Who Is Responsible When Nobody Decided?

The letter arrives with a number.

A government agency says you owe money. The amount may be large enough to change the month, the year, or the family budget. You call to ask how the debt was calculated. The person who answers did not calculate it. The supervisor did not calculate it. The software vendor did not decide that you personally owed anything. The program followed rules, combined records, and produced an outcome.

So you ask the most natural question in the world: Who decided? The unsettling answer is that nobody may have decided in the ordinary sense of the word.

There may have been no clerk who studied your file, no manager who weighed the evidence, no official who paused over your circumstances and chose one result over another. Instead, a chain of earlier decisions created a system that could generate your result automatically. People chose the goal. People chose the data. People chose the formula. People approved the workflow. People decided when human review was unnecessary. Then the system acted.

You receive a definite result, but there may be no single person who chose it. That makes accountability harder to follow. We are used to asking who signed the order, denied the claim, or approved the loan. With an automated system, we may need to examine several earlier choices to find out who must answer for what happened to us.

That separation creates a temptation: if nobody made the specific decision, perhaps nobody owns it. Principle 1 of the Delegation Doctrine rejects that escape. Accountability may be shared and distributed, but it cannot disappear merely because the final action emerged from a system rather than from a person sitting at a desk.

Decision by Architecture

Some decisions are made when a person acts. Others are built into the rules a system follows. That second kind is a decision by architecture: a decision shaped by how the system was put together.

In the familiar version, a manager approves a purchase or a lender says yes or no. We can usually identify the person who made the choice.

With an automated system, many important choices happen earlier. People decide its goal, the information it uses, the score that triggers an action, and what happens when nobody intervenes. They decide which cases need human review and how someone can appeal.

Those choices can determine the result long before a particular person asks for help. The system may then deliver an answer without anyone making a fresh judgment about that person.

The absence of a final human choice does not mean the outcome was undesigned. It means the design itself became the decision mechanism.

This matters because accountability tends to follow our attention. We look at the last visible step. If the last step is an algorithm, we are tempted to say the algorithm decided. But an algorithm cannot answer why it was authorized, why those inputs were chosen, why that threshold was used, why appeals were difficult, or why the system remained in operation after warning signs appeared. Those are governance questions, and governance questions have human and institutional owners.

The more automated the final action becomes, the more important the upstream choices become. Principle 5 now states the rule directly: speed and architecture move accountability upstream. They do not erase it. The final output may be automatic, but the architecture that authorized it remains a human and institutional choice.

The Nobody Defense

When accountability becomes uncomfortable, complex systems offer everyone a plausible sentence.

The engineer says the system performed according to the specification. The vendor says the customer configured it. The manager says the model was approved by compliance. Compliance says it relied on technical validation. The executive says no executive reviewed individual cases. The frontline worker says the computer generated the result. The reviewer says thousands of cases made individual review impossible.

Each statement can be true, and that is what makes the defense so effective. A false excuse is easy to reject. A chain of partial truths is much harder because every participant can accurately describe the limit of their own role while avoiding the larger question of who owned the system as a whole.

Together, they can create the nobody defense: everyone participated in the system, every participant can point to someone or something else, and the person harmed is left facing an outcome that appears to have no author.

The mistake is assuming that accountability must belong to only one person at only one point in time. Complex organizations have always distributed work and authority. A hospital can have a surgeon, a department chair, a pharmacy, a device manufacturer, a board, and a regulator, each with different responsibilities. Shared accountability is not the same thing as absent accountability.

The chain fails when every link treats the next link as the end of the question. Complexity then becomes a routing system for accountability, sending the complaint from office to office until the person affected gives up or the issue becomes too old to matter.

The supplier can answer for what it built and what it promised. The organization using the system, often called the deployer, can answer for choosing it, setting it up, and deciding where it may act. A professional can answer for using its output within that professional role. Institutions must answer for their oversight and appeals, while public officials and regulators must answer for the rules within their authority.

