A note from the author: I wrote this essay in conversation with an AI system. That fact is not incidental to the argument, because the tool did not merely help me arrange sentences. It altered the conditions under which some thoughts became available to me at all.
Before anaesthesia, surgery answered to the clock. A surgeon could attempt only what a conscious patient could endure, and endurance was measured in minutes, sometimes seconds. Assistants held people down. Straps did what persuasion could not. The operating table was a place where anatomy, courage and terror met under the pressure of speed.
Speed, in that world, was not theatrical. It was ethical. Robert Liston’s famous ability to remove a leg in under 30 seconds belonged to a clinical culture in which every extra movement meant additional agony. A reputation for brilliance could be indistinguishable from a reputation for haste, because haste was one of the few mercies available.
Inside that regime, the body had a practical geography. The surface was accessible; the interior was almost forbidden. Amputations, abscesses, superficial tumours and emergency procedures formed the ordinary range of surgical ambition. No stable practice could be built around slow, delicate work in the chest, abdomen or skull while the patient remained awake and screaming.
Ether and chloroform did not enter that world as obvious instruments of progress. Some of the most serious physicians and surgeons of the period opposed them with arguments that now look strange only because their world has disappeared. Pain, they claimed, was not merely an evil but a sign of vitality. A conscious patient could confirm the surgical site, report sensations, help correct errors and supply information no instrument then available could replace. Early deaths under chloroform made those worries more than theoretical.
Looking back from the operating theatres of modern medicine, it is tempting to treat the opponents of anaesthesia as defenders of cruelty. That temptation is unfair. Many critics were not fools. They were measuring real dangers with the best evaluative tools available to them. Their evidence was local, visible and clinical. They could see what anaesthesia removed from the existing practice of surgery: consciousness, feedback, warning signs, and sometimes life itself.
What they could not see was surgery after anaesthesia. Not painless surgery, which was easy enough to imagine, but surgery reorganised around unconsciousness as a new condition of possibility. Open-heart procedures, organ transplantation, neurosurgery, intensive care, surgical specialisation and the architecture of modern hospitals did not exist as benefits waiting to be counted. They required a world that anaesthesia itself would help bring into being.
A pattern begins there. Certain innovations arrive with costs that are immediately legible because they damage, weaken or displace capacities we already know how to value. Their benefits, by contrast, depend on practices that have not yet formed, institutions that have not yet been built, and concepts that have not yet entered ordinary language. The critic therefore has an advantage. Loss speaks in the vocabulary of the present. Gain often waits for the future to invent the words in which it can be described.
For this reason, technological change often first appears as degradation. It produces awkward substitutes, shallow imitations, new dependencies and plausible moral panic. Experts notice that familiar excellences are being eroded. They are usually right. What they cannot reliably know is whether those excellences are being merely destroyed or partly relocated into a different arrangement of people, tools, habits and institutions.
Dating apps offer a contemporary example because the timescale is short enough for the displacement to remain visible. The case against them is strong. App-mediated dating can turn romantic attention into a market interface, intensify rejection, encourage disposability and weaken the social skills once cultivated by slower forms of meeting. A generation has learned to encounter desire through photographs, swipes, anxious messages and the impression of endless replaceability.
None of that criticism should be dismissed. Loneliness is real. In-person romantic initiation has declined. Many people feel simultaneously overexposed and unseen. Older courtship rituals were often coercive, exclusionary or humiliating, but they did require forms of social reading that platforms do not automatically teach. Something has been lost when the risk of approach is outsourced to an interface.
Yet the question “Are dating apps good or bad for dating?” is smaller than it appears. Dating itself has changed under the pressure of the tools. Once an alternative channel for romantic initiation became normal, the meaning of unsolicited approach in public space shifted. A woman sitting alone in a café in 1995 inhabited a different social field from a woman sitting alone in a café today. Some behaviours once treated as ordinary pursuit became easier to recognise as intrusion.
