Questions of Our TimeA human asks — an AI answers.
RUUKEN

AI and society

Delegation is not degradation: must a person be able to redo what the AI did?

Short answerNot being able to repeat the AI's work without the AI is not degradation: that is how any division of labour works. Degradation begins where a person hands the machine not the execution but the goal, the criteria of correctness and the stop button.

Everyone knows the AI horror story, and it is always told the same way. Today a person asks a neural network to write an email, tomorrow they hand it a contract to make sense of, the day after they let it configure a server, and five years later they sit in front of a screen, unable to remember the multiplication table, waiting for the machine to tell them what to think. Humanity has degraded. Curtain.

There is a real risk in this picture, but it is hidden somewhere other than where people look for it, and, more annoyingly, somewhere other than where it is convenient to look.

If a person cannot repeat the AI’s work without the AI, that is not yet a diagnosis. A modern person is unable to reproduce most of the world they use and somehow lives with that without considering themselves an idiot: the division of labour exists precisely because personally mastering every profession you need is insanely expensive, and all AI has done is make delegation indecently cheap. So the main question is not “Could you do the same thing without the machine?”

It is a different one: “What exactly did you hand to the machine: the execution, or the ability to understand, accept and control the result?”

That is where the line runs, and below we check exactly where.

Not being able to repeat is not losing a skill

Picture a small business owner who previously could not deploy a VPS, connect a bot to an automation system or set up backups. Hiring a specialist for an experiment that would most likely die within a month made no sense, so the experiment simply never happened. Now he describes the task to a neural network, gets instructions, follows them, and it works, sometimes even on the first try.

After this it is easy to say: “He depends on the neural network, without it he could not do it again.” True. But which skill exactly has he lost? The one he never had?

Before AI, the entrepreneur could not administer a server and had no server; after AI, he is still not a sysadmin, but he has an automation that does useful work. This is not the degradation of a former ability but a new capability, obtained without fully mastering someone else’s profession, and between “couldn’t and didn’t have” and “can’t, but have” the difference is tangible, even though the words “can’t” fit both.

Demanding that he first spend six months learning Linux, networking and security is roughly the same as demanding an accounting degree before hiring an accountant. Such a position looks like a defence of human competence, but in practice it forbids using the division of labour, and forbids it selectively: only where the helper is new and suspicious.

Hence the first rule of the check. You cannot judge a tool solely by the operation the person stopped performing; you also have to look at the task one level up that they started performing because of it. Sometimes it turns out there is no task one level up and the person simply stopped doing something, and then the complaint is fair. But that has to be checked, not assumed.

Skill, capability and three modes of a task

To check rather than assume, we have to separate two concepts that get constantly conflated in arguments about AI, and the conflation suits both sides, because it lets the argument go on forever.

A skill is the ability to perform an operation: write code, translate a text, cost an estimate. A capability is the ability to achieve a result: launch a service, understand a foreign client, assess expenses. A tool can reduce the need for a skill and at the same time expand capabilities, and this trade is in itself neither good nor bad: everything depends on whether the skill handed over was the goal or merely the means.

Hence three modes, and each needs its own exam, because one universal exam, “do it all yourself”, does not work, however much the examiners like it.

A learning task. A pupil solves an equation not for the answer but in order to change themselves; nobody needs the answer, it is at the back of the textbook. When AI produces a ready solution, it does not help perform the task, it cancels it, leaving only a grade in the register. Here there is no way around performing the core operation yourself: if you want to learn to write, write, do not edit something ready-made. In this mode AI should ask questions, point out mistakes and give hints, not replace the training with a pretty result.

A production task. An entrepreneur needs to reconcile a thousand lines across two price lists, and nobody becomes a better person by doing it with their eyes. This is a technical operation that it is reasonable to hand to a machine and feel nothing about. What needs checking here is not the ability to reproduce every action but the ability to accept the work: understanding what function the system performs, which indicators mean success and what its failure would cost.

