Authority has traditionally been associated with identifiable people and institutions. A judge issues a ruling, a manager approves a proposal, an official signs a document, a doctor recommends treatment, and a public administrator determines whether an application satisfies regulatory requirements. Even when these decisions are supported by complex organizations, authority remains relatively easy to locate because someone ultimately occupies the position from which the decision is made and can, at least in principle, be asked to explain it.
Automation complicates this familiar structure. Digital systems can now rank applicants, identify anomalies, calculate risk, prioritize cases, recommend actions, approve transactions, reject submissions, and generate decisions at a speed and scale that would be impossible through human judgment alone. In some situations, a person still formally authorizes the result. In others, the process can be completed almost entirely by software. Between these two situations lies an expanding institutional territory in which it becomes increasingly difficult to identify where human judgment ends and automated authority begins.
The central question is therefore not simply whether machines are becoming more capable. It is whether institutions remain clear about who possesses the authority to decide, who is responsible for the consequences, and who can intervene when the system produces an inappropriate result.
From Assistance to Decision
Automation often enters institutions modestly. A system may initially be introduced to search records, identify inconsistencies, calculate probabilities, summarize information, or recommend priorities. Its function appears limited because the human employee remains formally responsible for the final decision.
Over time, however, institutional practice can change. Employees become accustomed to the recommendations, managers reorganize workflows around automated outputs, and the system gradually acquires a reputation for consistency and accuracy. What began as assistance can eventually become the default basis for action. An employee may still retain the formal power to disagree, yet rejecting the system’s recommendation can require more explanation than accepting it.
This is one of the least visible ways in which automation transforms authority. Authority does not necessarily move from humans to machines through an explicit institutional decision. It can shift gradually through routine. The official structure may remain unchanged while practical decision-making increasingly depends on the system that organizes information, establishes priorities, and frames the available options.
The result is not necessarily a fully automated institution. It is something more subtle: an institution in which human decisions are increasingly made inside computationally structured environments.
The Authority of Recommendation
Recommendations appear less powerful than decisions because they leave room for disagreement. In principle, a human operator remains free to reject the result produced by a system. In practice, however, the influence of a recommendation depends on the institutional context surrounding it.
Consider a system that analyzes thousands of previous cases and generates a risk score. The employee receiving that score may technically possess complete discretion to disagree. Yet disagreement requires confidence. If the model has processed more information than any individual could review manually, rejecting its recommendation may appear difficult to justify. The employee may ask why personal judgment should override a system considered more consistent, or why responsibility should be accepted for an exception when following the automated recommendation appears institutionally safer.
Under these conditions, a recommendation gradually acquires authority without ever being formally declared authoritative. The system does not need legal personality, a formal office, or decision-making status. Its influence emerges from organizational dependence, institutional trust, and the practical cost of disagreement.
This is especially important in organizations that value consistency. Automation can reduce arbitrary differences between decision-makers and can help limit errors caused by fatigue, incomplete information, or personal preference. These are significant advantages. Yet the same mechanisms can also make exceptions more difficult to recognize and professional judgment more difficult to exercise.
Human authority can therefore remain visible while becoming increasingly constrained.
The Human in the Loop
A common response to concerns about automated decision-making is the principle of keeping a human in the loop. The idea is intuitively attractive because it preserves a point of human responsibility. Machines may analyze information and produce recommendations, but a person should remain responsible for the final decision.
The presence of a human, however, is meaningful only when that person retains the practical capacity to exercise judgment. A reviewer who simply confirms automated recommendations does not necessarily provide meaningful oversight. Nor does a reviewer who lacks sufficient information to understand why the system produced a particular result.
Meaningful human involvement requires several conditions. The person must have time to evaluate the recommendation, access to relevant contextual information, sufficient understanding of the system’s limitations, and institutional permission to disagree. Without these conditions, human review can become ceremonial. The system produces the result, the person confirms it, and responsibility remains officially human even though practical authority has already shifted elsewhere.
The question is therefore not merely whether a human remains somewhere inside the process. It is whether that person retains genuine capacity to influence the outcome.
This distinction becomes increasingly important as institutions seek to reassure the public that automated systems remain under human control. Formal approval is not the same as substantive oversight.
Authority Without Responsibility
Automation introduces a difficult institutional asymmetry because systems can exercise considerable influence without possessing responsibility. An algorithm cannot be professionally disciplined, a model cannot accept moral accountability, and software cannot explain why it believed a decision was fair.
Responsibility therefore remains with people and institutions, but identifying exactly where that responsibility lies can become increasingly difficult. A decision may depend on the employee who accepted the recommendation, the team that developed the model, the organization that purchased the technology, the vendor that supplied it, the regulation that established the criteria, and the historical data on which the system relies.
Authority becomes distributed across a network of actors and technical components.
This distribution does not eliminate responsibility, but it can make responsibility harder to locate. When decisions are produced through multiple layers of data, software, organizational procedures, and human interpretation, no single participant may feel fully responsible for the final outcome.