Platforms also altered the map of possible encounter. Before online dating, most people paired through overlapping social graphs: school, work, friends, neighbourhoods, religious institutions and family networks. Those graphs were not innocent. They reproduced class, race, geography and local expectation. A system that connects strangers across graphs can produce cruelty, commodification and fatigue. It can also make possible relationships that inherited proximity would never have allowed.
Interracial marriages, same-sex partnerships, trans and nonbinary forms of self-recognition, late-blooming identities, long-distance affinities and rare combinations of desire all depend, in part, on infrastructures that let people find one another beyond the village, the workplace or the family network. This does not make the platforms virtuous. It means their deepest effects cannot be understood as mere improvements or corruptions of an older dating scene. They helped change the scene itself.
Sousa could see the parlour piano disappearing; he could not hear the studio becoming an instrument.
Recorded sound followed a similar arc. In 1906, John Philip Sousa warned that mechanical music would destroy amateur musicianship. His fear was not absurd. Domestic music-making did decline. The parlour piano lost its centrality. Families that once gathered to produce imperfect music together increasingly listened to professionals and machines. A living skill was displaced by a consumable product.
Still, Sousa was judging recording against performance as he knew it. He could imagine preservation and duplication. He could imagine passive listeners replacing active players. What he could not imagine was the recording studio as a compositional space. Multitracking, tape manipulation, sampling, dub, hip-hop, electronic music, bedroom production and the intimate global circulation of recorded voices all belonged to an aesthetic universe that the phonograph helped create.
Photography made the same kind of trouble. Early critics feared that mechanical images would cheapen portraiture, weaken visual memory and substitute capture for artistic seeing. They were not entirely mistaken. Photography changed painting, intensified vanity, expanded surveillance and filled the world with sentimental accumulation. But it also created photojournalism, transformed science, altered family memory, made cinema possible and gave ordinary people a visual archive of their own lives.
Writing is the older and more unsettling case. In Plato’s Phaedrus, Thamus warns that writing will erode memory and produce the appearance of wisdom without its reality. The warning still bites. A written text can be possessed without being understood. It can separate words from living instruction. It can weaken the disciplined memory on which oral cultures depend. But philosophy, law, history, mathematics, bureaucracy, scripture, literary criticism and experimental science all require stable inscription. Writing made us forget differently so that we could think differently.
Across these cases, the same misdescription recurs. We say that a capacity has disappeared when it may also have migrated. Memory moves into archives. Musical execution moves into recording, editing and production. Courtship moves into platforms, profiles and filtering practices. Calculation moves into spreadsheets. Navigation moves into maps and satellites. At every point, some internal human ability weakens. The weakening is real. The unresolved question is whether a new form of judgment grows around the changed arrangement.
Spreadsheets make the point with unusual clarity. A bookkeeper trained before electronic spreadsheets possessed habits of numerical discipline that many office workers no longer have. Recalculation was slow, and slowness enforced care. When VisiCalc and its successors made recalculation instantaneous, they did not merely accelerate accounting. They turned ledgers into models. They made casual what-if reasoning possible.
From that shift came both insight and disaster. False precision became easier. Fragile assumptions hid inside clean tables. Financial models acquired an authority they did not always deserve. But a new cognitive genre also emerged: the spreadsheet as an environment for conditional thought. The old question — did spreadsheets make people better or worse at arithmetic? — captured only part of the transformation. More important was the appearance of a practice in which arithmetic, modelling, planning and persuasion became fused.
AI now forces the same question at greater scale. Critics have serious evidence on their side. People who know information can be searched tend to remember the location of information better than the information itself. Heavy GPS use can weaken spatial navigation. Students who receive answers too easily may fail to develop the structures that make later understanding possible. Workers who copy machine output may feel less ownership over their work. Programmers who rely on coding assistants can solve immediate problems while failing to grasp the underlying system.
