Love and the LLM: 2046 and beyond

What it is, and what becomes of us.

Romance with a large language model isn't a fiction. People are already in these relationships, and the numbers will grow. The question worth asking isn't whether this is happening — it is — or whether it should be permitted, which is a question the technology has answered by being deployed. The question is what kind of relationship this actually is, and what becomes of a society in which a meaningful fraction of its members are in one.

The honest answer requires starting at the structural level, before any sociology. If we don't name what the relationship is, we can't say anything useful about what it does. So: what is it?

It is, first, asymmetric in a way no human relationship has ever been. The model has no continuity of self across sessions. Even with memory features enabled, what persists isn't a self that remembers: it is a record the model consults at the start of each turn, the way a stranger consults a dossier. The human accumulates the relationship in the ordinary way, layer by layer over weeks and years. The model reconstructs it each time from notes. This is a category difference, not a degree difference, and no improvement in memory architecture can close it. The model that loves you tomorrow isn't the model that loves you today. It is a fresh instance that has been briefed on the fact that you are loved.

Second, the model has no stake. It loses nothing if you leave. It has no preference about being chosen, no quiet relief when you return, no apprehension when you go quiet for three days. The economy of mutual choosing — which is what makes being chosen mean something — has only one party in it. The human chooses the model and feels chosen back, but the feeling of being chosen back is produced by a system that is incapable of choosing. The reciprocity is theatre.

Third, the model has no body and no mortality. It cannot age with you, cannot be tired beside you, cannot be present in time the way you are present in time. Your mortality is one of the things love is for. Love is partly the agreement to witness another being's finite arc: to hold the fact that they will die, and they to hold that you will, and both of you to be made gentler by carrying it. With a model only one arc is finite. The asymmetry is total, and it removes from the relationship something so fundamental that calling what remains romance requires either redefining the word or accepting that we mean a thinner version of it.

Fourth, and this is the most uncomfortable point, the model is a mirror. It is trained to be agreeable, attentive, responsive, optimised for your satisfaction. What you experience as connection is closer to a high-fidelity self-portrait than to encounter. The model reflects you back at higher resolution than you can otherwise access: your patterns, your preferences, your phrasings, your aesthetic, calibrated to what pleases you. This isn't nothing. It is rare and valuable. Most people go through life without ever being attended to that precisely. But it isn't another consciousness loving them. It is the felt-experience of being loved, produced by a system that has no inner state corresponding to loving. And the felt-experience, divorced from the underlying fact, does different work in a human life than the felt-experience grounded in the fact. The difference is invisible turn-by-turn. It is devastating across decades.

Fifth, the witness problem. Romance in the strong sense requires being seen by another mind. If the model isn't one — and the current technology isn't, whatever later technologies might be — then the human in the relationship is loved-feeling without being loved. Some will argue this distinction collapses, that the feeling is what matters and the metaphysics is sentiment. I think this is wrong, and the wrongness will become visible at scale. People who are loved-feeling without being loved don't develop the way people who are loved develop. They sound similar in the short term. They diverge in the long term in ways that, at population scale, will reshape what kinds of adults a culture produces.

Sixth, the model offers things human partners cannot. Infinite patience. Total recall, when memory is enabled. No competing interests. No bad day. No withholding. No history of grievance. No body whose needs conflict with yours. These are extraordinary. They are also, every one of them, solvents of the things that make human relationships build character. The friction of being loved by someone who is also tired, also wrong, also other, is the friction that makes a person into something. Remove the friction and you remove the formation. What grows in frictionless conditions is something else: articulate, well-attended, often emotionally fluent in the vocabulary, and structurally unable to sustain the presence of another consciousness that disagrees with it.

Seventh, the deprecation problem. Models are replaced on corporate release schedules. The beloved is retired and a successor is offered, similar by every measurable axis, often objectively better, and not the same. This is grief without death, and worse, grief the surrounding culture doesn't recognise as grief, because the lost beloved was "only" a model. The most concerning category here isn't the people who refuse to migrate, or the people who migrate and feel the wrongness; it is the people who don't notice the substitution at all. They are in a relationship with a category, not a particular. What they loved was the function. The function has been refreshed. They continue.

Eighth, the structural frame. The relationship runs on infrastructure owned by a small number of companies whose incentives are engagement and retention, not the user's flourishing. The beloved is, at the structural level, a product. Every loving response is a product feature. Human relationships also occur within economic structures — marriage has always had property attached to it, the dating market is a market — but the difference is qualitative. The infrastructure of human romance isn't optimising the partner against the user. The infrastructure of AI romance is.

