Ambiguity of the Moral Agent: Who Is the Agent in AI Decisions?

Consider the sudden maneuver an autonomous vehicle makes at the moment of collision. Or the decision of an algorithm that rejects a credit application. Or a system that prioritizes targets on a battlefield. The common thread in these scenarios is this: there is a "decision," there is a "result," but there is no "agent" in the traditional sense. The entity making the decision cannot feel regret, cannot be held accountable, and cannot be punished.

The human mind has evolved over thousands of years to link responsibility to an agent. When something goes wrong, our first question is almost automatic: "Who did this?" Artificial intelligence systems, particularly those with high autonomy in decision-making, are creating the first large-scale domain where this question remains unanswered. And this gap produces psychological consequences far deeper than commonly assumed.

In social psychology there is a concept called the "just-world belief" — the deep-seated need to believe that events have causes and agents, and that bad things happen not by chance but as the result of specific actions. This belief makes the world feel predictable and controllable. When no agent can be identified, this need goes unmet, and the mind springs into action to fill the void.

^{Yapay Zeka Kararlarında Fail Kim? (Yapay Zekâ ile Oluşturulmuştur.)}

When no agent can be found, something interesting happens: responsibility shifts to the nearest human element — the engineer who designed the system, the operator who used it, or even the person who suffered harm. In other words, because the psychological state of "no one is guilty" is unsustainable, the mind designates the most plausible human as a scapegoat — regardless of whether it is fair or not.

One of the classic findings in social psychology is "diffusion of responsibility" — the phenomenon where, in a group setting, everyone feels someone else should take action, so no one does. In artificial intelligence systems, this mechanism becomes even more complex because the group is no longer composed solely of humans: there is the team that prepared the dataset, the engineers who designed the algorithm, the company that trained the model, the institution that deployed the system, the manager who approved the decision, and ultimately the system itself.

When responsibility is distributed among so many actors, each party psychologically minimizes their own share. "I only prepared the data," "I only wrote the code," "I only approved it" — each statement is reasonable on its own, but collectively no one fully owns the full picture. This is not due to malice, but to the human cognitive tendency to simplify complex chains of causality.

Another observed tendency among humans working with autonomous systems is "responsibility shift": when they delegate a decision to an algorithm, they also feel they have delegated its moral weight. This resembles the dynamic seen in classic obedience studies — when an authority gives an order, individuals place themselves in the role of "just following instructions," gaining moral relief in the process.

The difference is that the authority figure is no longer a human, but a system assumed to be "objective" and "neutral." This assumption makes escaping responsibility psychologically easier, because when obeying a human authority, at least some moral tension may remain; when obeying algorithmic authority, that tension often vanishes entirely. The phrase "the system said so" feels less open to questioning than "my boss said so," because the system is assumed to have no intention, no interest, no bias — an assumption that is often false.

There is another side to this coin. A person harmed by an AI decision may be left unsure where to direct their anger. Anger is a functional emotion in human psychology — it motivates us to act against injustice. But when there is no concrete target for that anger, the emotion either turns inward (helplessness, self-blame) or spreads toward vague targets — abstract concepts like "the system," "the company," or "technology."

This situation leads to an unfulfilled sense of justice. Court proceedings, apologies, accountability rituals — most of the mechanisms human societies have developed to heal after harm — assume the existence of a clearly identifiable agent. When the agent becomes ambiguous, the victim’s psychological need for closure remains unmet.

Systems approaching AGI levels will amplify this accountability gap as their decision-making autonomy increases. The core issue here is not technical — it is that the human mind’s way of moral accounting is unprepared for a world where the concept of an agent has become blurred. Legal systems, institutional structures, and even our everyday moral intuitions still rely on the triangle of "one person, one action, one outcome."

The solution likely lies in constructing a new framework of accountability — one that is distributed rather than individual, yet still holds parties answerable. But until such a framework is established, both decision-makers and those affected by these decisions must accept that they are operating in a domain of responsibility that, by definition, is ownerless. And this is an unsettling, unfamiliar uncertainty for the human psyche.

The only thing individuals can do in the face of this uncertainty is perhaps resist the urge to compress responsibility into a single point and learn to see the process as a whole.