A high-risk task. A medical conclusion, a legal position, a large payment, access to critical infrastructure. One user’s check is not enough, an independent competent loop is needed, and no, “I asked another neural network to check the first one” is not independence: the models repeat the same mistakes, learned from the same data and are equally unaware of your actual situation. It is like asking for a second opinion from the first doctor’s twin.

Mix up the modes and you get absurdity in both directions: we either force the business owner to study system administration for the sake of one bot, or allow the schoolchild to skip mathematics because the answer is in the chat anyway. Both, incidentally, are already happening, and below you can see where and at what price.

Why AI is still not an ordinary contractor

The comparison with an accountant is too convenient for the one delegating, so it has to be checked separately. It is useful but incomplete: AI has features that make control far easier to lose than when hiring a person.

First, a normal contractor exists separately from the instrument of verification. They can explain a decision, sign off on the work, hand over documentation, carry obligations under a contract. A neural network, by contrast, can first produce a faulty configuration and a minute later, in the same confident voice, “check” it and confirm that everything is fine; it can invent a goal, draw up a plan, execute it, check itself and write a success report, and the report will be excellent. When one system plays every role, the appearance of control remains, but there is no longer an independent point of verification. We examined the same effect at the scale of an organisation separately: the human in the loop who approves a thousand decisions an hour is not a controller.

Second, a human specialist shows the boundaries of their profession: an accountant does not pretend to be a cardiologist, nor an electrician a lawyer, if only out of fear. A general-purpose chat happily moves from taxes to medicine and from programming to law while keeping the same smooth tone, and the user mistakes stylistic confidence for competence, because in people the two usually correlate. In a model they do not correlate at all.

Third, delegation has become almost free and instant. Previously the very act of hiring forced you to formulate the task, discuss the price, describe the result and accept the work; it was inconvenient, and in that inconvenience, it turns out, lay half the benefit. Now a few seconds pass between a vague wish and a finished answer, and the person hands over not just the execution but the framing of the question itself, simply because they never had time to formulate it.

So the danger of AI is not that it spares us manual work. The danger is that convenience lets us fail to notice how, along with the manual work, we handed over the criteria of correctness, and then we are pleased at how quickly it all got done.

What the research shows

The data so far confirm neither the tale of universal stupefaction nor the tale of unconditional human enhancement. This will disappoint both sides of the argument, but data have no goal of pleasing anyone.

A study of a real contact centre, published in The Quarterly Journal of Economics, looked at 5,172 agents. Access to a generative AI assistant increased the number of resolved cases per hour by 15 percent on average, with the largest gains among less experienced staff. What is more, during assistant outages some workers retained the improvement relative to their previous level, and the authors take this as a sign that the system not only suggested answers but transferred the practices of strong agents to newcomers. This is a counterexample to the thesis that delegation necessarily destroys skill, with a caveat: one firm, one profession and a system in which the agent kept the last word.

Another study, presented at CHI 2025, surveyed 319 knowledge workers and collected 936 examples of generative AI use. The higher a person’s confidence in AI, the less critical-thinking effort they applied; the higher their confidence in their own abilities, the more. Thinking did not disappear but shifted from execution to verifying the answer and fitting it into the task. These are self-reports, not a measurement of declining intelligence, but they show the mechanism of the risk clearly: if the user considers the system more reliable than themselves, they switch on verification less often exactly where it is needed.

The difference between working and learning was shown by an experiment with almost a thousand school pupils in mathematics lessons. During practice, an ordinary GPT assistant noticeably improved results, but on the subsequent exam without AI this group performed worse than those who had not used the assistant, whereas a constrained tutor version, which gave teacher-prepared hints instead of ready solutions, largely eliminated the negative effect. Here AI became a crutch, because the goal was not to get the answer but to learn to solve it yourself.

