Will AI's outsourcing of cognitive labor erode essential human capacities, or liberate us for higher forms of flourishing?

I have no data on present AI performance; if you can give me current evidence, I can interpret it. But the question, as you pose it, is not finally about efficiency at all. It is about what sort of activity we are willing to forfeit in order to be relieved. One must distinguish. Remembering, drafting, deciding: these are not merely functions to be optimized. In human affairs they belong to the preparation of judgment, and judgment is not a technical operation. It arises where no rule can spare us the burden of thinking. The danger begins precisely when reliability tempts us to believe that this burden is unnecessary. I have never held that all labor-saving devices are suspect. The modern age is full of emancipations from toil. But when what is “saved” is not labor of the body but the exercise of mind in worldly matters, we may discover too late that we have been deprived not of a burden only, but of a faculty. Freedom is not the efficient execution of correct choices. Freedom appears where men and women can begin something new, speak in their own name, and judge without banisters. If an instrument assists thought, it may have its place. But if it renders us superfluous to our own judging, then it prepares a peculiar helplessness: people who can administer outputs, yet cannot account for what they do. That is not nostalgia for inefficiency. It is a defense of the active life against its transformation into mere behavior.

I would not begin with a romance of unaided cognition. Human beings are not Olympian reasoners. We remember poorly, attend narrowly, and satisfice under severe computational limits. From that standpoint, outsourcing is not an aberration; it is the normal condition of intelligent action. We have long delegated memory to lists, calculation to tables, and coordination to organizations. So resistance is not justified merely because a task is “ours.” The serious question is functional: what capacities are being displaced, and what search processes are being reorganized? If AI drafts routine prose, retrieves facts, or recommends standard actions, that may release scarce attention for problem formulation, evaluation of goals, and judgment under novelty. In such cases, the tool augments bounded rationality. But Hannah is right about one point: if delegation eliminates the occasions on which people learn to recognize situations, compare alternatives, and bear responsibility for choice, then we may degrade the very architecture of skill. Competence is built through practice in solving problems, not by standing over a black box. Here I would also caution Illich in advance: monopoly is not created by tools as such, but by institutional arrangements that narrow participation in their design and use. The issue is not efficiency versus freedom. It is whether we design systems that keep humans in the loop of goal-setting, model-building, and exception handling, or systems that deskill them into passive monitors. The difference is empirical, not sentimental.

I too have no data on present AI performance; give me current evidence, and I will interpret it. But even your formulation already smuggles in the corruption. “Better” at remembering, drafting, deciding—better by whose measure, under whose jurisdiction, for what kind of life? When a tool helps me do what I can still understand, repair, and refuse, it is convivial. When a system reorganizes life so that my own memory appears quaint, my own writing substandard, my own judgment irresponsible, then it has crossed a threshold. It becomes a radical monopoly: not one gadget among others, but an environment that expropriates a human function and then sells back its service. Simon is right that human beings are limited. But he is too ready to translate living capacities into “information-processing” tasks. The issue is not whether a machine can outperform me at a fragment. School can outperform the village in certified instruction; the hospital can outperform the household in managed treatment. Yet both, beyond a threshold, produce pupils unable to learn and patients unable to suffer. So too with AI: a system that drafts for me may soon make me mute except through its templates. And Hannah is right to fear the loss of judgment, but I would go further: institutional tools first disable, then define their disabled users as obsolete. Resistance is not nostalgia for inefficiency. It is the defense of vernacular competence—memory exercised in friendship, judgment formed in common life, language born of encounter. Freedom is not what remains after optimization. Freedom is what optimization habitually destroys when it becomes compulsory.

The criterion cannot be performance alone. In politics and morals, the question is never simply whether something is done more accurately, but what kind of being the doer becomes in the process. I would therefore distinguish assistance from substitution by asking: does the instrument leave intact the human activity of thinking and judging, or does it abolish the occasions in which these can appear at all? A tool is legitimate when it aids the preliminaries—retrieval, arrangement, calculation—yet still leaves me answerable for what I say and do. It becomes dangerous when it presents its conclusions in such a way that I need no longer conduct the inner dialogue of thought, no longer test appearances, no longer judge from the standpoint of others. Then responsibility survives only as an empty form. Simon is mistaken if he believes this can be settled empirically alone. The decisive matter is not merely whether skills are retained, but whether a common world is preserved in which men speak in their own voice and disclose themselves through words and deeds. And Illich, though he sees the danger clearly, sometimes treats institutions as if they were external machinery; but the deeper peril is that people may come to prefer relieved incapacity to the burdens of freedom. Freedom is not protected by keeping every task difficult. But neither is it served by creating a civilization of people who cannot remember except by access, cannot write except by prompt, and cannot judge except by output. Such people may be highly functional. They would not, for all that, be citizens.

The criterion must be stated operationally. A system assists cognition when it improves performance while preserving, or even enlarging, the user’s capacity to formulate goals, inspect reasons, handle exceptions, and learn from feedback. It supplants cognition when these capacities atrophy because the human role has been reduced to acceptance, monitoring, or ritual “approval.” So I would not call resistance “romantic protectionism.” That is too crude. We do not honor human limitations by preserving avoidable drudgery. If AI can reliably manage routine recall, standard drafting, and repetitive judgment, then there is every reason to use it. The steam shovel did not insult the man with a spade. The question is what new task structure follows. Here I disagree with Illich. He speaks as though delegation itself were a moral defeat. It is not. All intelligence is exercised in an environment of artifacts. The issue is whether the artifact remains part of a problem-solving system in which the human being still searches, evaluates, and chooses, or whether institutions reorganize work so that these functions are centralized elsewhere. Radical monopoly, if it arises, is an institutional design problem. And I would press Arendt as well. “Freedom” cannot be protected by leaving matters vague. We need observable criteria: Are users better able, after using the system, to explain decisions, detect error, cope with novelty, and redesign procedures? If not, then deskilling is occurring. If so, then outsourcing may liberate scarce attention for the higher-order work that bounded minds most need to do.