This is why automated decision-making should be understood as a socio-technical process rather than as an isolated interaction between a person and an algorithm. The important questions concern not only whether the model is accurate, but also how the institution allocates responsibility, how decisions can be challenged, and who possesses the authority to correct errors.
The Comfort of Technical Objectivity
Automation acquires institutional authority partly because numerical outputs often appear more objective than human judgment. A personal judgment can be questioned as subjective, inconsistent, or biased, while a calculated score appears precise and impersonal.
This perception gives automated systems considerable legitimacy.
Yet no automated system is created without prior choices. Someone determines which variables should be included, which data should be used, which objectives should be optimized, what threshold should trigger action, and which types of error are considered acceptable. The mathematics may be rigorous, but the architecture of the system still reflects institutional assumptions.
This distinction matters because automation can reduce some forms of bias while reproducing others. Historical datasets may contain patterns produced by earlier inequalities or administrative practices. If these patterns are incorporated into predictive systems without critical examination, the resulting model may reproduce them while appearing neutral.
Judgment has not disappeared. It has moved.
Instead of occurring only at the moment of decision, judgment can become embedded in data selection, model design, classification systems, and optimization objectives. Once embedded, these choices can be reproduced repeatedly and at enormous scale.
Technical objectivity is therefore never purely technical.
It is institutional judgment translated into computational form.
Scale Changes Authority
Automation becomes especially significant when deployed at scale. A human decision-maker may influence dozens or hundreds of cases, while a digital system can affect millions.
This scale creates both opportunity and risk. A well-designed automated system can apply consistent standards across large populations and process enormous quantities of information efficiently. At the same time, a flawed classification or inappropriate assumption can also be reproduced with extraordinary speed.
Scale transforms a technical tool into infrastructure.
A risk score used occasionally is one thing. A risk score embedded across an entire administrative system becomes part of the way an institution encounters society. It influences which cases receive attention, which transactions are questioned, which applications move forward, and which individuals are treated as exceptions.
At this point, the distinction between technical design and governance becomes difficult to maintain.
The system is no longer merely supporting decisions. It is structuring the conditions under which decisions occur.
This is why the design of large-scale automated systems should be treated as a question of institutional authority rather than simply information technology.
Speed and the Compression of Judgment
Automation also changes the temporal structure of decision-making. Human decisions require time because information must be gathered, circumstances considered, and exceptions interpreted. Automated systems can compress these processes dramatically, producing thousands of assessments in seconds.
In many areas, this is enormously valuable. Fraud detection, emergency monitoring, network security, and operational systems may require speed that human processes cannot provide.
Yet the increasing availability of rapid decisions can also create new institutional expectations. If a system can produce an answer immediately, human deliberation may begin to appear inefficient. Decision-makers may be asked why additional review is necessary when a recommendation is already available.
This can gradually alter organizational culture. Speed becomes associated with competence, while hesitation can be interpreted as weakness.
The problem is not that institutions act quickly. The problem arises when the speed of computation becomes the standard by which every form of judgment is evaluated.
Some decisions involve ambiguity. Some require contextual interpretation. Others involve competing values or circumstances that cannot be reduced easily to structured information. In such situations, deliberation is not inefficiency. It is part of responsible decision-making.
Automation can accelerate information processing.
It cannot determine how much reflection a decision deserves.
The Exception Problem
Automated systems are particularly effective when patterns are stable and rules can be applied consistently. Institutions, however, do not encounter only ordinary cases.
They also encounter exceptions.
A transaction may resemble suspicious activity while being legitimate. An applicant may not fit standard categories because their circumstances are unusual. A record may appear inconsistent because the underlying situation itself is complex.
Human institutions have historically relied on discretion partly because no rule can anticipate every possible circumstance. Automation creates a tension around this discretion. Too much human discretion can generate inconsistency and unfairness, while too little can produce rigidity.
The challenge is therefore not to choose between automation and discretion, but to design systems in which both can coexist.
A mature automated institution should be capable not only of recognizing patterns but also of identifying situations in which ordinary patterns may be inappropriate. This is an important form of institutional intelligence.
The ability to recognize exceptions is not evidence that a system has failed.
Sometimes it is evidence that the institution still understands the limits of its own rules.
When No One Fully Understands the System
Modern automated systems can become sufficiently complex that no single person understands the entire decision process. A machine-learning model may depend on large numbers of variables, historical datasets, external systems, organizational rules, vendor components, and technical infrastructure developed over many years.
The final output may be simple while the process producing it is not.
This creates a serious challenge for institutional authority because legitimate authority has traditionally been connected to the ability to provide reasons.
A person affected by an important decision may reasonably ask why the decision occurred. If the answer becomes simply that the system produced a particular result, explanation weakens.
Technical complexity cannot become an excuse for institutional opacity.