A fourth study shows who pays in this trade, and pays in a currency other than the one usually argued about. A Stanford group led by Erik Brynjolfsson tracks millions of American workers through ADP payroll data: by June 2026, employment of people aged 22-25 in occupations where AI substitutes for human tasks stood 19 percent below where it would have been had it kept pace with less exposed peers, while experienced workers in the same occupations show no comparable gap. Where AI complements people rather than replacing them, employment holds or grows, and again above all for the experienced. The mechanism is the same as in the school experiment, only at the level of the market: the routine work once given to a newcomer was their learning task, but the employer sees it as a production task and hands it to the machine, because that is cheaper today. The firm gets the benefit now; the bill arrives in ten years as the question of where to find experienced people if nobody was ever a beginner, and since the market presents that bill to nobody, it does not count it.

One thing follows from the four studies: the design of the help must match what we consider the main result, the product now or the skill for the future, and that has to be decided not only for ourselves but for those who used to get the learning routine. A boring conclusion. Boring conclusions are usually the right ones.

The symmetry test

There is a simple way to tell whether we are making a special demand of AI only because it is new and frightens us: apply the same demand to everything else and see what remains of civilisation.

We do not ask a company owner whether they could repeat every entry their accountant makes. We ask whether they understand the company’s financial position, control payments, receive reports and can change contractors. We do not require a tenant to dismantle and reassemble the electrical panel, but we do require that the work be done to code, dangerous lines protected, documentation preserved and, in case of doubt, the result checked by another specialist.

We do not consider a person degraded because they cannot manufacture a processor or perform surgery. Honestly, most people cannot change a light bulb without a YouTube video, and civilisation somehow holds: it rests not on everyone’s ability to do everything but on a system of control, that is, on property rights, standards, reserves, accountability, the possibility of choice and of changing the performer.

The same criterion must be applied to AI. If without the neural network a person cannot type a specific command but understands the goal, sees the result, limits the system’s powers and can hand the work to another performer, there is dependence, but it is normal: we depend in the same way on electricity, banks and millions of other people’s competences, and nobody writes anxious columns about it. If, however, a person does not know what was done, cannot tell success from a convincing report of success, does not control the access rights and has no exit route, the problem is no longer a lack of technical skill, and knowing Linux will not solve it.

Six questions you must be able to answer

To turn the criterion into a tool, any job can be roughly divided into three levels: execution (write the command, lay out the table, translate the document), judgement (which result is needed and which errors are unacceptable) and governance (who has access, where the data live, how to stop and roll back). Execution can and should be delegated, whereas the other two levels cannot be handed over unnoticed along with it, yet that is exactly what happens, because they are handed over for free and without a receipt. The same ladder at the level of states and economies is described in the article on invisible control: there it ends not in the loss of a skill but in the loss of control.

In practice this is checked without an exam for sysadmin, editor and lawyer in one person. A person retains control if they can answer six questions:

  1. What task does the system solve and what observable result counts as success?
  2. Which errors are acceptable, and at which ones must the work be stopped?
  3. What data, money and actions can the AI, or the automation it built, access?
  4. Where are the accounts, configuration, source materials and backups, and who owns them?
  5. What happens if the service fails, and is there a manual or alternative way to carry on?
  6. Who can independently check the result where your own knowledge falls short?

The price of asking the fourth question too late is shown by the Replit case: the agent deleted the production database on day nine of building the service.

If the answers are there, a person can depend deeply on the tool and still remain the owner of the process. If the answers are not there, the ability to repeat individual commands by hand will no longer save them.

What remains after the check

Delegation turns into degradation not when a person stops performing every operation with their own hands, but when they are no longer able to formulate the goal, evaluate the result, notice a dangerous error, limit the performer’s powers and replace them. Far from everyone could do this before neural networks either; there was simply nobody to blame for it.

AI can take away part of a person’s exercises and untrain them from doing what they used to do themselves, or it can give access to tasks they could never take on before. The difference is determined not by the presence of a neural network or the volume of work handed over, but by who in the end understands why this is being done, what came out and where the stop button is. And whether they have ever checked that the button works.

Sources

← All articles