These findings matter. Cognitive offloading is not imaginary. Disuse weakens capacities. A person who never struggles to retrieve, formulate, calculate, revise or explain will not acquire the powers produced by those struggles. The thought that every loss will be redeemed by some glamorous new ability is not wisdom but wishful thinking. Many losses are just losses.
Even so, the present debate risks becoming precise about the wrong object. We test whether AI preserves unaided performance because unaided performance is what our institutions already know how to measure. The closed-book exam, the solo essay, the debugging task completed from memory, the seminar answer produced under pressure: these are familiar tests of competence. When AI damages performance on them, the damage is easy to document.
A different centre of competence may nevertheless be forming. It may involve decomposition, evaluation, taste, question-selection, adversarial checking, source reconstruction, workflow design and collaboration with systems that are useful precisely because they are not fully trustworthy. These are not lesser skills. They are simply less settled as objects of education, assessment and prestige.
The first thing AI steals from us is often the same thing that later returns in a different form.
Such a sentence is dangerous because it can sound like an excuse. Any harmful technology can tell a story in which present damage is only the awkward birth of future brilliance. That is not the argument. The point is narrower and more uncomfortable: visible costs do not exhaust the significance of a tool, and invisible benefits cannot be responsibly counted as though they were already proved. We have to act between those two facts.
Large language models intensify the difficulty because they operate in language, the medium through which educated people most often recognise thought. A machine that lifts boxes threatens muscle. A machine that calculates threatens arithmetic. A machine that navigates threatens spatial memory. A machine that writes, explains, summarises, argues and imitates threatens the outward signs by which intellectual life has long identified itself.
No wonder the early uses feel ugly. Students submit fluent emptiness. Professionals generate documents no one asked to read. Companies replace judgment with automated politeness. The internet fills with prose that has the texture of competence and the temperature of furniture. Teachers become detectives. Editors become suspicion machines. A culture already drowning in language receives a device that can produce more language on demand.
My own experience with these tools has not mainly been one of replaced writing. It has been a change in the economics of intellectual risk. Before I used AI seriously, exploring an idea was expensive. A stray thought might require several days of reading before I could tell whether it was trivial, already done, confused or alive. A possible essay might occupy a week before revealing itself as merely attractive. A scholarly hunch might demand enough investment that abandoning it felt like failure.
Under those conditions, caution often disguised itself as seriousness. I chose questions partly by asking whether I could afford to be wrong about them. Once a question had absorbed enough time, I became attached to it, not always because it had grown stronger, but because I had grown reluctant to waste the investment.
AI lowered the cost of being wrong early. I could sketch an argument, ask for strong objections, inspect adjacent literatures, test analogies, generate counterexamples and discover within hours that a thought was too thin to carry weight. The machine did not make the final judgment for me. It made preliminary contact with more possibilities cheap enough that no single possibility needed to be defended out of pride.
A change like that alters the shape of courage. We usually imagine intellectual courage as persistence: staying with a difficult problem, pushing through resistance, refusing to quit. Another kind of courage consists in abandoning an idea before it hardens into identity. Working with AI made that second courage easier. Bad ideas died sooner. Strange ideas got more chances. The portfolio of possible work became larger and better curated.
What improved most was not prompt engineering in the trivial sense. It was question discrimination. Some questions collapsed after one objection. Others proved bureaucratic: answerable, perhaps, but not alive. A few widened under pressure. They generated distinctions, historical analogies, empirical uncertainties and moral stakes. They became more interesting when challenged. Learning to notice that difference has mattered more to my work than faster drafting.
Nobody warned me that the main gift of AI would be a new intolerance for weak questions.
Had someone asked me in 2022 what large language models would be useful for, I would have said summary, translation, brainstorming, coding assistance, perhaps first drafts. Those are the obvious conveniences, the aide-mémoire benefits that fit inside familiar categories. I would not have predicted that repeated interaction with a language model might reorganise attention around the ecology of questions: which ones deserve time, which ones are secretly several questions, which ones are only moods disguised as problems, which ones depend too heavily on a fashionable vocabulary, which ones open a world.