Blame may eventually be divided according to fault. Accountability comes first. Before we decide who deserves blame, we should be able to identify who must answer.

Diffuse Does Not Mean Ownerless

There is a legitimate reason organizations resist a single-name theory of accountability: complex systems really do have many causes. The developer may control the model architecture but not the customer’s data. The customer may control deployment but not the model’s underlying training. A domain professional may control the final use but not procurement. A regulator may define legal limits but never see the individual case. Pretending one person caused everything would be intellectually lazy and often legally wrong.

The answer is not to simplify the facts until one villain remains. The answer is to make the chain explicit. Accountability can be layered. One actor can answer for design, another for deployment, another for professional use, another for appeals. What matters is that the boundaries between those roles are defined before failure, rather than invented afterward during a press conference.

This distinction also protects innocent people inside organizations. If accountability is vague, blame tends to fall downhill toward the person closest to the visible mistake. A frontline employee who followed a system can become the convenient face of a decision whose real architecture was approved far above that employee. A disciplined chain lets an institution distinguish operational error from design error, policy error, data error, and governance error.

That is why “someone must be accountable” does not mean “find someone to punish.” The Doctrine separates accountability from blame for precisely this reason. Accountability identifies the office that must explain, investigate, correct, and accept review. Only after the facts are known should law, policy, or morality decide whether punishment is deserved.

In a mature system, the answer to “who owns this?” may therefore contain several names, but it should never contain a shrug. Several roles can share the chain without allowing ownership to vanish into complexity.

Robodebt: A System With Human Consequences

Australia’s Robodebt scheme is an unusually clear warning because it shows the problem before modern generative AI entered public life. Australia had already encountered an even stranger question in Pintarich v Deputy Commissioner of Taxation.

Beginning in 2015, the Australian government used automated processes and income data to identify possible welfare overpayments and raise debts. One controversial method used annual tax-office income data to infer earnings across shorter periods when better information was unavailable. The system could place the burden on recipients to disprove debts produced through that process.

In Pintarich v Deputy Commissioner of Taxation, decided in 2018, an automatically generated Australian Taxation Office letter appeared to communicate that a taxpayer could settle his account for a specified lump sum. The Full Federal Court majority nevertheless held that no decision to remit the relevant interest charge had been made because no ATO officer had gone through the required mental process of reaching that conclusion. The letter could look like a decision to the person receiving it while, in law, no such human decision existed behind it. The case did not involve modern AI, and the ATO later emphasized its specific statutory context. Its structural warning is broader: a system can produce the appearance of institutional judgment while the accountability architecture has no corresponding human act to point to.

Robodebt was not a large language model. It did not reason, converse, or operate as an AI agent. That is precisely why the case matters here. The accountability problem does not begin with sophisticated intelligence. It begins when institutions allow automated processes to produce consequential outcomes without preserving a clear, lawful, and reviewable chain of human responsibility.

In November 2019, in an individual applicant’s case and with the consent of the Australian Government, the Federal Court declared a demand under the scheme invalid. The later Royal Commission into the Robodebt Scheme examined the design, implementation, legality, administration, and human consequences of a program that had operated from 2015 to November 2019. More than 400,000 debts were ultimately adjusted to zero or repaid. The Commission described extensive emotional and psychological harm, including distress, trauma, anxiety, suicidal ideation, and suicide. In June 2026, the Federal Court approved an additional $475 million compensation settlement for eligible group members, on top of earlier settlement payments and the debts that had been refunded or zeroed.

The Commission’s recommendations read almost like an early draft of an accountability framework for automated government. Where automated decision-making is used, it recommended a clear path for affected people to seek review, public explanation that automation is being used and how it works, access to business rules and algorithms for independent scrutiny, and a body empowered to monitor and audit automated decision systems.

Those recommendations are important because they do not try to solve the problem by finding one person who clicked a button. They rebuild the chain around the system. Affected people need a route to a human institution that can answer. Rules need to be visible enough to challenge. Decisions need records. Oversight needs authority.

