Use Is Not Evenly Distributed

Discussions of AI adoption often speak in generalities: organizations are using AI, people are benefiting from AI, society is adapting. This language hides an important asymmetry. AI systems are not used equally across roles, and their consequences are not borne equally across populations. The people who choose to deploy and rely on these systems are often insulated from their failures, while the people most affected by errors have little say in their design or use.

Understanding AI usage requires tracking where power sits in the workflow, not just where the system appears.

Upward Efficiency, Downward Risk

In many institutions, AI tools are adopted by those with managerial or administrative authority to increase efficiency: summarizing reports, drafting communications, triaging cases, generating evaluations. These uses save time at the top of the hierarchy. But the outputs often act on those lower in the hierarchy—workers, applicants, customers, students, patients.

When an AI-generated summary misrepresents a worker’s performance, the worker bears the consequence. When a classification system routes a case incorrectly, the applicant experiences delay or denial. When automated language deflects a complaint, the customer absorbs the frustration.

Efficiency flows upward. Risk flows downward.

The Error Budget Is Unevenly Assigned

All systems have an error rate. The critical question is not whether errors occur, but who absorbs them. In AI-mediated workflows, errors are frequently treated as acceptable noise if they do not disrupt institutional goals. A small percentage of incorrect outcomes can be tolerated as long as aggregate metrics improve.

But for individuals, errors are not statistical. They are personal. A mistaken label, an incorrect summary, or an automated denial can have lasting consequences. The system’s success is measured in averages; harm is experienced one case at a time.

This mismatch allows institutions to declare systems effective while individuals experience them as arbitrary or unjust.

Disempowered Subjects, Not Empowered Users

Much of the rhetoric around AI emphasizes empowerment: tools that help people work better, learn faster, decide more wisely. But many of the people most affected by AI systems are not users at all. They are subjects.

Subjects do not interact with the system directly. They encounter it through outcomes: a decision letter, a performance review, a risk score, a delayed response. They cannot inspect prompts, question assumptions, or provide feedback. Their only interface is the result.

This distinction matters. Empowered users can adapt, compensate, or opt out. Subjects cannot.

Blame Flows Toward the Least Powerful

When AI-mediated decisions go wrong, responsibility is rarely assigned upward. Executives cite systems. Managers cite tools. Frontline workers are left to explain outcomes they did not meaningfully control. Individuals affected by decisions are often told that “the system” produced the result, as if that absolves anyone of accountability.

This pattern is familiar from earlier forms of bureaucratic automation. AI intensifies it by adding opacity and speed. The more complex the system, the easier it becomes to deflect responsibility.

Blame flows downhill because power does.

The Illusion of Neutrality

Because AI systems often present outputs in neutral, procedural language, their decisions can appear objective even when they encode institutional priorities. A denial letter generated by a model feels impersonal, but impersonality does not equal fairness. It often means that human discretion—along with human empathy—has been removed.

For those on the receiving end, this neutrality can feel dehumanizing. There is no one to appeal to, no explanation to contest, no clear source of authority to challenge.

Neutral language becomes a shield behind which power hides.

Why This Pattern Persists

This asymmetry persists because it aligns with institutional incentives. Systems that reduce labor costs and increase throughput are rewarded. Harms that are diffuse, individualized, or difficult to aggregate are discounted. There is little pressure to redesign workflows around those who bear the costs.

As long as the people experiencing harm lack voice or leverage, their experiences remain anecdotal rather than actionable.

Looking Ahead

Understanding who uses AI—and who pays for its mistakes—clarifies why accountability debates so often stall. Authority is exercised upward, harm is absorbed downward, and responsibility is diffused in between.

The next essay will examine how organizations respond when these harms become visible—and why the language of “oversight” often masks the absence of meaningful control.