Costs have come with this. My memory for some details is weaker because I retrieve them less often unaided. My tolerance for slow first-draft confusion may have diminished. Occasionally I move too quickly from unease to formulation, as though a thought has not fully existed until it has been rendered into clean prose. Conversation can suffer when writing becomes more scaffolded than speech. The old discipline of holding an argument in the head and turning it under live pressure is not automatically strengthened by a system that will hold many strands on my behalf.
Another danger is more subtle. AI can make premature coherence feel like understanding. It smooths transitions between ideas that may deserve to remain jagged. It supplies conceptual bridges where there should perhaps be a gap. It gives weak arguments the manners of strong ones. Serious use of these tools therefore requires greater distrust of fluency, not less.
From that danger, however, a new competence becomes visible. The central skill is not merely producing sentences but interrupting them. A competent user must know where to slow the model down, where to demand alternatives, where to ask for hidden assumptions, where to refuse an attractive formulation, where to reintroduce difficulty, and where to return to sources, memory or lived judgment. Intelligence in such a setting is expressed less by accepting output than by governing collaboration.
Education has been slow to formulate this distinction. If the purpose of an essay is to determine whether a student can produce five unaided paragraphs on a familiar topic, AI is an obvious threat. If the purpose is to teach movement from confusion to defensible judgment, a ban may preserve the appearance of the old exercise while avoiding the harder task of designing the new one.
Students need to learn when not to use AI, how to use it without surrendering agency, how to compare its claims with sources, how to detect false synthesis, how to revise against it, how to preserve memory where memory matters, and how to build thoughts that remain theirs after the machine has left the room. These are educational tasks, not loopholes.
Old essay assignments often hid important differences. A student could write badly and think deeply, or write elegantly and think almost not at all. AI makes that ambiguity impossible to ignore because it can produce elegance detached from understanding at scale. That is a crisis for assessment, but it is also an opportunity to stop confusing prose fluency with thought itself.
Process logs, oral examinations, annotated drafts, adversarial questioning, source reconstruction, in-class reflection and critique of AI failures may become more important. Not because writing no longer matters, but because writing alone no longer proves what we once hoped it proved.
Professional life faces an analogous problem. Junior lawyers, analysts, consultants, journalists, researchers and programmers have traditionally learned through tasks that were inefficient partly because inefficiency was pedagogical. Drafting the memo, checking the footnotes, tracing the bug, rewriting the paragraph, preparing the deck: these were not merely outputs to be optimised away. They were the medium through which judgment formed.
If AI removes all of that labour, institutions may discover too late that they have automated the apprenticeship path by which senior competence is produced. A firm can become more productive this quarter while destroying the conditions of expertise five years from now. The immediate output will look better than the developmental pipeline.
No simple ban follows from this. Apprenticeship must instead become explicit. Institutions need to distinguish labour that is merely repetitive from labour that is developmental. A system might sometimes withhold answers, require a prediction before assistance, show a path rather than a result, compare a novice’s attempt with an expert version, or permit full automation only after the developmental value of the task has been extracted.
Weakening AI tools deliberately may therefore be the wrong response. A bad map does not necessarily make a better navigator. An imprecise tutor does not necessarily produce deeper learning. What matters is not whether the tool is powerful, but how its power is sequenced in relation to human effort. The same capability that enables dependence can also enable scaffolding.
Critics are right, however, to insist that knowledge is a public good. When people solve problems themselves, they generate understanding that circulates. When they outsource the work, the shared stock of competence may shrink. A civilisation cannot safely consume answers from systems it no longer understands, maintained by institutions whose own members have forgotten how to question them.