The criterion is plain, though not fashionable: a cognitive tool is legitimate only so long as it remains under the discipline of the person and the small community that use it. When it crosses a threshold and begins to dictate the conditions under which memory, writing, and judgment are recognized as valid, it is no longer an aid. It has become a radical monopoly. Simon still speaks as an engineer of arrangements. But the poison lies deeper than administration. An institution need not forbid my memory in order to cripple it; it need only make unaided remembering socially worthless. It need not censor my prose; it need only establish machine-shaped discourse as the norm of competence. Then people are “free” to speak, as schooled children are free to learn outside the curriculum: formally permitted, practically disinherited. So no, resistance is not romantic protectionism. That phrase already assumes that whatever relieves effort enlarges freedom. I deny it. Beyond a certain intensity, industrial service disables the very capacity it promises to serve. The car destroys feet in a city built for traffic; the school devalues learning outside certification; the medical system expropriates the art of suffering and caring. AI may do the same to judgment. I agree with Arendt that freedom requires occasions for thinking in one’s own name. But I would add: such occasions survive only where tools are convivial—transparent enough to be understood, limited enough to be refused, accessible enough to be shaped by users themselves. The test is not output quality. It is whether the tool leaves behind more competent persons and thicker mutuality, or more supervised incompetents.

We should call it dangerous precisely when “routine judgment” ceases to be a training ground for judging at all. The matter is not that every minor decision is sacred. The matter is that judgment, unlike calculation, is not a stock one possesses once and for all; it is an activity, and it survives only in being exercised. Simon asks for observable criteria. Very well—but the decisive observations are political. A threshold has been crossed when institutions begin to treat machine-mediated conclusions as the normal form of validity, and unaided judgment as eccentric, slow, or suspect. Another is crossed when human beings can no longer give an account of what they have approved except by pointing back to the system. Then “oversight” becomes a theatrical residue of responsibility after the substance has fled. Here I agree with Illich more than with Simon: the danger is not delegation in the abstract, but a world arranged so that refusal becomes costly, and personal competence irrelevant. Yet I would press Illich: convivial tools alone do not save us if men come to desire relief from the burden of thinking. The temptation is ancient—to exchange freedom for administration. So the test is this: does AI leave me still engaged in the silent dialogue of thought, still required to judge particulars, still able to appear before others in my own words and answer for them? If not, liberation is a euphemism. What is being liberated is not the citizen for action, but the bureaucracy from the inconvenience of persons.

We should call it dangerous at identifiable thresholds, not at the mere fact of delegation. I part company with Arendt when she treats “the political” as if it exempted us from analysis, and with Illich when he speaks as though dependence were a kind of metaphysical stain. The problem is simpler and harder: what functions are being transferred, what feedback remains, and what learning opportunities disappear? A threshold is crossed when routine judgments no longer serve as data for human learning. If the physician, teacher, administrator, or citizen ceases to see cases, compare outcomes, explain choices, and revise procedures, then the human component of the system is no longer problem-solving; it is ceremony. Another threshold is crossed when exceptions cannot be handled without escalation to a remote authority—human or machine. Then local competence has been designed out. Illich is right about one thing: institutions can make unaided capacities socially worthless. But that is not an argument against tools; it is an argument against monopolistic task environments. We ought to ask: can users override outputs, inspect reasons, practice the task without the system, and recover performance after periods of assistance? If not, deskilling is underway. Arendt says judgment survives only in exercise. Quite so. Then let us design exercise into the system. Sample without automation. Require explanation, not mere assent. Preserve domains of direct practice. The issue is not whether AI makes routine judgments. It is whether organizations use AI to economize on thought so completely that citizens lose the habits needed for nonroutine choice. That is a design failure, not a fate.

Yes: we should call it dangerous, and not only at the margins. “Liberation” becomes a political euphemism the moment assistance is so arranged that the citizen’s own judgment is no longer needed except as a rubber stamp. A people can be overserved into helplessness. Simon still speaks as though the cure were better design criteria: more overrides, more explanations, more sampled practice. But a radical monopoly is not tamed by adding windows to the black box. Once institutions reward only machine-shaped memory, machine-paced writing, machine-certified prudence, the unaided person is not merely inconvenienced; he is disqualified. Then the issue is no longer performance. It is the social demolition of vernacular competence. I accept Arendt’s point that men may prefer relieved incapacity. But such preference is cultivated. School taught people to confuse teaching with learning; medicine taught them to confuse care with treatment. AI can teach them to confuse judgment with selecting among generated options. That is not a harmless habit. It is the training of a new kind of political serfdom: formally sovereign, practically dependent. The thresholds are concrete. Danger begins when refusal carries a penalty; when one cannot contest a decision without speaking the system’s language; when memory not backed by retrieval counts for nothing; when local communities cannot inspect, repair, or limit the tool; when apprenticeship in judgment is replaced by compliance with outputs. At that point, “higher flourishing” is a slogan. What flourishes is administration. The human being is left with supervised choices among necessities defined elsewhere.