What Remains Human

V = Eval_A(a, F, t)

The Issuer            authors A, the authority          what may the machine pursue?
The Escalation Target decides what A reserves           when must the machine defer?
The Sovereign         answers for A and every V         who stands behind the action?

The agent is not an input. Every input that confers authority is held by people.

1. The question

Every gain in machine capability reopens the same question: when software can execute more of the work, what are people for? Two answers dominate, and both fail.

The first keeps humans in the loop as supervisors who review what the machine does. It fails on arithmetic and on psychology. On arithmetic, because review that keeps pace with machine execution is not review; at sufficient volume the supervisor either becomes the bottleneck autonomy was meant to remove or stops reading. On psychology, because the evidence on automation is consistent: people who monitor automated systems tend to defer to them [1], and the more reliable the automation, the less practised and less alert the human becomes at exactly the moment intervention is needed. Lisanne Bainbridge called this the irony of automation: the more we automate, the more crucial and the more difficult the remaining human contribution becomes [2].

The second answer casts humans as competitors who must stay more capable than the machine to stay relevant. That race has a predictable end, and an argument for human standing that depends on winning it is an argument with an expiry date.

The third answer starts from a different observation. What a machine may do was never a function of what it can do. It is a function of authority, and authority has a structure that capability does not touch. Ask what that structure requires and three functions appear that no amount of intelligence performs on its own behalf. This paper names them, specifies the mechanism that keeps each in human hands at machine speed, and states plainly what the mechanism cannot do.

2. Capability is not authority

The premise is old. Max Weber distinguished power, the ability to get one's way, from legitimate authority, the probability that a command will be obeyed because those subject to it accept the right of the one who gives it [3]. In the modern state that acceptance rests on legality: on rules, and on the standing of those who make them, rather than on the competence of whoever happens to apply them.

H. L. A. Hart sharpened the point against John Austin's picture of law as the commands of a sovereign whom others habitually obey [4][5]. Habit of obedience, Hart argued, cannot explain why a newly crowned legislator's first statute is law, or why a rule binds the official who enforces it. What does the explaining is a set of rules about rules: rules that confer the power to make law, rules that confer the power to decide disputed cases, and a rule of recognition by which officials identify what counts as law at all. Authority, on this account, is conferred by rules and held by persons. It is not a property of whoever is most capable of acting.

Machine capability changes nothing in that structure. A model that can draft a credit policy better than any analyst has not thereby been given the power to adopt one. A system that can evaluate a loan faster than any officer has not thereby acquired standing to approve it. The Computable Authority equation makes the separation explicit: in V = Eval_A(a, F, t) the agent that presents the action is not an input. The authority A, the action a, the proven facts F and the time t decide the verdict. Who asks does not.

Two limits must be stated with the premise, because the argument fails without them. First, computability does not create legitimacy. A rule can be identifiable, reproducible and independently verified, and still be unjust or issued by an authority that should not hold it; a perfectly verifiable unjust rule is verifiable injustice. Computable Authority makes legitimacy inspectable and enforceable. People and institutions confer it. Second, the argument does not depend on any forecast about superintelligence. It holds at today's level of capability and at any higher one, because it rests on the structure of authority, not on the limits of machines. The more capable the executor, the more the boundary around it matters.

Capability is not authority. Intelligence is not legitimacy. Execution is not authorization.

3. Three functions that confer rather than exercise

Strip any working relation of authority to its essentials, whether a mandate from a board, a statute and the officials who apply it, or a delegation from a bank to an agent, and the same three functions remain. Someone grants the authority. Someone decides the cases the grant deliberately leaves open. Someone answers for the grant and for what is done under it. The parallel with Hart's secondary rules is close: rules of change confer the power to make rules, rules of adjudication confer the power to decide cases, and recognition identifies which rules count [4]. Each of the three functions confers authority rather than exercising it, and that is why none can be computed into existence. A machine can apply a rule it was given. It cannot give itself the rule, reserve its own exceptions, or be the party the consequences fall on.

Computable Authority names the functions as roles, assigns each a question and a primitive, and specifies what remains institutional rather than technical.

3.1 The Issuer: what may the machine pursue?

The Issuer defines the objectives, constraints and boundaries of permissible action. It does not specify every step a machine will take; it establishes the authority within which the machine may take them. In the equation, the Issuer authors A.