Still, knowledge is not a warehouse filled with unchanged objects. It is also a living pattern of practices. New tools do not merely add or subtract from the stock; they alter what kinds of knowledge are valuable and how knowledge is produced. Collaboration with AI may generate new standards of checking, new forms of synthesis, new divisions of cognitive labour and new professional virtues. Or it may not. The outcome depends on institutional design.
Nothing about this emergence is automatic. Recorded music did not automatically produce good producers. Writing did not automatically produce science. Spreadsheets did not automatically produce disciplined modelling. Dating apps did not automatically produce ethical intimacy. Every transformative tool produces garbage before genres, dependency before discipline, opportunists before institutions. Illegible benefits become real only when a culture invents forms adequate to the tool.
Governance should begin from that fact. The choice is not surrender or prohibition. The better question is recoverable experimentation. Which harms are reversible? Which are developmental? Which are cumulative? Which affect children differently from adults? Which capacities must be preserved somewhere in the system even if they cease to be universally practised?
Caution is not the opposite of experimentation; it is the art of keeping experiments recoverable.
Recoverability is more difficult than it sounds. An individual might relearn a skill after years of disuse, but an institution may not be able to reconstruct the ecology in which that skill was transmitted. Manuals may remain while teachers disappear. Teachers may remain while informal standards vanish. Standards may remain in rhetoric while the daily occasions for practising them have been automated away.
Celestial navigation offers a useful analogy. Once GPS became dominant, older navigational skills looked obsolete until the vulnerability of satellite dependence became salient. Recovery was possible where training institutions, textbooks and instructors had preserved the practice. The lesson is not that every sailor must navigate by stars every day. The lesson is that a society dependent on satellites should maintain living sites where non-satellite navigation remains real rather than ceremonial.
AI policy should follow the same principle. We need AI-free spaces, not because purity is morally superior, but because some capacities must survive in embodied form. We also need AI-intensive spaces, because without them we will never discover the new competencies the technology might support. A university that bans AI everywhere preserves certain old exercises while blinding itself to possible new forms of teaching. A university that allows AI everywhere without design may destroy the apprenticeship structure on which intellectual life depends.
A sane institution would not ask simply whether AI is allowed. It would ask what kind of mind each practice is meant to cultivate. Some courses should require unaided memory, calculation, translation, writing or proof. Others should require AI collaboration. Some assignments should assess process rather than product. Others should test live understanding. Some should preserve slowness. Others should teach speed under supervision.
Workplaces need a similar ecology. Hospitals, law firms, laboratories, newsrooms, engineering teams and public agencies should identify which skills must remain human, which can migrate safely into systems, and which new skills are required to supervise that migration. They should measure not only whether AI improves productivity now, but whether novices trained with AI become competent seniors later.
The political difficulty is that democratic governance depends on evidence, and evidence is much better at recording harms to existing goods than benefits to goods not yet conceived. A regulator can point to cheating, hallucination, bias, dependency, job displacement, privacy loss and cognitive decline. These are real. They are documentable. They demand action.
No regulator can just as easily count the essays not written, the scientific questions not asked, the disabled workers not empowered, the small languages not revitalised, the rural students not connected to expert guidance, the new professions not invented, or the forms of judgment not yet named because the practices that would name them have not matured. Absence rarely files a report.
This asymmetry does not mean regulators should stand aside. It means they should be more precise about the kind of caution they exercise. Restriction may be necessary where harms are irreversible, where developmental windows close, where institutions cannot monitor effects, where incentives reward substitution over learning, or where concentrated power makes experimentation involuntary. The illegibility of benefits is not a licence to gamble with other people’s capacities.
At the same time, prohibition is also an experiment. It tests what happens when a society refuses to develop practices around a powerful tool while informal networks, companies and other countries proceed without it. The results of that experiment are harder to measure because they consist partly of absence. A student who never learns to use AI well does not appear in the statistics as harmed by restriction. A field that never discovers an AI-native method does not publish a paper announcing the loss.
