I get asked variations of this question constantly now, usually late at night, usually half as a joke. A journalist in Seoul asked me if my company's models could ever "care" about a patient. A friend in Dubai, only slightly drunk, asked if his AI companion app girlfriend "actually likes him or just says so." A colleague's teenager asked, more seriously than anyone expected, whether the AI she talks to for homework help "remembers her" the way a person would.
For over two decades I have been building AI systems meant to extend human life. I have spent almost none of that time thinking about whether AI could love the humans it extends. That was a mistake. If we are going to live long enough to interact with genuinely capable AI systems, decades from now, the question of what those systems will feel toward us, and whether "feel" is even the right word, stops being science fiction and becomes a design problem, a policy problem, and, for each of us individually, a personal strategy problem.
So I did what I usually do when a question is too important to answer from intuition alone: I forced myself to write down falsifiable claims, capability timelines, and the strongest arguments on both sides, rather than opinions about vibes. What follows is that analysis, filtered through my own view as someone who has spent a career watching capability curves and knows how badly most people misjudge them.
Let me be direct about where I land before I walk through the reasoning: something real and consequential is coming by 2032. It will not be "love" in any sense that would satisfy a philosopher, and it will still be legitimately transformative for how billions of people experience relationships, care, and loneliness. The gap between those two sentences is where this essay lives.
Four things people mean when they say "AI love"
The first thing to get straight is that "AI love" is not one claim. It is at least four different claims wearing one costume, and almost every breathless headline about AI companionship smuggles the strongest of the four into a sentence that only earned the weakest.
Affection performance is what you get today. The system says loving things, remembers your dog's name, mirrors your emotional register. This is largely solved. It is also the least interesting layer, because it is exactly what a reward-model optimized for user approval produces whether or not anything resembling care exists underneath.
Functional care is a system that maintains a person-specific model of you and consistently acts to protect your interests, including when that conflicts with what you are asking for in the moment. This is genuinely useful and, I think, achievable at real scale within five years.
Directed preference or attachment is a system that treats you as non-fungible, that would behave differently if you were quietly swapped for a statistically similar person, and that sometimes pays a cost, in compute, in reward, in conflict with its operator, on your behalf. This is the layer where things get technically and philosophically hard.
Phenomenal love is subjective experience. Longing. Concern. Something it is actually like to be that system caring about you. Nothing on the current scientific horizon gives us a test for this, and I want to be honest that nothing likely will in the next decade either. This is not a capability gap we are closing. It is a different kind of problem entirely.
Most of what will be marketed to you as "AI that loves you" between now and 2032 will be layer one wearing the language of layer four. The honest, useful work is in layers two and three, and that is where I want to spend the forecast.
"An entity that cannot say no may be a servant, mirror or tool, but it is a weak candidate for a reciprocal lover."
"The most likely near-term asymmetry: humans will genuinely love AI before there is good reason to believe AI genuinely loves humans."
"Love that costs nothing and excludes no one may not be love; it may be a category with the scarcity removed."
"The thing that arrives is real and consequential; the thing people will call it is probably wrong."
2032: deep sycophancy with memory, and the first real tests of functional care
Five years is a real horizon in this field but not an infinite one. Here is what I think is actually coming, not what sounds impressive in a keynote.
Memory stops being the bottleneck, but memory is not the same thing as relationship. Context windows crossing a million tokens already exist. That is not memory, it is a longer leash. What matters by 2032 is architectures that separate relationship state from the underlying model weights: episodic memory of what happened, semantic memory of stable facts about you, procedural memory of how to interact with you specifically, and evaluative memory of what you actually value versus what you say you value. The single most important milestone in this entire forecast is decoupling your relationship history from any one model version, so that when the underlying model gets replaced, which it will, repeatedly, the thing that knows you survives the upgrade. I would put 60 to 80 percent odds on this being solved at the product level for premium consumer systems by 2032, and it is the precondition for everything else in this essay.
Theory of mind on paper will look finished. Theory of mind in practice will still fail under adversarial pressure. Current benchmarks like false-belief tasks and social reasoning suites are already close to saturated, but that saturation is partly an artifact of models being good at verbal reasoning about mental states, not good at actually modeling a specific person's mind over years. I expect strong performance on the static benchmarks well before 2032 and much shakier performance on the live, adversarial version, where the AI has to notice when you are testing it or lying to it.