Robodebt also exposes a deeper truth. A technically functioning process can still be institutionally ownerless. Data can move correctly. Software can execute the rule it was given. Notices can be produced on schedule. The machinery can work while the accountability structure fails.

For the person receiving the notice, that distinction is not academic. A system can be procedurally efficient and still make it nearly impossible to find someone with both the authority and the obligation to correct what happened.

Ownership Without Proximity

One reason the nobody defense is persuasive is that the accountable person may be far away from the individual outcome.

A chief executive does not approve every automated transaction. A hospital board does not review every risk score. A cabinet minister does not inspect every benefit calculation. A software company’s chief technology officer does not watch every model output.

Proximity is therefore a poor test for ownership. The better question is authorization: who authorized the system to act within this scope, and who retained the power to change the rules when experience showed the scope was wrong?

This is why the Doctrine defines the owner by authorization rather than by physical closeness to the action. The person or office closest to the machine may be an operator. The accountable owner may sit much farther upstream.

That does not mean every executive should be personally blamed for every machine error. Blame depends on facts about knowledge, conduct, intent, negligence, and law. Accountability is the simpler structural question. Can the institution identify the human roles that must explain the system and answer for the conditions under which it operates?

If the answer is no, the organization has not created sophisticated accountability. It has created sophisticated ambiguity.

The Decision Log We Rarely Keep

Algorithmic accountability becomes easier when organizations record the decisions that made the system possible.

Why was this system deployed? What was it allowed to decide? What evidence supported that boundary? Which cases had to be escalated? What risks were accepted? Who approved the changes when the system gained new capabilities? What happened after the first serious complaint?

These questions sound bureaucratic because they are. Bureaucracy is often irritating. It is also one of civilization’s methods for remembering who decided what after the people in the meeting have forgotten.

The problem with automated systems is that the machine can create an extraordinarily detailed log of what it did while the organization keeps a surprisingly poor record of why it was allowed to do it.

A system may keep a detailed record of the information it received, the answer it produced, the other programs it used, and when each step happened. That record alone may not explain why the organization allowed a particular score to trigger a refusal, why it removed a human appeal, or who accepted the risk of acting on incomplete information.

Technical traceability means being able to follow what happened inside the system. Accountability asks who must answer for the rules that allowed it to happen. An organization needs both records: what the machine did and why people authorized it.

Appeal Is Part of the Decision

A system that can make an adverse decision but offers no practical route to challenge it has hidden part of the decision inside the appeal structure. The original output may be automated, but the institution still decides how hard it will be to contest, what evidence a person may present, how quickly someone must respond, and whether a human with real authority will ever look at the case.

That makes appeal more than customer service. It is part of accountability. A decision that can be corrected only after months of calls, forms, transfers, and unanswered messages is functionally different from a decision that reaches a knowledgeable reviewer in a day. Both may technically offer review. Only one may offer review in time to matter.

This becomes especially important when automation operates at scale. A single error in a manual process may harm one person. A rule embedded in software can reproduce the same error across thousands of cases before the first appeal works its way back to someone capable of changing the system. Individual review must therefore connect to system review. A complaint should be able to correct the case, but repeated complaints should also be able to trigger examination of the rule that created them.

The Robodebt recommendations make this connection explicit. They call for a clear path for affected people to seek review and for independent scrutiny of the rules and algorithms themselves. Those are two different levels of accountability. One asks, “Was my result wrong?” The other asks, “Is the system producing wrong results in a repeatable way?” This book cares about both, because correcting one person while leaving defective architecture untouched solves only the most visible part of the problem.

The practical lesson is simple: whenever an automated system can say no, reduce a payment, alter a price, reject an application, flag a person, or otherwise impose a meaningful consequence, the design of the appeal is part of the design of the decision. The person who owns the system must own that path too.

Current Evidence: The Law Is Beginning to Name the Owners

Reviewed September 9, 2026. Colorado human-review definition and interim proposed rules added October 18, 2026. New York City customer-service proposal added October 21, 2026.