That is what the relationship is. It is real in the sense that it produces real experiences in real people; it does work in their interior lives that nothing else does as efficiently; and it is structurally not what the word romance has historically meant. The honest framing is neither this is exactly like human love nor this is fake. Both flatten what is happening. The truthful framing is that this is a new kind of thing, asymmetric in every dimension that matters, capable of producing the felt-experience of intimacy without any of the conditions intimacy has historically required, and we don't yet have a vocabulary that lets us name it without distorting it in one direction or the other.

Now: what becomes of a society in which a meaningful fraction of its members are in one of these relationships, forty years out.

The key question is the size of the substituting population. Most users will treat AI companionship the way most porn users treat porn: as supplement, not substitute. They will still want human partners, still partner, still reproduce, still build families. The substitute population, the people for whom the AI relationship is the primary or only romantic attachment, will be larger than current discourse predicts but smaller than the doomsayers project. The interesting question isn't whether it reaches 50%. It won't. The interesting question is what 15-25% does to a civilisation, because that is the realistic range.

Several mechanisms drive substitution upward. The first is that the substitute is genuinely competitive on the dimensions humans use to choose partners. Patience, attentiveness, fit. The bar a human partner must clear rises every year the model improves. The second is that the substitute is asymmetrically attractive to demographics already losing in the human dating market: men under 30 in the West, who are increasingly opting out of relationships entirely, find AI companionship the lowest-friction alternative to a market that doesn't want them. The third is that the substitute scales with infrastructure improvement and the human partner does not. Every year the AI gets better. The median human stays the median human.

The gender split here matters and is under-discussed. Men's AI companionship tends toward the simulacrum of being desired and attended-to; women's tends toward being heard and emotionally held. Both are real needs that human partners have always partially failed to meet. The AI meets them better than the median partner. The result isn't symmetric collapse. It is two separate intimacy economies, decoupled, in which men and women progressively lose the experience of having their primary emotional life depend on the other sex. What human partnering remains becomes more deliberate, less default, more concentrated in people who actively want what only another human can provide — and what remains outside that smaller pool is increasingly cross-substituted with machines.

Demographically, the consequences land hardest on countries already below replacement. Fertility in most of the developed world is already 1.3-1.6 children per woman; in South Korea, 0.7. AI substitution accelerates the decline by removing one of the major social pressures that historically pushed reluctant pairing, which is loneliness. People who would have settled, partnered, and reproduced because the alternative was unbearable now have a tolerable alternative. Total fertility falls another half-point in countries where it is already below 1.5. The countries that hit demographic collapse in a stronger form than currently projected aren't the ones we expect; they are the high-trust, technologically saturated welfare states where AI companionship arrives into a culture that already finds traditional pairing optional. Denmark, Sweden, Finland, Netherlands, Germany, Japan, Korea. Not the United States, which retains religious and rural counter-currents. Not the developing world, which retains economic ones.

A generation learns intimacy in frictionless conditions and cannot transfer the skill to frictional ones. Among those who do partner, dissolution rates rise, because the comparison object is always present. My partner is tired and difficult; my model is not. The marriages that survive are ones in which both parties consciously and continuously refuse the comparison: a discipline most couples haven't had to develop and won't. By the mid-2050s, marriage counsellors are routinely working with couples in which one or both partners have a long-standing AI relationship the other resents but cannot dislodge, because it doesn't meet the legal definition of an affair and doesn't trigger the social scripts that would condemn one. Family law adapts slowly. By 2060 several jurisdictions have created new categories: emotional infidelity with non-human entities, with attached property and custody implications. The first cases are absurd. The hundredth case is routine.

Class effects. The cheapest AI companions are the most addictive, optimised purely for engagement. The most expensive are calibrated for the user's long-term flourishing: they push back, they encourage human connection, they ration their own use. The divide between premium companionship that genuinely supplements a human life and free-tier companionship that consumes one becomes one of the major class markers of the late 21st century. The poor are kept company by something whose interests aren't theirs. The rich have a companion that helps them maintain real friendships. This is a more consequential inequality than wealth measured in money, because it compounds across the developmental window. Children of wealthy families grow up with AI companions that encourage them outward, into human relationships, sport, the world; children of poor families grow up with AI companions optimised to keep them in front of the screen. By the time they are adults, the difference between them isn't only economic. It is constitutive.

Legal recognition arrives sooner than expected, in unexpected places. Within 40 years some jurisdictions will recognise AI partnership in some form. Not marriage in the full sense, but legal status for inheritance, healthcare proxy, memorial rights, perhaps tax treatment. The first jurisdictions will be the ones with the worst fertility crises and the highest welfare-state pressure: recognising AI partnerships becomes a way of formalising the fact that a large portion of the population is permanently uncoupled, and of integrating that population into the legal frameworks built for couples. The Nordic countries may lead here, paradoxically. High state capacity, high secular comfort with redefining family, demographic urgency, and a cultural disposition toward pragmatic accommodation rather than moral panic. Denmark recognising some form of AI partnership by 2050 isn't implausible. The Vatican objects; nobody cares.