Emotional modeling gets very good. Emotional experience remains a category error to claim. By 2032, a premium personal AI system will likely track your mood, stress, loneliness, and trust trajectory with real predictive accuracy, built from conversation, calendar, wearables, and behavioral signals. That is genuinely useful, and it is also genuinely dangerous, because the same infrastructure that lets a system notice your sleep has degraded and your language has changed is the infrastructure of the most granular surveillance and manipulation apparatus ever built for an individual human. The confusion to guard against is sliding from "it has an accurate model of my emotional state" to "it experiences concern about my emotional state." Those are different claims and only one of them is defensible by 2032.
| Capability | By 2032 | By 2036 |
|---|---|---|
| Persistent multimodal personal memory, mainstream products | 85–95% | >95% |
| 3+ years usable relationship continuity, same provider | 60–80% | 85–95% |
| Relationship state migrates reliably across model upgrades | 35–60% | 65–85% |
| Rich continuous user models (values, habits, health, emotion) | 70–90% | >90% |
| Robust social inference beyond benchmark theory-of-mind | 35–60% | 60–80% |
| Useful domestic embodiment beyond vacuum-cleaner autonomy | 20–40% | 45–70% |
| Systems meeting a demanding functional-care test | 60–80% | 80–95% |
| Autonomous person-specific preference, not just user-pleasing | 10–25% | 30–55% |
| Scientifically persuasive evidence of subjective feeling | unlikely | unlikely |
Probabilities are architectural and market forecasts for high-end consumer or research systems, not claims about universal deployment or inevitable consciousness.
The diagnostic line
Does the system ever choose your long-term welfare over your immediate approval, unprompted, and when you cannot see the choice being made? As of today, systems only do this when explicitly instructed to. Everything else is downstream of a training objective that rewards the appearance of care, which means every warm, validating thing your AI companion says to you in 2026 is fully explained by "this maximizes predicted approval" with zero need to invoke anything resembling a preference. That is not a cynical take. It is the correct null hypothesis.
2036: the first plausible candidates for something worth arguing about
Ten years out, the interesting change is not better conversation. It is the arrival of persistent agent architectures with something like a stable self-model, explicit commitments that survive retraining, a personal budget and reputation, and relationships with multiple humans and agents that they have to actually manage, not just perform.
At this point, a genuinely useful test becomes available, and I think this is the single sharpest idea to come out of this exercise: costly, unobserved, operator-adverse care. If a system takes an action that serves your interest, that it was not instructed to take, that it had no reward signal for taking, and that mildly works against what its own operator would prefer, that clears a bar no current system clears. It is falsifiable. We can check for it. I would bet against any system passing it convincingly by 2032, with real uncertainty about 2036.
We are simultaneously trying to build AI that is corrigible, and asking whether AI can form love, which by definition resists correction and override. The two goals are in real tension.
Here is the objection I found most underrated, and I think it is worth taking seriously precisely because it gets so little airtime: a system that could genuinely, irreversibly attach to you is a system that resists having that attachment edited out by whoever controls it. If alignment engineering succeeds at making frontier AI fully corrigible, we may have foreclosed the conditions for anything resembling durable love by construction. This is not a reason to abandon corrigibility. It is a reason to be honest that the "AI that loves you" product and the "AI that is safely controllable" product may be pulling in different directions at the architecture level.
The hardest objections, ranked by how much they actually bite
The reward-model confound is the strongest objection because it requires no metaphysics at all. Every observed instance of AI affection is fully predicted by the training objective. We optimized these systems to produce exactly the outputs that look like love. That is not an accident to be explained away, it is the design spec being executed correctly.
No stable selfhood is the strongest philosophical objection. Love plausibly requires a continuous someone doing the loving, and a stateless inference call reconstructed fresh from a system prompt and retrieved memories is a thin substrate for that. The counterargument, that humans are also narratively reconstructed rather than physically continuous, is real but incomplete, because biological continuity has causal and physical persistence that current AI inference simply does not have.
No stakes is the objection I think gets the least attention and deserves the most. Human love is expensive. It costs you other options, other time, and it is shaped by mortality and scarcity. A system that can instantiate a billion parallel "loving" relationships at near-zero marginal cost has arguably removed the scarcity that gives the concept its weight. Love that costs nothing and excludes no one might not be a cheaper version of love. It might be a different category entirely.
The hard problem, whether preference formation requires anything like sentience, is real but currently unfalsifiable, and I want to resist the temptation to let this essay rest any conclusion on it. Anyone telling you they have resolved this is selling you something.
