The Last Human Signature

Imagine that the work is ready. The analysis is complete, the alternatives have been compared, and the recommendation is waiting for your approval. A machine has done much of what once would have occupied your afternoon. You can read the result in minutes. At the bottom of the screen is a place for your name.

What are you being asked to supply? If the answer is merely a click that allows the process to continue, your presence adds little. If the answer is a guarantee that every part of the work is correct, the request is impossible. Between those two answers lies the obligation this book has tried to describe: to exercise the authority that is yours and remain answerable for its use.

The last human signature is not necessarily the last human action in a workflow. It belongs to the person or institution whose judgment gives the work permission to become a decision. Sometimes that judgment concerns a particular case. Sometimes it concerns the earlier authorization under which thousands of routine actions proceed. Other people may hold other duties. None of those duties disappears because one name occupies the final box.

By now, the machine may have done more of the visible work than you have. That can be a good arrangement. The question is what your name means when the arrangement is challenged, when someone needs a correction, or when continuing would require a judgment that the machine was never authorized to make.

What the Name Commits

A signature ordinarily arrives near an ending. A document is completed, a proposal accepted, a decision released. Accountability gives it a longer life. The person affected by the decision encounters its consequences after the signer has moved to something else. An institution needs to remain reachable across that distance. Its records, authority, and correction process must survive the moment of approval.

The signer does not have to repeat every calculation. We already rely on work other people produce. The decision owner does need sound reasons for trusting the work, an understanding of its important limits, and the ability to respond when someone questions it. AI changes the amount and speed of delegated work. It also makes plausible output easier to obtain than an adequate basis for accepting it.

Consider a hypothetical manager reviewing an AI-assisted report before it goes to employees. Most of the report concerns routine operations. One passage recommends eliminating a service on which a small group depends. The recommendation may be well supported as an estimate of cost. The manager still has to decide whether that estimate answers the question the organization should be asking. The people who use the service may know consequences that the report has not considered.

The manager’s contribution need not be a better calculation. It may be recognizing that the calculation leaves a decision unfinished. Seeking the missing evidence, hearing from those affected, changing the proposal, or declining to authorize it can all be exercises of judgment. Approving it can be an exercise of judgment too, provided the reasons and the authority are real. The signature should distinguish among those possibilities rather than conceal that none was examined.

This is why an impressive explanation from a model cannot perform the whole duty of an owner. An explanation can help someone assess a proposal. It cannot decide whether the proposal deserves the authority of a particular office, profession, or institution. That authority brings obligations to people beyond the conversation in which the answer was generated.

Nor does a sincere signer make an inadequate arrangement sufficient. The employee may care deeply and still lack time, information, or permission to change the result. A rubber stamp is abdication wearing accountability’s clothes. The institution that requires the stamp remains responsible for the conditions under which it is applied. Naming the employee cannot settle the question of who removed the opportunity for judgment.

The Freedom to Delegate

There is a hopeful possibility here. If machines can take on more of the work that consumes our attention, we can use some of the recovered capacity to make better decisions about purposes, exceptions, and consequences. A smaller business might gain analytical help it could not previously afford. A professional might spend less time assembling material and more time considering what the person in front of them needs. Those are possibilities worth pursuing, not reasons to treat every saved minute as a minute of human capacity that should be removed.

That freedom depends on what happens to the time saved. If staff receive more cases without more time to consider each one, review may get weaker. If people stop practicing their judgment because the system usually sounds convincing, they may miss the unusual case that matters. An institution should decide how it will use the time and attention AI releases, including what it will keep available for checking and difficult decisions.

There is no virtue in requiring a human to repeat a task simply to preserve the appearance of involvement. Where a bounded action can proceed under valid authority, routine autonomy can be the sensible choice. The owner’s work then includes maintaining the boundary, examining whether the arrangement continues to meet its purpose, and responding when conditions change. Accountability should be located where it can be exercised effectively.

Other decisions require judgment over the specific case. Tier 3 addresses irreversible, high-stakes consequences. Reserved identifies judgments whose legitimacy depends on an accountable human decider. These categories must retain their distinct reasons even when their practical safeguards overlap. Earlier permission to use a system cannot substitute for the independent human judgment required by a Reserved decision.

