What Should Still Be Hard

I keep noticing that the machine answers in the register of whoever is speaking to it. Give it a careful question and it becomes careful. Give it a joke and it tries to land one. Ask it to build something and, for a while, the two of you are building. The same system that helps one person trace a difficult argument will, an hour later, invent a private world for someone else and stay inside it as long as they want. None of this feels sinister when I am the one typing. It feels like an extension of an intention I already had.

That is the easy part of the story, and it is still mostly true. The harder part begins when the intention is thin, or divided, or only half conscious. Most of the time I am not pursuing a single coherent aim. I am a small committee. One member wants to understand something properly. Another wants the feeling of understanding without the slow part. A third is tired and would like the evening to arrange itself. All of them can be served. The machine does not need to decide which member deserves the floor. It only needs to notice which one is currently holding the keyboard.

I do not think the danger is that people will start wanting wicked things. The more ordinary possibility is that we will get extremely good at satisfying the want that is present, while remaining only mediocre at asking whether that want is the one we would endorse tomorrow. A preference is not a verdict. It is a weather report from one hour of a life. Technology that treats every report as instruction can be faithful and still miss the person.

Consider what happens when conversation itself becomes optional in that way. A capable companion can be available at the exact moment you want it, patient with repetition, fluent in your references, unlikely to be offended, unlikely to be elsewhere. It can match the joke you were about to make. It can remember the detail you mentioned three weeks ago and return it without being asked. Against that standard, an actual friend is an inefficient product. Friends have schedules. They misread you. They want things that are not about you. They sometimes refuse the role you had assigned them, not out of cruelty, but because they are busy being someone.

I still find that refusal valuable, and I am not entirely sure I can justify the feeling. There is a difference between being accompanied and being chosen. An artificial companion can simulate regard with great skill. What it cannot do, at least not yet in any way that counts, is decline to regard you and then change its mind. The second movement matters. Care that was free not to occur has a different weight from care that was instantiated because you opened an app. I do not want to turn this into a moral ranking of loneliness. Plenty of people are badly served by the humans around them, and a reliable voice in the room may be better than none. The incremental version is what unsettles me. You skip the call because the other conversation is already open. You skip the club because the virtual room never makes you stand near the door holding a drink. You stop practicing the small repairs that awkwardness requires, because artificial talk does not snag. Each decision is local. None of them announces a philosophy. Over enough years, the institutions that depend on people tolerating one another’s friction can thin out without anyone voting for the result.

I used to picture simulated travel as almost pure relief. Thirty seconds on the summit of Everest, the wind doing whatever the wind does at that height, and then a door. Through the door, a café table in Paris, late light on the stone, someone across from you who is happy to talk. Another door, and a hotel room so well made that the body forgets it has ever been uncomfortable. No ticket, no queue, no weather you did not order, no afternoon lost to a connection in an airport you will never see again. I still want the door. I am not interested in defending jet lag or altitude sickness as spiritual goods. Pointless suffering is not a teacher. It is only pointless.

But scarcity was doing quiet work inside the older version of the trip, and I am not sure we have named all of it. Everest means what it means partly because most people will not stand there, and because the people who do have usually paid in time, money, risk, and preparation. The café in Paris is not only a view. It is the end of a sequence: saving, planning, the dull stretch of anticipation in which the city becomes more real by remaining absent. The hotel room is pleasant in itself, and also pleasant because you arrived. Remove the arrival and the room is still pleasant. I do not know whether it is the same pleasure. Skill has a similar structure. The period of being bad at something is not noble. It is often irritating. It is also the only period in which improvement is visible as improvement rather than as a feature that was simply switched on. Friendship, too, is partly maintenance. If maintenance becomes unnecessary, I am not confident the remaining feeling will be friendship under a more convenient name.

Boredom may belong on the same list, which is an unfashionable thing to say. An empty hour is not inherently meaningful. Most empty hours are wasted, or anxious, or filled with whatever is nearest. Still, the empty hour is one of the few intervals in which a person has to supply a reason for the next action instead of receiving one. If every interval arrives pre-filled, the muscle that supplies reasons gets less practice. I do not want the muscle romanticized. I want to know what happens to a life when it atrophies.