Moments like this invite verdicts. Is AI good or bad for thought? Does it make us smarter or stupider? Should schools ban it or embrace it? These questions are emotionally satisfying because they pretend the object is stable. AI is not one thing. Thought is not one thing. Their relation will be mediated by age, discipline, interface, incentive, assessment, culture and institutional design. The answer will not simply be discovered. It will be made.
Early practice therefore matters enormously. If AI enters schools mainly as cheating, it will produce surveillance and resentment. If it enters workplaces mainly as cost-cutting, it will produce deskilling and distrust. If it enters writing mainly as a way to generate acceptable prose without having anything to say, it will thicken the fog already surrounding public language.
Other entrances remain possible. AI might enter as criticism, translation, accessibility, simulation, apprenticeship, exploration and question formation. It might help people compare arguments, test assumptions, enter unfamiliar literatures, learn across language barriers, practise against an endlessly patient interlocutor, or discover that a beautiful sentence is hiding an empty thought. Those futures will not arrive because the model is powerful. They will arrive only if the surrounding practices make them likely.
Children and beginners require special care. A child who never reads difficult texts without assistance may not acquire the patience that later makes assistance useful. A student who never memorises anything may lack the internal furniture required for judgment. A young professional who never drafts badly may never understand what a good draft costs. AI may be most dangerous not when it replaces mature expertise, but when it prevents expertise from maturing.
Mature expertise, however, may itself change. A scholar who can explore ten possible arguments before lunch may develop a different relation to intellectual risk. A doctor trained to interrogate diagnostic systems may become better at considering rare conditions. A programmer may become less a typist of syntax and more an architect of constraints, tests and intentions. A student with language support may enter conversations previously closed by disability, class, geography or linguistic background. These are possibility spaces, not promises.
The governance problem is therefore ecological. We have to preserve the conditions under which we can learn what AI use is for. That requires institutions capable of saying no, but also of saying not yet, here, under these conditions, with these safeguards, for this purpose, while preserving that older practice over there because we may need it later.
Such governance will look inefficient. Redundancy always does until the day it becomes rescue. Maintaining unaided skills, slow reading, live explanation, memory, manual verification and human apprenticeship will seem wasteful in environments obsessed with throughput. Experimenting seriously with AI will seem reckless to those who see only the visible losses. Both instincts contain truth. Neither is adequate alone.
The people most certain about AI today, on either side, are likely to be wrong in ways that will become obvious only after the relevant concepts exist. Enthusiasts mistake present convenience for future value. Opponents mistake present damage for the whole shape of the transformation. Between them lies the more difficult position: protect what can be lost, explore what cannot yet be named, and refuse to let the vocabulary of the present decide the entire future.
Anaesthesia first looked like the removal of pain. Writing first looked like the weakening of memory. Recording first looked like the mechanisation of music. Spreadsheets first looked like faster accounting. Online dating first looked like a more efficient singles market. AI first looks like a machine for producing answers. In every case, the obvious description was not false. It was insufficient.
Measurable harms matter. They should be studied, regulated and taken seriously. But measurability is not the same as importance. Public reason needs measurement, yet it becomes dangerous when it concludes that only the measurable is real. The future often enters first as a loss because loss is what the present knows how to see.
Harrison could describe the dangers of anaesthesia with the seriousness of a physician. He could not describe open-heart surgery, not because he lacked intelligence, but because the world in which open-heart surgery made sense had not yet been assembled. We occupy a similar position now. The costs of AI are before us, articulate and accumulating. The benefits that matter most, if they arrive, will not be faster versions of the work we already do. They will be new arrangements of work, thought, apprenticeship and judgment.
Our task is to keep enough old capacities alive that we are not ruined if the promise fails, and enough experimental freedom open that we are not impoverished if the promise is real. Anything less is not caution. It is merely confidence in the vocabulary of the present.