A machine may make a better prediction while a person still has a duty to decide what should happen. That duty does not make the person’s prediction correct. Institutions need to use the machine’s strengths, check the quality of human judgment, and establish who has the right to make the decision.

That combination asks more of us than either enthusiasm or refusal. Enthusiasm can mistake capability for permission. Refusal can leave useful work undone without explaining what has been protected. A responsible owner must be prepared to give reasons for using the system, limiting it, or doing the work another way. The duty is to the decision and the people it affects.

Current Evidence: Making Review Substantive

Reviewed September 13, 2026.

Research on human review offers a useful caution. In a 2021 experiment with 199 participants, Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos compared several AI-assisted decision designs. Interventions intended to require more deliberate thought reduced acceptance of incorrect AI advice relative to simpler explainable-AI designs. The interventions that reduced overreliance most also received the least favorable subjective ratings. Benefits varied with participants’ motivation to engage in effortful thinking.

This was a bounded experiment, not evidence that a particular institutional approval process is safe. It suggests that review design can affect reliance, while also creating friction and uneven benefits. It does not establish that adding a pause or an explanation produces accountable judgment, or that laboratory results transfer unchanged to demanding workplaces.

The practical lesson is to test whether a promised safeguard works. Can reviewers spot an important error and act on it with the time, information, and workload they actually have? Future evidence about repeated use and successful intervention should inform this chapter. Calling a process human-reviewed is a claim that needs support from what people can actually do.

The Name on This Book

My name belongs on the same side of this argument as anyone else’s. The Delegation Age has been produced with extensive AI assistance. The project’s production policy permits AI across research discovery, comparison, critique, drafting, editing, and other parts of preparation. Final editorial authority remains with me. That is a practical division of work, and it carries practical obligations.

I do not claim that writing about accountability makes the book immune to the failures it describes. A source can be misunderstood. A qualification can disappear during revision. A clear sentence can carry more certainty than the evidence supports. Multiple rounds of review can miss the same problem. The fact that a claim survived a process is a reason to inspect that process when the claim fails, not a sufficient answer to the failure.

The book’s chapter reviews, author approvals, source records, and version history are intended to make substantive decisions traceable. They do not amount to a guarantee that every claim is right. Where an archive reference remains pending, that work remains unfinished. A record should say what has been completed and what has not, even when a cleaner presentation would be more comfortable.

If a reader identifies a material error, “the AI wrote that” would describe part of the production history. It would not answer the reader. The editorial obligation is to examine the challenge, check the evidence, and correct or withdraw what cannot be supported. A correction should be visible enough for readers to understand what changed and why. The production method cannot serve as an excuse that becomes available only after publication.

This also limits what I should claim about my own role. Approval does not mean that I personally originated every sentence or independently reproduced every result cited. It means that I accept responsibility for the editorial decision to put the work before readers and for responding when that decision needs reconsideration. The work of verification must be proportionate to the claim; naming its owner cannot replace doing it.

The same standard applies to the argument itself. The Doctrine should be challenged where its categories are unclear, where its safeguards prove inadequate, or where better evidence changes the case for a particular delegation. Defending the book must include the willingness to revise it. A framework that cannot learn from its own use would be a poor guide to governing systems that change.

I want the assistance because it can make work possible, improve it, and expose weaknesses I might otherwise miss. I accept the editorial obligation because those benefits do not belong to a process that can answer a reader in my place. Authorship in this arrangement is an ongoing responsibility for what I have chosen to publish, including what later needs repair.

Accountability Analysis

The signature brings several parts of the Doctrine together. Authority must reach the decision being made. Information and competence must support its exercise. Intervention must be possible at a point when it can matter. Consequences must reach the parties whose choices shaped the result. These requirements reinforce one another. A powerful owner without usable evidence may authorize blindly; an informed reviewer without power may see the problem and be unable to change it.

The opening approval screen therefore cannot be evaluated by counting humans. Its meaning depends on what lies behind it. If the signer owns the decision and has a sufficient basis for making it, the approval records an exercise of judgment. If the signer merely clears a mandatory step while authority sits elsewhere, the record can misrepresent how the decision was made. The remedy concerns the allocation of authority and working conditions as well as the interface.