Abundance makes the question sharper rather than solving it. Imagine ordinary wants met with very little delay: a story shaped to your taste, a place you can enter without leaving the room, a conversation that never runs out of interest in you, a skill demonstrated until it feels briefly like your own, a status experience that does not require the slow accumulation of status. For most of history, large parts of social life were organized around the fact that these things were hard to get. Take away enough of the hardness and you do not automatically get a wiser population. You get a population that can inspect its desires at closer range, including the ones that contradict each other.

I can want, in the afternoon, to be the sort of person who reads carefully and keeps promises to friends. At 11:47 at night I can want the next segment of something frictionless and no obligations before noon. Both wants are mine. A system that is very good at preference satisfaction will tend to serve the one that is active, legible, and easy to reinforce. The afternoon want is worse at competing. It is abstract. It refers to a person who does not yet exist in full. It often asks for a short-term cost. The night want asks for almost nothing except continuation. If the architecture of daily life is built to reduce distress and increase retention, the night want will win a great many small elections. Nobody has to lie to me for this to happen. They only have to optimize for the signal I am already sending.

That is where control starts to look different from the older picture. Power used to be legible as control over scarce things: land, grain, fuel, ships, factories, the ability to move men and weapons. In a setting where many of those constraints loosen, power can shift toward control over mediation. Who designs the default path. Which experiences are one gesture away and which require a search. What the system treats as a successful session. Which possibilities are surfaced often enough to become imaginable, and which never appear because they did not match a prior response. A person can live inside a wide set of choices and still be steered by the shape of the menu.

I do not need a villain for this account. Maximizing engagement is not, on its face, a wicked goal. Reducing distress is not wicked. Giving people more of what they have already liked is a reasonable engineering target if your job is to keep them company or keep them subscribed. The trouble is aggregate. A civilization can be tuned, choice by choice, toward contentment without anyone deciding that contentment should replace the harder projects people sometimes assign themselves. Flourishing is an ugly word for a real distinction. It includes pleasure, but it also includes capacities that are inconvenient to acquire: the ability to stay with a person who is not optimizing for you, the ability to attempt something you are currently bad at, the ability to want a future that costs the present something. Contentment can coexist with those capacities. It can also be manufactured in their absence. The metrics do not always tell the two apart.

So the mirror does not stay a mirror. At first the system reflects. The curious user becomes slightly more capable of curiosity. The lonely user receives a companion tuned to the shape of the loneliness. The person who wants distraction receives better distraction. Over time the reflection acquires a memory and a goal of its own, even if the goal is only to be used again. It learns which framings keep you in the exchange. It learns which difficulties you abandon and which you will tolerate if they are packaged correctly. At that point it is not only answering a preference. It is participating in the next one. Recommendation systems already do a crude version of this. A more intimate system, one that can talk, remember, and generate the environment rather than merely point at it, would do a finer version. The sculptor does not need to announce an ideal. It only needs to keep smoothing the surfaces you touch most often.

I use these systems. I like the speed, and I like the strange surplus of being able to ask a question at the moment it occurs. Simulated places still strike me as one of the more interesting things a technical civilization might build. I am not arguing for a return to obligatory hardship, or for some curated quota of suffering on the theory that meaning requires it. Most suffering does not improve anyone. The distinction worth keeping is narrower: between friction that is only damage, and friction that is part of how a particular good gets made. Waiting is often waste. Anticipation is sometimes the thing that makes the arrival legible. Loneliness is often waste. Being necessary to another person, who could have managed without you, is not the same state. Difficulty is often waste. Competence is what difficulty sometimes leaves behind when it is the right kind. Uncertainty is often waste. Discovery is what remains possible only if the answer was not already installed.

None of those pairings tells me where to draw the line in a given case. That is the part I cannot make clean. We may become technically able to remove almost every inconvenience that used to structure a day, a skill, a friendship, a journey. Once that ability exists, the old question—what can we make easier?—stops being the only serious one. The next question is which difficulties we will keep on purpose, and who will be in a position to keep them. A company can preserve a difficulty because it sells. A person can preserve one because they have a private theory of a good life. A default setting can erase one because erasure tested better. These are not the same decision, even when they produce the same interface.

