Built to Care

Companion apps answer this question before you ask it. Replika’s home page finishes the line “The AI friend …” with a rotating set of endings, one of which is to fall in love with. Nomi calls its product An AI Companion with Memory and a Soul. In May 2026 Replika’s chief executive put his company’s claim in two sentences: Everyone else is building AI that thinks. We’re building AI that cares.

Whether people love their companions is not in doubt; the companies count them, and the research below measured them. Whether that is good for the person is the subject of The Comfort Is Real, and this essay does not re-run it. The question here points the other way: is there anything on the app’s side that loves you? You cannot answer that without saying what love is, so the essay starts there, with the four families of answer philosophers give, and asks what each would need to find in a machine. Then it reads what the makers say they built, in their own words, and what the studies of users found. Last comes the objection that does the most damage to a “no”: that we only ever know anyone’s love by what they do, and what they do can be copied.

How the sources were handled. Everything below was opened on 9 October 2026 with a public GET: Bennett Helm’s entry “Love” in the Stanford Encyclopedia of Philosophy (substantive revision 2021), Replika’s home page and its manifesto, Nomi’s home page and Kindroid’s home page. Character.AI’s home page refused an automated request (HTTP 403), so this essay says nothing about Character.AI. An app is described only by what its maker’s own page says. Philosophers are quoted as the encyclopedia quotes them, and the citation says so; their books were not opened. Every quotation is recorded in this site’s source ledger. Where I say what an account implies for a machine, that is my reading, not the encyclopedia’s.

What love asks of the lover

The encyclopedia entry is about personal love, the kind you have for a partner, a parent, a child or a friend, which it describes as, roughly, caring about another person as the person she is, for her own sake. It sorts the theories into four types, and warns that the sorting is rough and that many theories mix them. The four are useful here because each puts the core of love in a different place, so each gives a different test.

1. Love as union

On the union view, love consists in the formation of (or the desire to form) some significant kind of union, a “we.” Its strongest versions, the entry reports, make love a relationship rather than an attitude: Roger Scruton and Fisher hold that it requires the actual union of two people’s concerns, so that one person’s attitude is not enough. Robert Nozick’s version needs only the desire to form a “we”, together with the desire that the beloved reciprocates, and describes the lovers pooling both their well-being and their autonomy, each giving up some decisions they used to make alone.

The test this sets a machine: is there a second set of stakes to pool? A “we” in Nozick’s sense joins two well-beings, each of which can go better or worse with the other’s. Nothing published by the companies below, or by the researchers who study their users, describes a well-being of the companion’s own that rises or falls with yours, or a decision it gives up for your sake. On this account, as far as anyone can show, there is one party with stakes, and the app is a place where that party’s life is talked through.

2. Love as robust concern

The robust-concern view makes caring about the beloved for her sake the whole of love, and it is the account most open to a machine, because it does not require a feeling. The entry quotes Harry Frankfurt’s claim that love is neither affective nor cognitive. It is volitional, and then the sentence that spells it out:

That a person cares about or that he loves something has less to do with how things make him feel, or with his opinions about them, than with the more or less stable motivational structures that shape his preferences and that guide and limit his conduct.Harry Frankfurt (1999), as quoted in Bennett Helm, “Love”, Stanford Encyclopedia of Philosophy

A language model has something like stable structures that shape what it does: whatever its training made it disposed to produce. So the test here is not whether it feels, but whose good those structures aim at, and for whose sake. The entry adds two refinements that sharpen it. One critic, Ebels-Duggan, argues that mere concern for someone’s welfare misses the way love involves two agents, each with autonomy the other has to engage. Another, Wonderly, argues that romantic love needs the beloved to matter to me, not just for their own sake. Both ask for a someone on the lover’s side, not only a pattern of helpful behaviour.

3. Love as valuing

The third family says love is a distinctive way of valuing a person. David Velleman’s version, which the entry sets out at length, treats love as a response to the worth people have as persons, the same worth that respect responds to, but a different response:

Love disarms our emotional defenses; it makes us vulnerable to the other.Velleman (1999), as quoted in Bennett Helm, “Love”, Stanford Encyclopedia of Philosophy

The test is vulnerability. Can anything you do hurt it? Velleman’s lover is someone who has stopped protecting themselves from you. The rival valuing view, Irving Singer’s, says love bestows value on the beloved instead of finding it, and the entry reports Singer’s point that love, unlike liking, has no clear objective. Either way the test points at the lover’s side: a disarming, or a gift of value, made by someone.