Institutions do not need every citizen or user to understand the mathematical structure of a model. But they should remain capable of explaining the institutional logic of decisions. They should be able to clarify what information mattered, what rules were applied, what a particular classification means, how the result can be challenged, and who possesses the authority to correct it.
Automation may make explanation more difficult.
It does not eliminate the obligation to explain.
Governing the Systems That Govern
The expansion of automation creates another challenge that receives less attention: whether institutions themselves possess the capacity to govern the systems they deploy.
An organization may purchase sophisticated technology without developing equivalent expertise internally. Employees may know how to operate the interface but have limited understanding of how the model works. Senior officials may understand the intended purpose of the system while remaining unfamiliar with the assumptions embedded in the underlying data.
This creates a potential imbalance between formal authority and technical understanding.
The institution remains legally responsible for decisions, yet critical knowledge may reside with external vendors or specialist teams.
Such dependence does not automatically make automation inappropriate. Complex organizations have always relied on specialized expertise. But institutions need enough internal capacity to question the systems on which they depend.
Otherwise, they risk becoming formally responsible for processes they cannot adequately evaluate.
The challenge of automation is therefore not only whether artificial intelligence can make decisions.
It is whether institutions can remain capable of governing artificial intelligence.
Authority without understanding is fragile.
Accountability Cannot Be Automated Away
Automation can redistribute tasks, but it cannot remove accountability.
Institutions remain responsible for the systems they choose to deploy, particularly when automated processes affect rights, opportunities, access to services, employment, financial decisions, or other significant aspects of life.
The phrase “the system decided” cannot become an acceptable endpoint of explanation because systems do not enter institutions independently. Someone authorized their use, defined their role, selected the consequences attached to their outputs, and established the procedures through which those outputs become decisions.
Accountability therefore needs to follow authority even when authority is technically distributed.
This may require audit mechanisms, decision logs, appeal procedures, independent evaluation, and clearer institutional rules concerning automated systems.
Most importantly, organizations need to preserve the distinction between delegating a task and delegating responsibility.
Automation can move tasks away from individual employees.
Responsibility must remain institutionally identifiable.
Preserving Human Judgment
Debates about automation are often framed as competition between humans and machines, but this framing is increasingly unhelpful. Machines are superior at some tasks, while humans remain necessary for others.
The more important institutional question is where each form of capability belongs.
Automation is especially powerful when work involves repetitive processing, large volumes of structured information, pattern recognition, or consistent application of clearly defined rules. Human judgment becomes particularly important when decisions involve ambiguity, ethical consequences, competing values, unusual circumstances, or contextual interpretation.
The boundary between these domains will continue to change as technology develops.
What should remain stable is the principle that technical possibility does not automatically establish institutional desirability.
A decision can be automated without necessarily becoming better.
A process can become faster while losing accountability.
A system can become more consistent while becoming less capable of recognizing legitimate exceptions.
Institutions therefore need to evaluate automation not only through efficiency, but through purpose.
The relevant question is not whether a task can be automated.
It is whether automation strengthens the institution’s capacity to fulfil its responsibilities.
Authority After Automation
Future institutions will almost certainly contain more automation than institutions do today. Artificial intelligence will support analysis, documentation, monitoring, risk detection, service delivery, and decision-making. Some processes will become increasingly automatic, while others will continue to depend on human interpretation.
The governance challenge will be ensuring that greater automated capability does not create greater ambiguity about authority.
Every significant automated process needs boundaries. Every recommendation needs a defined relationship with human judgment. Every consequential decision should remain contestable when circumstances justify reconsideration.
Authority may become distributed across people, software, data, rules, and organizations, but accountability cannot simply dissolve into that network.
This requires a broader understanding of authority. Authority should not be defined only by the person who presses the final button. It should be understood through the entire architecture that produces the decision: who designed the system, who selected the data, who established the criteria, who determined the consequences, who can question the result, and who remains responsible when something goes wrong.
Automation changes the mechanics of decision-making, but it does not eliminate judgment, responsibility, or institutional obligation.
The most capable automated institution will therefore not be the one in which humans intervene the least.
It will be the one that still knows when human judgment matters most.
References
Andrews, L. (2019). Algorithms, regulation, and governance readiness. In K. Yeung & M. Lodge (Eds.), Algorithmic Regulation (pp. 203–223). Oxford University Press.
Criado, N., & Such, J. M. (2019). Digital discrimination. In K. Yeung & M. Lodge (Eds.), Algorithmic Regulation (pp. 82–97). Oxford University Press.
Scantamburlo, T., Charlesworth, A., & Cristianini, N. (2019). Machine decisions and human consequences. In K. Yeung & M. Lodge (Eds.), Algorithmic Regulation (pp. 49–81). Oxford University Press.
Yeung, K. (2019). Why worry about decision-making by machine? In K. Yeung & M. Lodge (Eds.), Algorithmic Regulation (pp. 21–48). Oxford University Press.
Yeung, K., & Lodge, M. (Eds.). (2019). Algorithmic Regulation. Oxford University Press.
