically a risk control. If the signer cannot inspect the system, change the workflow, or refuse the act, the signature merely gives the insurer a familiar defendant.

As cognition becomes easier to obtain, firms with licenses, balance sheets, proprietary data, physical access, and recognized standing may command a greater share of returns. Yet these assets are not one membrane, and their scarcity is not equally durable. Regulation can widen entry. Standards can make verification portable. Public insurance can socialize risks private markets will not bear. Institutional design can separate the valuable office of answerability from an incumbent's exclusive claim to perform the underlying work.

The decisive scarcity is not permission in the abstract. It is governed permission: authority joined to evidence, an answerer, and a forum able to provide relief. Computation increases the pressure upon that arrangement because the act may occur long before any institution is ready to hear the dispute.

Where liability collects will shape where returns collect. The two should not be confused. A society may choose to place responsibility upon those with control even when markets would place revenue elsewhere. If it does not, the cheapest architecture will often be the one that captures the benefit, distributes the act, and leaves the last recognizable human holding the loss.