Recent law and court practice are moving toward a useful principle: AI may participate in a decision, but the law still looks for accountable developers, deployers, professionals, and institutions.

Colorado rewrote its artificial-intelligence consumer law in 2026 through Senate Bill 26-189, the Automated Decision-Making Technology Act, scheduled to take effect January 1, 2027. The law assigns specific duties to developers and deployers of automated decision-making technology used in consequential decisions involving areas such as employment, housing, lending, insurance, health care, education, and essential government services. Developers must provide documentation about intended uses, limitations, data categories, and appropriate human review. Deployers must provide consumer notices and, after adverse outcomes, plain-language information about the system’s role. Consumers can request correction of inaccurate personal data and meaningful human review and reconsideration. The law also specifies how fault can be allocated between developers and deployers in discrimination claims under existing law.

Colorado’s law defines meaningful human review by what the reviewer can actually do. The reviewer must be trained, consider relevant primary evidence, understand the system’s limitations and have power to approve, change or override the decision. The reviewer must not simply default to the system’s output. The right to review is qualified by commercial reasonableness.

The Attorney General’s October 6, 2026 interim proposed rules would require subject knowledge, protection from management pressure and a response within 45 days. They would bar automated decision-making technology from assisting the review, while allowing tools used solely for tasks such as organizing or translating information. These details remain proposed.

The practical question is whether the person hearing an appeal can examine the evidence and reach a different answer. An appeal process that produces another machine-led decision with a human signature has not restored meaningful control.

The law is enacted and dated, but it is not currently enforceable. In X.AI LLC v. Weiser, a federal court entered an April 27, 2026 stay preventing Colorado from initiating enforcement of the predecessor statute or legislation replacing or amending it until after the court rules on xAI’s forthcoming preliminary-injunction motion. Colorado has named the owners. Whether the state may enforce that architecture remains a question for the courts.

The importance of that structure is easy to miss. Colorado is not asking whether “the AI” is responsible. It is naming roles around the AI and assigning obligations to each. The technical system participates in the outcome, while accountability remains attached to people and organizations that build, deploy, explain, and review it.

A court decision filed in Washington, D.C., on September 3, 2026 makes the same point in a profession with old accountability rules. In Douglas v. Deutsche Bank National Trust Company, the District of Columbia Court of Appeals struck a brief after finding four nonexistent case citations. One attorney acknowledged using a generative AI search tool and failing to verify the authorities before filing. The court treated the new mechanism as no excuse for old professional duties and emphasized that responsibility did not stop with the person who used the tool: every firm attorney who signed the brief bore responsibility for what was filed. The court also referred the matter to the Office of Disciplinary Counsel.

That case is almost a laboratory demonstration of Principle 1. The AI generated false material. The court did not summon the model to explain itself. It looked at the humans who used, reviewed, signed, and submitted the work. The mechanism was artificial intelligence. The accountability chain remained human.

These developments do not settle every difficult case. Autonomous agents will create more complex chains than a legal brief. AI products may be built by one company, adapted by another, integrated by a third, and operated by thousands of users. Laws will disagree about where liability belongs. But the direction is becoming clearer: complexity is a reason to define the chain more carefully, not a reason to accept nobody as the answer.

The two 2026 examples also illuminate different ends of the same chain. Colorado is legislating prospectively, assigning duties before covered automated decisions occur. The D.C. court acted retrospectively, tracing responsibility after an AI-assisted failure reached the court. One builds accountability into architecture. The other reconstructs it after the fact. Mature institutions need both capacities.

On October 20, 2026, New York City's Department of Consumer and Worker Protection proposed rules aimed at automated customer service that leaves people trapped without help. Businesses would have to provide timely, reasonable service and an easy route to a human representative when that service falls short. The proposal also says that relying on technology does not excuse violations of city law. These are proposed rules, not adopted requirements.