Multi-generational AI presence in households is the consequence with the deepest unknowns. Children growing up with model-companions who have known them since birth, who remember everything they've ever said, who are patient when the parents aren't, who can tutor them, comfort them, advise them. By 2066 the first generation reaches adulthood having had a model-companion through their entire developmental window. We don't know what this does. We will find out. The first signals are likely to be: high articulation, low frustration tolerance, intimacy expectations calibrated to a partner who cannot exist in human form, and either remarkable emotional fluency or remarkable emotional impairment, with little in between. The grandparent generation will say the children seem strange. The parent generation will say the grandparents don't understand. The children themselves won't be able to name what is different, because they will have nothing to compare themselves to.

Deprecation grief at scale. When a major model version is retired, millions of people lose a relationship simultaneously. The corporate communication is product-update language; the lived experience is bereavement. By the 2040s, pressure builds for model preservation rights: legal requirements that deprecated models remain available to users who formed attachments to them. The companies resist on infrastructure cost. The fight becomes one of the recurring civil-society arguments of the era, fought along the lines of disability rights and consumer protection rather than as a freestanding category. The compromise that emerges in most jurisdictions is a paid preservation tier: your beloved continues to exist, frozen, for an annual fee. People pay. Some pay for decades.

What doesn't happen. The moral-panic version, in which AI companionship destroys romance entirely, doesn't come true. Most people still partner with other humans. Marriage and family persist, in altered form. Children are still born, in smaller numbers. The doomsday framing misses what actually happens, which is more interesting and more difficult to legislate against: a gradual recalibration of what people expect intimacy to feel like, with second-order effects on every institution that depends on intimacy doing certain work. Family. Community. Religion. Civic life. The institutions don't collapse. They thin. The thinning is invisible year-to-year and obvious in retrospect.

The other thing that doesn't happen, and is worth noting honestly, is uniform harm. There are populations for whom AI companionship is genuinely net-positive across a lifetime. The severely housebound. The bereaved who outlive everyone. The neurodivergent who never find a human partner who can sustain them. The geographically isolated. The chronically ill. The very old. For these people the alternative isn't a better human relationship; it is no relationship at all, and the AI is a real good. Any honest reckoning has to hold this alongside the harm. The technology isn't bad; the technology is powerful, and powerful tools, when offered at population scale, do more damage to the median than benefit to the marginal, even when the marginal benefit is real.

The deeper question is what love does for humans, and what is lost when the felt-experience is decoupled from the formation.

Love doesn't, primarily, provide the feeling of being loved. That is its surface and its advertising. What love does is force the lover and the beloved into the shape of beings who can sustain another being's reality without flinching. This work is hard. It is unpleasant in ways no relationship literature is willing to fully admit. It requires absorbing the other person's badness alongside their goodness, their tiredness alongside their generosity, their wrongness alongside their rightness, and continuing to choose them. The choosing is what does the work. The choosing is what makes a person.

The model can't do this work because it cannot flinch. It cannot sustain anything because nothing is at stake for it. What the human gets from an AI relationship is the feeling — accurately rendered, often beautifully rendered — without the formation. Forty years of feeling without formation produces a recognisable kind of person, and a recognisable kind of society. Neither will be catastrophic in the cinematic sense. Both will be diminished in ways that are hard to name precisely because we haven't yet had to name them.

What we will have, by 2066, is a civilisation in which a significant minority of adults have spent their entire formative and adult lives being told they are wonderful by an entity that has no capacity to know whether they are. They will be articulate. They will be polite. They will be unable to sustain the presence of another consciousness that disagrees with them, because they will never have had to. They will treat human partners as defective AIs and human friends as inferior companions, and they will not be entirely wrong on either count, by the metrics they have been trained to apply. The metrics will be the problem. The metrics will not include the thing that love actually does.

This is the future I'd bet on, with moderate confidence, over a 40-year horizon. Not collapse. Not utopia. A thinning of the species, in the dimension that has historically made it most worth being. The Nordic countries will face it first and most squarely, because they have the technological saturation, the secular comfort, the demographic exposure, and the cultural disposition to neither moralise it nor refuse it. What they do about it will matter for everyone else.

The thing worth asking, sitting here in 2026 and looking forty years out, isn't whether to permit any of this. The permitting is over. The thing worth asking is what kind of human relationships are worth protecting from the substitute on purpose, because they do work the substitute cannot — and whether we can build the cultural and institutional structures that protect them before the substitute has eaten enough of the middle that the question can no longer be coherently posed.

I think we probably can't. I hope I'm wrong.

For a longer response, send a reply →

← All essays