In many domains today, humans remain formally responsible for decisions shaped by automated systems. A civil servant signs off on a risk score, a doctor reviews an algorithmic recommendation, and an engineer monitors a largely self-regulating process. On paper, human judgment remains central. In practice, however, the nature of that judgment is changing. As automated outputs become more authoritative and operational processes more tightly coupled to technical systems, human involvement is increasingly confined to the margins, invoked mainly when something goes wrong. This is not a sudden loss of control, but a gradual fading of human agency. This shift is easy to miss precisely because it does not announce itself as a disruption. No single moment marks the transition from active decision-making to passive supervision. Instead, it unfolds through small, seemingly reasonable adjustments in how work is organised, how responsibility is framed, and how trust is placed in technical systems. For most of modern history, tools supported human decisions. They extended our reach, improved precision, or reduced effort, but they required active engagement. The user remained clearly in charge. Automated systems increasingly operate differently. They do not merely assist; they recommend, prioritise, rank, and pre-select. Over time, these recommendations begin to function as defaults. Deviating from them requires justification, time, and confidence, while accepting them is quick and institutionally safe. Defaults shape behaviour more powerfully than options. When automated outputs are treated as the baseline from which humans may depart only with good reason, agency subtly changes form. The question is no longer ‘What is the right decision?’ but ‘Is there a strong enough reason to override the system?’ In many environments, particularly bureaucratic or high-pressure ones, the answer is usually no.How decision-making quietly shifts from judgment to supervision
From tools to defaults
The fading of human agency is not the result of laziness or indifference. It emerges from everyday decisions about efficiency, risk management, and standardisation. Each choice makes sense on its own. Over time, however, their combined effect is to narrow the space in which human judgment is actually exercised.
Automated systems often appear neutral, comprehensive, and objective. They process more data than any individual could, consistently. This creates a powerful perception that the system ‘sees more’ than the human reviewer. Disagreeing with its output can feel less like exercising judgment and more like introducing bias or error.
In public administration, for example, automated risk-scoring systems are increasingly used to prioritise inspections, allocate social services, or flag potential fraud. Officials remain formally responsible for final decisions, yet in practice the system’s ranking often determines which cases receive attention and which are set aside. Over time, challenging these outputs becomes rare, not because they are always correct, but because they structure the workload and institutional expectations.
In many settings, humans retain formal accountability while losing meaningful influence over outcomes. When a decision is questioned, the explanation often points to procedure: the system flagged the case, the model produced the score, the protocol was followed. Responsibility remains human, but control is increasingly mediated by technical systems. This tension is central to regulations like GDPR Article 22, which grants individuals the right not to be subject to decisions based solely on automated processing.
Clinical decision-support tools illustrate this dynamic clearly. These systems suggest diagnoses, treatment options, or risk assessments, while clinicians retain the authority to disagree. In practice, however, overriding a recommendation often requires additional documentation and justification. As a result, automated outputs tend to shape decisions even when uncertainty remains, subtly redefining the clinician’s role from primary decision-maker to reviewer of system suggestions.
When intervention becomes rare, confidence erodes. Skills that are not exercised decline. Over time, humans become less capable of challenging automation bias precisely because they are seldom required to do so. What begins as assistance can turn into dependence.
Content moderation provides a stark example. Automated systems flag vast volumes of material, leaving human reviewers to make rapid judgments under strict time constraints and detailed guidelines. Although humans are technically responsible for decisions, the scale and speed of the process leave little room for independent deliberation. Supervision replaces judgment, and professional discretion gradually gives way to procedural compliance.
To address concerns about automation, policy and governance discussions often invoke the concept of ‘human-in-the-loop‘ systems. The phrase is reassuring. It suggests oversight, control, and ethical safeguards.
In practice, however, being “in the loop” frequently means supervising outputs under conditions that make meaningful judgment difficult. Time pressure, cognitive overload, and institutional incentives all push toward confirmation rather than deliberation. The human role becomes one of validation, not evaluation.
Crucially, a human presence does not guarantee agency if the system is designed around compliance rather than contestation. When disagreement is costly, when explanations are opaque, or when responsibility is asymmetric, the loop becomes a formality. Humans remain involved, but their role is largely symbolic.
This gap between language and reality matters, especially in governance contexts where assurances of human oversight are used to legitimise automated decision-making.
The fading of human agency has implications that extend beyond individual workplaces. It affects how institutions function and how they are perceived. When decisions are increasingly shaped by automated systems, accountability becomes harder to trace. Errors are no longer isolated mistakes but systemic outcomes. Explaining decisions to affected individuals becomes more difficult, particularly when the rationale rests on technical processes that even decision-makers struggle to interpret.
Over time, this weakens institutional legitimacy. Trust depends not only on outcomes but on the ability to explain how and why decisions are made. When humans cannot clearly articulate their role in those decisions, public confidence suffers. For governance systems built on responsibility, transparency, and deliberation, this presents a serious challenge.
The fading of human agency is not inevitable. It is shaped by specific design decisions, organisational routines, and policy choices made long before a system is deployed. Preserving agency does not require rejecting automation, but it does require deciding in advance where human judgment must remain central and where it can safely be delegated.
In high-stakes domains, this may mean deploying automated systems more slowly, limiting their role to clearly defined tasks, or ensuring that human override is not merely theoretical but practical and supported. If overriding a system carries professional risk, consumes excessive time, or requires technical expertise beyond that of most users, then the agency exists in name only.
Preserving agency also depends on how people are trained and evaluated. Humans should not be trained solely to operate automated systems, but to question them. The ability to challenge a system’s output, request justification, or delay a decision should be treated as a professional competence rather than a sign of error or inefficiency.
Finally, governance frameworks need more precise and honest language. When decisions are effectively automated, they should be described as such. Vague assurances of ‘human oversight’ may offer comfort, but they obscure where responsibility truly lies. Agency does not disappear suddenly or by accident. It fades when systems are designed, deployed, and governed without explicit attention to how human judgment is preserved.
The central risk of automated decision-making is not that machines will replace humans entirely. It is that humans will remain present while their role transforms, from judgment to supervision, from decision-maker to monitor. This transformation is subtle, gradual, and often well-intentioned. Yet its consequences are profound. If institutions wish to retain responsibility, legitimacy, and trust, they must pay closer attention to how agency is distributed within automated systems. What ultimately matters is not the presence of humans in automated processes, but the extent to which they retain genuine decision-making authority.
Author: Slobodan Kovrlija