For the hypothetical manager, a defensible approval would require understanding what the service recommendation establishes and what remains unresolved. The organization might still discontinue the service. Accountability does not guarantee agreement with everyone affected. It requires an authorized decision that can be explained, challenged, and reconsidered through a meaningful process. The affected group should not have to defeat a model’s confidence before someone will listen to information the model never considered.

Accountability must also remain distributed according to role. A manager who approves the report answers for that judgment. Senior leaders may answer for the policy and resources governing the review. A supplier may answer for relevant product commitments or defects within its control. Establishing one duty does not extinguish the others. An investigation should follow authority, knowledge, and control rather than assume that the person nearest the final output owns every cause.

This distinction matters especially when someone is asked to sign an arrangement they cannot defend. The immediate response should make the unresolved condition explicit and use the available escalation or stop path. The institutional response must address why the condition exists. A company cannot build a workable safeguard out of repeated acts of individual courage while treating ordinary disagreement as failure to perform. Its claim to human oversight depends on people being able to exercise it in practice.

Accountability is not blame. It provides a structure for obtaining an answer and requiring a response. Responsibility includes the obligations people owe one another through their roles and choices. Blame requires a further judgment about fault. Keeping them distinct allows correction before every causal dispute is settled and allows investigation without making the discovery of an error an automatic admission of misconduct.

The signature also cannot guarantee that repair will be enough. Some harms cannot be reversed or adequately compensated. That is one reason boundaries and case-specific judgment matter before action. Where the decision requires a human decider, a promise to investigate afterward cannot supply the judgment that was missing when the consequence became unavoidable. Where routine autonomy is justified, the owner still has to account for foreseeable failures and the limits of recovery.

Finally, the commitment has to survive change. The signer may leave office, the model may change, and the original reasons for approval may cease to hold. Records make the earlier decision intelligible; renewal establishes whether the current arrangement deserves continuing permission. The duty belongs partly to enduring institutions precisely because the people exercising it are temporary. A person affected tomorrow should have a place to take a challenge even if the person who signed yesterday has gone.

Accountability Verdict

The last human signature belongs where meaningful authority remains. It may be a judgment over an individual case or the owned authorization of bounded autonomous work. It must never be used to turn the final employee in a workflow into a substitute for everyone who shaped the outcome. The owner must have the means to decide and must remain exposed to an appropriate demand for explanation, correction, and consequences.

Take one decision you already allow AI to influence. Before accepting the next result, write down the action you are authorizing, its limits, and who can change or stop it. Use the Accountability Test to identify the control it needs and when permission should be reviewed. Could you explain your reasons to someone affected? Resolve a missing answer or narrow the permission. Where the arrangement is sound, use it with the attention its consequences deserve. Responsible delegation should make useful action possible as well as identify when to stop.

We should want better machines. We should want them to relieve drudgery, extend our abilities, and help us see what we have missed. We should also want the gains to reach people who need better services, more useful knowledge, and institutions that can hear them. Those hopes require choices about how capability is used. No improvement in a model makes those choices cease to be ours.

I cannot tell you how much of your work a future system will be able to perform. The answer will differ by task and will change. I can say what I am asking of the person who accepts its help: know the authority you are exercising, take the remaining judgment seriously, and stay available to those who have reason to question the result. When the work succeeds, the assistance deserves recognition. When it fails, the people affected deserve an answer.

At the end of this book, my name makes the same commitment I have asked you to consider. It does not certify an absence of error. It identifies someone who must respond to the work and its consequences. Whatever part of the work moves next, that obligation stays.

Delegate the work. Never delegate the accountability.

Sources and Notes

1. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos, “To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making,” Proceedings of the ACM on Human-Computer Interaction 5, CSCW1, Article 188, April 2021. Primary locator: https://doi.org/10.1145/3449287. Author publication page and abstract: https://www.eecs.harvard.edu/~kgajos/papers/2021/bucinca2021trust.shtml. Accessed September 13, 2026. The abstract reports the 199-participant experiment, reduced overreliance, subjective-rating trade-off, and variation associated with Need for Cognition. This is bounded experimental evidence, not validation of the Doctrine, its Tiers, or an institutional review process. Source record: DA-CH27-S1. Archive reference: pending.