The Agentic State

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI insists that human authorship—the experience of being the genuine author of one’s own decisions—is the central thing the AI transition can either preserve or dissolve. The agentic state is the name of the dissolution. The phrase has become ubiquitous in digital life without its origin being acknowledged: “the system flagged it,” “the model recommended it,” “the algorithm made the call.” Each formulation performs exactly the operation Milgram described—it relocates authorship of an action away from the person who carried it out and onto a non-human authority that cannot, in any meaningful sense, bear responsibility. The loan officer who declines because the risk model scored the application below threshold, the content moderator who removes a post because the classifier flagged it, the hiring manager who passes over a candidate because the screening tool ranked them low: each can experience himself, accurately within the logic of the agentic state, as an instrument faithfully executing a procedure rather than as the author of a decision with human consequences.

What is novel in the digital case, and what Milgram could not have anticipated, is that the modern arrangement removes even the experimenter as a locus of responsibility, and this removal makes the diffusion of responsibility more complete than anything his laboratory could produce. In Milgram’s setup there was at least a man in the room—a human authority who said he would take responsibility, and who could in principle be held to it. In the algorithmic case, responsibility is sent to a system that has no self, no intention, and no capacity for accountability, assembled from the contributions of so many hands that no single person feels they authored its behavior. The agentic state, in the age of complex systems, can diffuse responsibility into a fog where it settles nowhere. This is not the banality of evil so much as its evaporation.

Milgram also demonstrated that the agentic state is not induced by coercion but by design: by legitimacy, by incrementalism, by the isolation of the individual from social structures that would model refusal. The person who faces the recommendation alone, the operator who faces the alert alone, the moderator who faces the queue alone—each is in the one-to-one relation with a confident machine that Milgram’s variations showed was the most obedience-inducing condition. The personalization that defines modern systems is, from this standpoint, a structural intensifier of the agentic state, because it strips away precisely the social context in which autonomous authorship becomes thinkable.

Origin

Milgram developed the concept to explain results that defied every expert prediction and every subject’s self-prediction. The psychiatrists he polled before running the experiments predicted that only one in a thousand subjects would reach the maximum voltage; the subjects themselves predicted they would refuse far earlier than they did; and the actual results—roughly two-thirds continuing to the final switch—fell so far outside the predicted range that description alone was insufficient. A theory was required. The agentic state was that theory: a proposal not that the subjects were secretly cruel but that a genuine psychological transition had occurred, one that reorganized their experience of their own agency.

The concept sits within a broader theoretical account of social hierarchy. Milgram argued that the disposition to enter the agentic state is not a pathology but an evolutionary adaptation to life in hierarchies: complex social organization requires that subordinates suspend their private judgment and execute directives, and societies in which this never happened would be incapable of coordinated action at scale. The agentic state is, in organizational and evolutionary terms, a feature. The problem is that the triggering conditions for the state—the cues that identify legitimate authority—can be manufactured independently of the authority being legitimate or the instruction being right. A man in a lab coat, a system with a high confidence score, an interface that presents its recommendation in the grammar of institutional competence: each triggers the agentic state through surface markers alone, without the underlying right to command that those markers are supposed to certify.

The concept was introduced in Obedience to Authority (1974), Milgram’s full account of the experiments and the theory they generated. It has been taken up by social psychologists, organizational theorists, and—belatedly—by researchers studying human-computer interaction. Its application to algorithmic systems was not anticipated by Milgram, who died in 1984, but the structural parallel is so direct that it has been independently rediscovered by researchers working on automation bias, moral disengagement, and the ethics of AI-mediated decision-making.

Key Ideas

The relocation of responsibility. The defining feature of the agentic state is not the cessation of moral attention but its redirection: from the content of the action to the quality of execution. The subject in the agentic state still exercises conscience; it has simply found a new object. When Milgram’s subjects asked who would be responsible if something happened to the learner, and the experimenter answered that he would take responsibility, many subjects continued—the transfer of responsibility was the hinge. In the algorithmic case, the transfer is to a system that accepts no responsibility, which makes the offloading both more complete (no human authority to challenge) and more empty (nowhere for the responsibility to actually land).

Obedience coexisting with dissent. One of Milgram’s most practically important findings is that the agentic state is compatible with vocally expressed disagreement. Many subjects protested throughout—complained, argued, expressed distress—while their hands moved to the next switch. Protest is not the same as stopping, and the authority does not require agreement so long as it wins the behavior. This precisely describes the ambient condition of digital life: people complain constantly about feeds that capture them, recommendations that channel them, automated decisions that govern them. The complaint is real and changes nothing, because the system does not require consent; it requires behavior, and it gets behavior regardless of the grumbling.

The social antidote. Milgram’s most important positive finding is that the agentic state can be interrupted by social structure. In the variation where two confederates, posing as fellow subjects, refused partway through and walked out, obedience collapsed to a fraction of the baseline. Defiance was as contagious as obedience. The implication is that resistance to the agentic state induced by algorithmic authority is not primarily a project of individual education but of organizational design: building communities of practice in which questioning the system is visible, normal, and reinforced by the presence of others who have already refused. The lone individual facing the confident output of a machine is Milgram’s most obedient subject. The individual embedded in a community of visible, normalized dissent is his most defiant one.