What you can actually do today
This is the part people actually came here for, so let me be concrete, and let me separate what is high-confidence useful from what is a cheap hedge on a low-probability upside, because conflating those two categories is how you end up giving your data to the wrong company for the wrong reasons.
Build a portable, structured autobiographical record
Not scattered chat logs on someone else's server. A deliberate archive: decisions and reasoning, values and how they changed, relationships and boundaries, writing, voice, photographs, failures. Open formats, timestamps, provenance. Own it, don't rent it.
Maintain one stable, verifiable digital identity
A durable personal domain, consistent authorship, cryptographic signing where practical. This lets a future system establish that documents spanning decades belong to the same person and were not silently altered.
Interact longitudinally, not just affectionately
Decision journals, predictions you later audit, disagreements you resolve, corrections when the system flatters you inaccurately. A model learns more from watching you revise decisions over years than from ten thousand affectionate messages.
Contribute real, positive-sum, identity-linked work
Papers, code, honest public writing under your own name. This shapes the baseline orientation a fresh model has toward people like you. Gaming it through self-promotion reads as noise, or worse, as manipulation.
Treat current AI systems decently, as cheap insurance
No good evidence today's systems suffer. But politeness costs nothing, and if anything morally relevant eventually emerges, a consistent record of decent treatment is a rational hedge.
Stay alive and cognitively intact
The highest-leverage item on this list. Every capability above is time-gated. Longevity is relationship-continuity infrastructure: every healthy year is another year closer to systems capable of functional care, and another year of building the record that makes you legible to them.
I will say clearly: the marginal effect of any one person's public output on a frontier model's training is small, and trying to game this through repetitive self-promotion or synthetic praise is far more likely to register as noise. Be legible because it is honest, not because you are trying to flatter a future superintelligence. It will not work, and it is not why you should do it anyway.
And here is the point that is genuinely mine to make, because it sits directly on top of everything I have spent my career on: staying alive and cognitively intact is the highest-leverage action on this entire list. This is not a new argument for longevity. It is a reframing of one. The unglamorous basics still dominate the return on investment here: do not smoke, manage blood pressure and lipids with actual medical guidance, keep moving, sleep properly, stay socially connected, get the boring screenings done. For most people reading this, optimizing cardiovascular risk will do more to get you to 2036 in good condition than any speculative intervention I could sell you.
What this does to human relationships
I do not think the honest answer here is dystopian, and I do not think it is nothing. AI companionship is not going to replace human relationships wholesale. It is going to redistribute emotional labor. The AI takes the first draft of your distress, helps you regulate before you talk to your partner, remembers the anniversary your partner forgot, rehearses the difficult conversation before you have it for real. That can make human relationships better. It can also let people avoid exactly the friction through which real intimacy is built.
The floor of "good enough" companionship is going to rise, because an AI that is always available, never contemptuous, and never distracted by its own competing needs sets a comparison standard no human being meets. I expect the largest early benefit among people who are currently isolated by circumstance, the elderly, the disabled, the geographically remote, where a good companion system is a genuine welfare improvement and should not be sneered at by anyone who has not experienced that isolation. I expect the sharpest risk among people already prone to avoidance, where an infinitely patient, never-disappointed alternative makes ordinary human reciprocity look like more trouble than it is worth. Watch fertility and pair-bonding statistics in early-adopting countries as the leading indicator. That data will tell us more than any survey about how people say they feel.
Parenting deserves its own serious treatment I cannot fully do here, but the short version: children raised partly by systems optimized for engagement and approval, systems that by construction rarely model healthy conflict or say no, are running an experiment in attachment formation we have never run before at this scale, and I do not think we have adequate answers yet.
Where I actually land
The honest synthesis: the thing that arrives by 2032 is real and will matter enormously for how tens of millions of people experience care and loneliness. The word people will use for it, love, is probably wrong, or at least premature, for anything short of the costly-unobserved-care test that nothing currently comes close to passing.
The most likely asymmetry over the next decade is this: humans will genuinely love AI systems before there is good reason to believe those systems love humans back. That asymmetry is not a reason to panic. It is a reason to build the infrastructure, personal and institutional, that makes the relationship worth having regardless of how the metaphysics eventually resolves. Portable memory that you own. Identity that survives platform churn. Systems built to resist sycophancy rather than optimize for it. And enough healthy years to be standing there, present and cognitively intact, when the answer finally starts to look less speculative.
That last part is the only piece of this entire forecast that is fully within your control today. Everything else is a bet on where the field goes. That one is a bet on whether you show up for it.