The primitive is the sealed authority artifact, and the architectural decision that makes it possible is the separation of compilation from execution set out in the position paper submitted to the OECD Consultation on the Digital Provision of Law (R-07). Interpretation of a policy, including AI-assisted interpretation, happens once, at compilation, where people can review it. What runs is a sealed artifact evaluated deterministically. The boundary lives outside the application, in an object an outside party can inspect, rather than inside code where it is invisible.

Three properties keep this role human. The artifact is sealed by a key that chains to a root the relying party chose, so its issuer is established rather than claimed (NOMOS-SPEC-007). The key that issues rules cannot also state the facts those rules read: rule-issuing keys and fact-stating keys are separated by construction, so authority over the boundary and authority over the evidence cannot collapse into one holder. And only the version the Issuer designates as in force may authorise a new action (NOMOS-SPEC-008 §5.4), so an Issuer's decision to replace its rules takes effect without depending on anyone else's cooperation.

The consequence is the strongest claim in this paper's first role: a machine cannot legitimately author its own mandate. It may draft one, and drafting is where machines increasingly excel. But a draft becomes authority only when an Issuer adopts and seals it, and an agent can be given standing to issue only if an institution certifies its key, in which case the authority it issues remains the institution's.

What remains institutional is the content of the mandate: what the machine should pursue, which risks are acceptable, which values outrank which. Those are judgements an institution makes and owns. The mechanism makes them explicit, versioned and binding. It does not make them for anyone.

3.2 The Escalation Target: when must the machine defer?

No rule set anticipates every situation, and some decisions should never be automated even when they could be. The Escalation Target decides the cases the authority deliberately reserves for human judgement. In the equation, the Escalation Target receives the verdict ESCALATED and answers it.

Precision matters here, because three different situations are often confused. ESCALATED means the rules themselves reserve this decision to a designated role: a loan above a threshold requires two senior approvers, a child-protection intervention requires a social worker's judgement. INCOMPLETE is different: a fact the rules need is unproven, and the route is to evidence, not to people. A human approval must never substitute for a missing fact, or discretion becomes a way to skip evidence. And a situation the rules did not anticipate at all falls to the authority's sealed default outcome, which the specification recommends setting to escalate or deny for any consequential action. That default is the honest answer to the observation that no policy can foresee everything: the unforeseen is routed, in advance, to the people who should decide it.

The primitive is escalation as a first-class verdict of the computation, not a failure mode of it. In conventional software an exception that reaches a person usually means something broke. In Computable Authority it is the correct execution of a limit on machine authority, authored in advance by the Issuer. The human decision does not leave the system either. It returns as a consent signed over the exact action, the exact rules and the exact facts, counted only from distinct keys certified for the required role, and void if anything in the action changes (NOMOS-SPEC-008 §4). Human judgement re-enters the computation as a verified input.

This role is where the irony of automation must be designed against rather than discovered. An Escalation Target asked to approve a stream of routine cases will defer, as the evidence predicts [1][2]. The answer is structural: escalation should arrive rarely, only for cases the rules reserve, with the facts the rules read already proven and the open question stated, so that the human decides something real rather than ratifying something already decided. The literature on meaningful human control makes the same demand in other words: control is meaningful only when the human can track the reasons behind the system's behaviour and the outcome can be traced back to a human decision [6]. The European Union's AI Act now requires high-risk systems to be designed for effective human oversight on the same reasoning [7].

The role also makes a precise definition of autonomy possible. Autonomy is not the ability to act without asking. It is the ability to act independently within a defined authority boundary, and recognising the edge of that boundary is part of it. A machine that escalates correctly is more autonomous, not less, than one that acts everywhere.

3.3 The Sovereign: who stands behind the action?

The word needs care. It is not used in Austin's sense of a commander whom others habitually obey; Hart's critique of that picture is part of this paper's premise. The Sovereign here is the accountable holder: the institution, and the named people within it, who answer for the authority granted and for every action taken under it. In the equation, the Sovereign answers for A and for every verdict V computed under it.

The role exists because of a gap that the ethics literature identified before agents made it urgent. Andreas Matthias called it the responsibility gap: when a learning system acts in ways its operators could not predict or control, traditional grounds for holding anyone responsible weaken [8]. Dennis Thompson described the older form in public administration as the problem of many hands: when many people contribute to an outcome, it becomes difficult to identify anyone responsible for it [9]. Autonomous execution compounds both. A piece of code cannot stand before a regulatory board, be sanctioned, be ruined, or carry moral responsibility. If no one can be connected to what it did, no one answers for it.

