From Hammer to Chatbot: What Are We Actually Delegating?

Part 3 of 30 · Series date: July 5, 2026

A hammer does not decide whether the house should be built. A spreadsheet does not, by itself, decide what a family owes its elderly members. A chatbot can discuss both decisions. That may tempt us to think it is just a more complex version of an older tool. But the kind of work being delegated matters as much as its difficulty.

We can delegate force, storage, execution, interpretation, recommendation, and the selection of goals. These activities overlap, but separating them helps reveal where human judgment remains necessary.

From execution to interpretation

A jack supplies force according to a mechanical arrangement. A notebook preserves what someone writes. A calculator carries out the math we ask for. These tools can fail or be misused, and their simplicity does not make them harmless. Still, the intended task can often be described relatively narrowly.

An interpretive assistant operates closer to the description of the problem itself. Ask it to explain a disputed passage, assess a plan, or suggest an action. In each case, it must choose what to stress. A helpful answer may explain assumptions. An unexamined answer may install assumptions the reader never noticed.

This is why calling AI “just a tool” is both partly right and insufficient. It reminds us that humans choose goals. But it can hide how a system helps shape those goals. Tools differ in how much they shape the user's next move.

Advice is not automatically domination

The case for broad delegation begins with ordinary cooperation. We rely on editors, accountants, teachers, and trusted friends. Nobody could live well if they had to master every subject before accepting help. Why should assistance from software be uniquely disqualifying?

The answer should not be that all machine advice is unacceptable. It should concern the basis of trust. An expert may have years of experience. They may also have duties to the client and rules they must follow. Software may help too. We need sound grounds to trust it, beyond its smooth words. A well-designed institutional use can provide stronger oversight than an informal human conversation. The comparison must be specific.

Can the person judge the advice? If not, is there a trusted way to check it for them? Suitable trust includes knowing when to seek another judgment. It does not require pretending to know everything.

Delegation moves upstream

Think of a charity deciding how to share a small emergency fund. At first, software adds the amounts requested. Next, it checks whether applications are complete. Then it estimates urgency. Eventually it recommends who should receive assistance. Finally, managers ask it to propose the organization's priorities. These are not simply five levels of speed. They move from carrying out a decision toward deciding what the institution should value.

The early steps still contain choices. A completeness check can disadvantage someone unable to obtain a document. An urgency estimate can confuse well-recorded need with greater need. The later steps make the choices clearer. But values were at work before the system seemed to give moral advice. A numerical process does not become value-free because its assumptions were encoded earlier.

This is why “the human chooses the goal” is an incomplete safeguard. Which human? The board may choose an admirable goal while a software setting determines its practical meaning. A vendor may define success as processing speed; staff may understand it as keeping people housed. Unless those meanings are reconciled, the system can carry an institution toward an outcome nobody openly defended.

Why a tool is not morally neutral in every sense

Calling a tool morally neutral may mean that it has no aims of its own. We do not blame it as we would a person. That does not mean its design makes every use equally easy or equally likely. A locked door favors some movements and prevents others. A form that requires one fixed category makes certain lives easier to describe than others. A ready answer can change our next step.

It may become easier to accept the chatbot's view than to look for other views.

A defender may reply that all environments shape behavior, including libraries and classrooms. Correct. That does not mean influence from a tool is always wrong. Influence deserves review wherever it occurs. Design choices can be judged without pretending a device has a conscience. Moral responsibility belongs to people and institutions whose choices shape the system and its use.

The distinction also prevents an easy escape for users. “I only used the tool” is no defense if someone knew it was unsuited to the decision. Yet a worker may lack the power, facts, or time to step in. Placing all blame on that worker would be unfair too. Duty should follow knowledge, control, and a fair chance to act. It should not fall by default on the last person to click.

A decision that can still be answered for

Imagine an applicant challenging a denial. The group should explain its rules and look at any missing facts. It should name who can change the answer. If an error keeps happening, it should fix the cause. It need not disclose every technical detail to every applicant. It must give clear enough reasons and a real way to appeal. That keeps people answerable for the decision.

The strongest version of delegation preserves that chain. Tools can do more while people remain able to explain the work and fix what went wrong. The weakest version produces an impressive answer surrounded by people insisting that nobody really decided. As tools gain power, does our sense of duty grow too? Or do we use the tool to make it harder to find anyone responsible? Part of the moral line lies there.

The missing responsibility

A dangerous form of delegation appears when a recommendation becomes a way to avoid giving reasons. “The system selected this person” says what happened; it does not explain why the selection was justified. Nor does it identify who must study an error.

People's duties should match their roles. A frontline user cannot fairly bear every consequence of a system they cannot inspect or change. Designers, purchasers, managers, and regulators may each have relevant duties. Shared responsibility must not become a condition in which everyone points elsewhere.

Consider a charity using software to prioritize requests. The tool might organize information efficiently. The charity must decide which needs matter. It must handle missing records fairly and provide a way to appeal. A score cannot settle those moral commitments just by being numerical.

The key is not just whether a machine helped. It is whether people still answer for how that help is used. We can hand over hard tasks while keeping control of their purpose. We can also surrender our purposes through something as ordinary as an unexamined metric. Complexity is not the dividing line. The real test is whether someone can answer for the decision.