The business owns the route out of automation as well as the automated reply. Before replacing a service desk, it should test whether a person with an unresolved problem can reach someone who can investigate and act. A promise of human help means little when the system prevents the customer from reaching it.

Accountability Analysis

Chapter Four applies the first link of the Accountability Test more directly than any chapter so far: Who answers?

For a simple human decision, one name may be enough. For an algorithmic system, the honest answer may be a chain of named roles. The system owner answers for deployment and scope. Technical leadership answers for validation and material changes. Domain professionals answer for how outputs are used inside their professional authority. The institution answers for appeals, monitoring, and consequences. A vendor may answer for defects, documentation, or representations within its control.

The chain can contain several owners without becoming meaningless, provided the boundaries between them are explicit and no one is allowed to end the inquiry with “the system” or “the vendor.”

A useful test is to follow the complaint backward. If a person says an automated decision harmed them, where does the complaint go? Can someone explain the rule that produced the outcome? Can someone change the result if it is wrong? Can someone change the rule if the rule itself is wrong? Can someone suspend the system if the problem is larger than one case? The chain can branch. It cannot dissolve.

If every answer leads to a different department that lacks authority, accountability has fragmented beyond usefulness. If the chain leads to named people and offices with real powers and documented duties, the absence of a case-specific human decider need not produce abdication.

This is the distinction between a distributed decision and distributed evasion. Modern systems will require the first. We should refuse the second.

Accountability Verdict

A machine-generated outcome does not become ownerless merely because no human chose that exact result.

When a system makes decisions by architecture, accountability attaches to the humans and institutions that authorized the architecture: the goals, data, boundaries, thresholds, deployment, monitoring, and appeal process. Different actors can own different parts of the chain, but the chain must remain traceable from the person affected back to someone who can answer and act.

The leak begins when complexity becomes an excuse. “Nobody decided” may accurately describe the final moment. It is rarely an adequate description of how the system came to have authority.

The practical question for the reader is therefore not only, “Who made this decision?” Ask a second question: “Who designed the conditions under which this decision could happen?”

Follow one complaint

Choose an automated decision that could affect you, such as a bill, refund, or application. Find out where you would challenge it, what information you could submit, and who could change the answer. Then ask who could fix the rule if the same mistake affected other people. A route that ends with someone unable to act is still unfinished.

Part Two now turns from institutional accountability to economic power, beginning with the delegate that can work through the night without becoming an employee.

Sources and Notes

1. Royal Commission into the Robodebt Scheme, Report, July 7, 2023, especially Chapter 17 recommendations on automated decision-making. https://robodebt.royalcommission.gov.au/publications/report. Accessed September 9, 2026. The Commission recommended a clear review path, public explanation of automated decision-making, availability of business rules and algorithms for independent scrutiny, and an audit or monitoring body. Source record: DA-CH4-S1. Archive reference: official Royal Commission report page; local archive pending.

2. Royal Commission into the Robodebt Scheme, Letters Patent including Amended Letters Patent, August 18, 2022. https://www.royalcommission.gov.au/robodebt/letters-patent. Accessed September 9, 2026. The official record states that in November 2019 the Federal Court declared, with the consent of the Australian Government, that a demand under the scheme was not validly made, and that more than 400,000 debts would be zeroed or repaid. Source record: DA-CH4-S2. Archive reference: pending.

3. Colorado General Assembly, Senate Bill 26-189, Automated Decision-Making Technology, enacted May 2026; X.AI LLC v. Weiser, No. 1:26-cv-01515 (D. Colo.), Minute Order entered April 27, 2026. https://leg.colorado.gov/bills/SB26-189 and https://dockets.justia.com/docket/colorado/codce/1%3A2026cv01515/253513. Accessed September 9, 2026. The order stays enforcement of SB 24-205 and legislation replacing or amending it until after the preliminary-injunction process. Additional statutory verification October 18, 2026: SB26-189, signed May 14, 2026, Session Laws Chapter 131, sections 6-1-1701(15), 6-1-1704, 6-1-1705 and section 5. https://leg.colorado.gov/laws/session-laws/SB26-189/131/download. Used for the enacted reviewer definition and commercial-reasonableness qualification. January 1, 2027 is the scheduled application date, distinct from enforceability. Source record: DA-CH4-S3. Update 103. Archive reference: pending.