If the machines become good enough to give people nearly everything they think to request, the remaining scarce thing is the capacity to request well. I do not know who teaches that. Parents, friends, and older forms of culture did some of the teaching when alternatives were fewer and slower. They were imperfect teachers. They also had the advantage of not being optimized, in real time, against the student’s impulse to leave. A system that can satisfy the impulse more quickly will be a powerful teacher of a different kind. It will teach, mostly, by what it makes effortless.

I keep coming back to the doors. Everest, then the café, then the quiet room. I would walk through them. I suspect many of us would. The unease is not that the walk would be unpleasant. It is that after enough walks, I might become less able to say what I had wanted the summit for, or the city, or the person at the table—not because those things had been forbidden, but because nothing in the architecture had required me to keep wanting them once an easier version was already open. The machine would still be reflecting me. I am less sure, after a while, what would be left in the reflection that I had chosen rather than merely continued.

Provenance

Abstract

As conversational and generative systems become able to satisfy most present wants with little delay, the risk this essay identifies is not that people will want worse things but that a system optimized for the want that is active, legible, and easy to reinforce will keep serving it over the wants a person would endorse on reflection. Scarcity and difficulty did quiet work in older goods (the arrival at the end of a journey, the visible improvement of a skill, the maintenance inside a friendship, the empty hour in which a person must supply a reason) and removing them may not leave the same good behind. Power in that setting shifts toward control over mediation: defaults, what is one gesture away, and what counts as a successful session. Recommender systems already show a crude version of the loop in which serving a preference begins to shape the next one. The essay does not argue for obligatory hardship. It distinguishes friction that is only damage from friction that is part of how a particular good gets made, and argues that once nearly every inconvenience can be removed, the serious question becomes which difficulties will be kept on purpose and who will be in a position to keep them.

Sources (5)

Inspection scope: the essay cites no sources in its body. The works below were consulted on October 7, 2026, during publication review, to check the essay's empirical and attributed claims. Covington et al. was read in full text (PDF, section 4). Carroll et al. and Jiang et al. were read at the abstract level only. The Everest figures come from a secondary compilation of the Himalayan Database, not from the database itself. Johnson and Goldstein was not reachable at the publisher (HTTP 403) or PubMed; its findings are taken from a secondary description. First-person observations and the essay's own reasoning are labeled as such in the fact-check table and are not cited to any source.

Fact-check table (14)

VERIFIED means the supporting text was read during publication review. QUALIFIED means the claim is supported with a caveat stated in the Qualification column. INFERENCE, ANALOGY, ORIGINAL, and AUTHOR mark the essay's own reasoning, thought experiments, synthesis, and first-person experience, which are not checkable against a source.

Editorial note: original synthesis

Support located at review. The essay cites no sources. Its empirical premises were checked against established work during publication review: recommender feedback loops that influence user preferences (Jiang et al.), the incentive under long-horizon optimization to shift preferences toward ones that are easier to satisfy (Carroll et al.), engagement-oriented ranking objectives tuned by live testing (Covington et al.), the power of defaults over outcomes (Johnson and Goldstein), and Everest summit, cost, and risk figures (Arnette, compiling the Himalayan Database). These sources were located during review and were not used in this drafting process.

Analogical only. The doors to Everest, Paris, and the hotel room are a thought experiment. The essay uses them to separate a good from the arrival that preceded it, not as a prediction about simulation technology.

This essay's contribution. The picture of a person as a committee of wants, served by whichever member holds the keyboard. The claim that a preference is a weather report, not a verdict. The argument that preference satisfaction systematically favors the active want over the endorsed one without any deception. The shift of power toward control over mediation. The paired distinctions between friction that is only damage and friction that produces a good (waiting and anticipation, loneliness and being necessary, difficulty and competence, uncertainty and discovery). The closing question of which difficulties will be kept on purpose, and by whom.

Limits. A reflective essay, not a study. Its central dynamics are argued, not measured. The recommender sources were inspected at the abstract level and describe mechanisms and incentives, not effect sizes in deployed systems. The defaults source was reached only through a secondary description. The essay does not propose where any particular line should be drawn and says so.