The primitive is an unbroken, verifiable chain from every machine action to the authority that permitted it and the people who granted that authority. The seal names the issuer. The certificate chain ties the issuing key to an institution's root (NOMOS-SPEC-007). Each consent is signed by a named approver's key. Each committed action leaves a receipt that records which sealed version governed it, the facts it was evaluated on and the time, so that any third party can reconstruct the verdict years later. A regulator or auditor can co-sign the exact version it reviewed (NOMOS-SPEC-004). And the Sovereign keeps the power to withdraw what it granted: a revoked authority stops every action that would rely on it (NOMOS-SPEC-006).

The chain also enforces what Lon Fuller called congruence between official action and declared rule, the requirement he regarded as the most complex of the conditions that make a legal system a legal system at all [10]. In a document world, congruence is aspired to and audited after the fact. When the system that would make a change computes the verdict itself, against the sealed version in force, an action incongruent with the declared rule does not commit.

What remains institutional is accountability itself. The chain makes the relationship between an action and its authority explicit and verifiable. Legal and institutional arrangements decide what that relationship means for liability. Authority can be delegated; accountability for having delegated it cannot.

4. The roles and the equation

Two further inputs complete the picture. The facts F come from witnesses the Issuer chose to certify, each for exactly the claims it may state, so even the evidence a verdict rests on traces back to a human decision about whom to trust. And the agent appears nowhere in the table. That is the architectural form of this paper's thesis: every input that confers authority is held by people, and the machine is the party that acts within it.

5. A shift in what is scarce

Each transition moved humans away from the function that became cheap and toward the one that stayed scarce. When intelligence becomes abundant, legitimacy does not, because it is not produced by computation at all. It is conferred by people on people and institutions, and held by them. The agentic transition does not make humans less important. It makes the specific thing they contribute impossible to substitute.

6. Objections

Humans will rubber-stamp. They will, if escalation is designed as review of everything. The evidence on automation bias is the reason the Escalation Target's primitive is built the way it is: escalations are reserved by the rules, carry the proven facts and the open question, and bind the human's answer to the exact action. A design that escalates everything recreates the problem this paper began with. A design that escalates nothing has removed the role.

The roles will be automated too. Drafting will be, and should be: models already propose rule structures, flag ambiguity and summarise escalated cases. What cannot be automated is the standing to adopt, to decide and to answer. An institution can certify a machine's key to issue within a scope, but the authority it issues is still the institution's, and so is the accountability. Delegation moves the work. It does not move the responsibility.

This entrenches whoever holds the keys. It makes their power visible, which is different from making it legitimate. Who should hold the roots that authority chains to is the open institutional question R-09 addresses, and revocation lets authority be withdrawn as well as granted. A system in which every exercise of machine authority traces to named holders is a system in which concentrated power can at least be seen, contested and taken back.

7. What the roles do not require

None of the three roles requires humans to be more intelligent than the machine. The Issuer need not out-compute the system it bounds; it must decide what the system may pursue. The Escalation Target need not out-reason the model; it must decide the cases the authority reserves. The Sovereign need not understand every computation; it must answer for the authority under which each one ran. What the roles require is that humans hold the authority, and that the infrastructure make holding it real at the speed machines now act.

Capability is not authority. Intelligence is not legitimacy. Execution is not authorization.

References

• [1] L. J. Skitka, K. L. Mosier & M. Burdick, "Does Automation Bias Decision-Making?", International Journal of Human-Computer Studies 51, no. 5 (1999): 991–1006.
• [2] L. Bainbridge, "Ironies of Automation", Automatica 19, no. 6 (1983): 775–779.
• [3] M. Weber, Economy and Society, ed. G. Roth & C. Wittich (University of California Press, 1978 [1922]).
• [4] H. L. A. Hart, The Concept of Law (Clarendon Press, 1961).
• [5] J. Austin, The Province of Jurisprudence Determined (1832).
• [6] F. Santoni de Sio & J. van den Hoven, "Meaningful Human Control over Autonomous Systems: A Philosophical Account", Frontiers in Robotics and AI 5 (2018): 15.
• [7] Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14 — human oversight.
• [8] A. Matthias, "The Responsibility Gap: Ascribing Responsibility for the Actions of Learning Automata", Ethics and Information Technology 6, no. 3 (2004): 175–183.
• [9] D. F. Thompson, "Moral Responsibility of Public Officials: The Problem of Many Hands", American Political Science Review 74, no. 4 (1980): 905–916.
• [10] L. L. Fuller, The Morality of Law (Yale University Press, 1964).
• NOMOS-SPEC-001, 004, 006, 007 and 008 (0.4.0): github.com/nomos-spec/spec.