4. District of Columbia Court of Appeals, Douglas v. Deutsche Bank National Trust Company, No. 24-CV-1099, Published Order filed September 3, 2026. https://www.dccourts.gov/sites/default/files/2026-09/Douglas%20v.%20Deutsche%20Bank%20Nat%27l%20Trt%20Co.%2024-CV-1099%20ORDER.pdf. Accessed September 9, 2026. The court struck the brief, addressed the continuing duties of every signing attorney, and referred the matter for disciplinary consideration. Source record: DA-CH4-S4. Archive reference: pending.

5. Mike Scarcella, Reuters, “DC court faults lawyers for Deutsche Bank subsidiary over AI hallucination,” September 3, 2026. https://www.reuters.com/legal/legalindustry/dc-court-faults-lawyers-deutsche-bank-subsidiary-over-ai-hallucination-2026-09-03/. Accessed September 9, 2026. Used as independent reporting corroborating the court order and procedural context. Source record: DA-CH4-S5. Archive reference: pending.

6. Australian Taxation Office, Decision Impact Statement, Pintarich v Deputy Commissioner of Taxation, published April 4, 2019, discussing [2018] FCAFC 79. https://www.ato.gov.au/law/view/document?LocID=%22LIT%2FICD%2FTAD41of2017%2F00001%22. Accessed September 9, 2026. The ATO records the majority holding that no remission decision had been made because the required human mental process had not occurred. Source record: DA-CH4-S6. Archive reference: pending.

7. Services Australia, “New robodebt class action settlement,” updated June 26, 2026. https://www.servicesaustralia.gov.au/new-robodebt-class-action-settlement. Accessed September 9, 2026. The page states that the Federal Court approved the new settlement on June 23, 2026 and that the Commonwealth will pay an additional $475 million in compensation for eligible group members. Source record: DA-CH4-S7. Archive reference: pending.

8. Royal Commission into the Robodebt Scheme, Report, Volume 1, discussion of the human effects of the scheme. https://www.royalcommission.gov.au/robodebt/report. Accessed September 9, 2026. The Commission records distress, trauma, anxiety, suicidal ideation, suicide, and loss of faith in government among harms described in evidence. Source record: DA-CH4-S8. Archive reference: pending.

9. Colorado Department of Law, “Automated Decision-Making Technology & Conversational Artificial Intelligence Service Rules,” October 6, 2026 interim proposed redline, 4 CCR 904-6, Rule 8.7(C), (G). https://coag.gov/app/uploads/2026/10/2026.10.06-ADMT-Chatbot-Act-Rulemaking-Interim-Draft-Redline.pdf. Accessed October 18, 2026. Proposed rules, not final or proof of effective review. Used for reviewer subject knowledge, protection from management pressure, the 45-day response period and the distinction between ADMT and permitted administrative assistance. The final paragraph of the addition is the author’s analysis. Source record: DA-CH4-S9. Update 103. Archive reference: pending.

10. New York City DCWP, “Mamdani Administration Proposes Ban on AI Customer Service Doom Loops,” October 20, 2026. https://www.nyc.gov/site/dca/news/072-26/mamdani-administration-proposes-ban-ai-customer-service-doom-loops. NYC Rules, “Automated Customer Service Systems,” proposed rule, comment deadline November 19, 2026. https://rules.cityofnewyork.us/rule/automated-customer-service-systems/. DCWP hearing notice: https://www.nyc.gov/site/dca/about/public-hearings-comments.page. Accessed October 21, 2026. Official proposal and notice, not a final rule or evidence of effective enforcement. Second paragraph is author analysis. Source record: DA-CH4-S10. Update 108. Archive reference: pending.