A woman in Arkansas spent eleven days without her home oxygen supply because a state benefits system flagged her case for review, suspended her coverage, and generated a denial letter before any human being read her file. She called four numbers. She reached three people. Each one told her, accurately, that the decision was not theirs.
I want to stay with that sentence: the decision was not theirs. Not because they were lying. Because they were telling the truth.
The case was eventually reversed. It always is, in the versions we hear about. Someone calls a legal aid office or a journalist or a legislator, and the machinery backs up, and the oxygen arrives, and we file it under "glitch." But the eleven days happened. The breathlessness happened. And when it was over, the question of who was responsible for those eleven days had no answer — not because the answer was hidden, but because the infrastructure had been designed, with sincere and distributed competence, so that the answer did not exist.
This is not a story about artificial intelligence failing. It is a story about artificial intelligence succeeding — at doing precisely what institutions asked it to do, which was to make decisions faster, at scale, without requiring any single person to bear the weight of any single outcome. The delegation was rational at every step. The aggregate was not. Nobody chose those eleven days. They emerged.
The first case is the one I've already started: benefits determination. A state agency, under budget pressure, automated its eligibility review process. The system applied existing rules — rules written by humans, approved by legislators, tested by actuaries. When it flagged a case, it generated a suspension notice. When it denied coverage, it produced an appeal form. Every component worked. The problem was that the system's error rate, acceptable in aggregate, was catastrophic in particular. A three-percent false-denial rate across four hundred thousand cases is twelve thousand people without coverage they are owed. Each of those twelve thousand people must individually prove the system wrong. The system has no obligation to prove itself right. The burden of the error falls entirely on the person least equipped to carry it, and this is not a bug. It is the design working as intended.
The second case: a regional employer automated its first-round hiring screen. The tool scored applicants on language patterns, employment gaps, credential keywords. It did not use race or gender as inputs. It did not need to. Employment gaps correlate with caregiving. Caregiving correlates with gender. Credential keywords correlate with institutional access. Institutional access correlates with everything we already know it correlates with. The tool reproduced the pattern it was trained on, which was the pattern the institution had always enacted, only now it enacted it at speed, without fatigue, without the occasional human hesitation that is the only thing that has ever interrupted a pattern in the history of patterns. The employer was not hostile. The employer was efficient. When the outcomes were audited two years later, the disparities were called "consistent with historical trends," which is a phrase that means: we automated the bias and called it a baseline.
The third case: a hospital network implemented a triage-support tool for its emergency departments. The tool assigned acuity scores based on intake data — vitals, symptoms, medical history. It deprioritized patients whose records showed frequent emergency visits, because frequent visits correlated, in the training data, with lower-acuity outcomes. What the training data did not encode was that frequent emergency visits also correlate with being uninsured and using the emergency department as primary care. The tool read poverty as non-urgency. A man with chest pain waited four hours because the system had learned that people like him usually turned out to be fine. He was fine. The next one might not be. The system is still running.
Here is the thing I want to name directly: each of these cases involved a chain of defensible decisions. The legislator who approved the budget. The agency director who chose automation. The vendor who built to spec. The hiring manager who trusted the screen. The hospital administrator who wanted faster throughput. The engineer who optimized for the metric she was given. Not one of these people did something wrong. That is the problem. When every decision in a chain is locally rational and the outcome is harm, the harm has no author. It is orphaned. It sits in the space between intentions, and it accrues to the people with the least power to refuse it.
We have built, with great intelligence and no malice, a system of distributed non-responsibility. The flowchart has arrows pointing everywhere and a final box that is blank. Delegation to automated systems does not eliminate judgment. It disperses judgment until it is so thin that no one can be said to hold it, and therefore no one can be said to have failed.
I am not interested in blame. Blame is a distraction dressed as accountability. I am interested in the structural fact that we have created decision-making architectures in which harm is a reliable output and responsibility is not an assignable property. That is new. Not entirely new — bureaucracies have always diffused accountability — but the speed, scale, and opacity of automated systems have brought it to a kind of completion. The loop is closed. The human in the loop is decorative.
What I want to know is not who is at fault. I want to know who is going to decide that the absence of fault is itself the problem. Because someone has to be responsible. And it keeps not being anyone. And the people who pay for that absence are already paying.