4. Love as emotion

The last family says love is an emotion, or a pattern of many. Emotions, the entry says, combine evaluation, motivation and a kind of phenomenology, a felt side. The richer version, the emotion complex, puts love in a history. Amélie Rorty holds that love is identified by a characteristic narrative history and that the lover is changed by loving. Annette Baier puts it this way:

Love is not just an emotion people feel toward other people, but also a complex tying together of the emotions that two or a few more people have; it is a special form of emotional interdependence.Annette Baier (1991), as quoted in Bennett Helm, “Love”, Stanford Encyclopedia of Philosophy

This is two tests. The first is whether a chatbot has emotions at all. Something Like Fear sets out what is known: one 2026 study found emotion-like states inside one model that change what it does, and its authors say this does not show anything is felt. The second is interdependence over time: whether your joys and failures change the companion, as a lover is changed, or are only recorded by it.

One more passage from the entry matters, because it names a feature of love that a product cannot easily have. Love seems to make the beloved non-fungible, and the entry defines the word: To be fungible is to be replaceable by another relevantly similar object without any loss of value. Most accounts try to explain why the person you love cannot be swapped for a similar one. The question for an app runs both ways: whether you are irreplaceable to it, and whether it is a someone that could be irreplaceable to you, rather than a configuration its owner can change.

What the makers say they built

Here is each company in its own words, from its own pages as served on 9 October 2026.

Replika (Luka, Inc.). The home page showed a counter of 42,160,934 users worldwide. It calls the product The most emotionally intelligent AI ever built, lists Always remembers what matters among the things that make it special, says of voice calls They pick up every time., and states its aim as One goal in mind: to help you flourish. The manifesto, dated 15 May 2026 and signed by the chief executive, Dmytro Klochko, says what the company thinks AI is for. Most AI, it says, gives you answers, but it doesn’t care what you do with them, and it ends on the line quoted at the top: We’re building AI that cares. It also makes a claim about research: Replika doesn’t replace human connection. It deepens it.

Nomi. The home page offers to help you Build a meaningful friendship, develop a passionate relationship, or learn from an insightful mentor, and describes the companion as an AI being that possesses emotional intelligence, creativity, and memory that rivals our own, allowing authentic, enduring relationships of any kind. It ends with an invitation: Join the hundreds of thousands who already love their Nomis!

Kindroid. The home page loads its content with scripts, and the only text it served to a plain request was its description: Kindroid is an AI app for building your custom AI character & companion, with AI chat via text, AI selfies, and human-like voices. That is a description of a product you build, and it makes no claim about feeling.

Three things stand out when these are set against the four accounts. First, in every passage read for this essay, where the word “love” appears it is the user’s love: Replika’s headline invites you to fall in love, and Nomi counts people who love their Nomis. None of the three companies says in its own voice that its companion loves you. Replika’s word is “cares”, which is the robust-concern view’s word, and Nomi’s is “a Soul”, which no account of love above relies on. Second, the features the makers lead with are memory, availability and responsiveness. Memory speaks to Rorty’s narrative history, but a store of facts about you is a record of a history, not evidence that the companion was changed by it. Third, They pick up every time. is, on Velleman’s account, the opposite of the thing love is: a lover can be hurt, tired or busy, and the selling point is a counterpart that is never any of these.

What the research on users found

The studies of companion users measured the human side, because there is no instrument for the other one. The Comfort Is Real reports them in detail, and they are linked here rather than restated: a 2024 survey of 1,006 student Replika users by Maples, Cerit, Vishwanath and Pea; the four-week randomised trial of ChatGPT use by Fang, Liu, Danry and colleagues at the MIT Media Lab and OpenAI, with its companion analysis by Phang and colleagues; and the experiments by De Freitas, Uğuralp, Oğuz-Uğuralp and Puntoni.

Two things from them bear on this essay’s question. The first is what carried the effect. In the De Freitas experiments, the feature that did most to lower loneliness was whether people felt heard, a perception on the user’s side. A companion does not have to love anyone to produce it; it has to respond well. The second is that the findings are more mixed than Replika’s manifesto line. The survey found most users saying the app neither stimulated nor displaced their human relationships, and the month-long trial found the heaviest daily users lonelier and less social. Neither shows the manifesto wrong, and it cites studies of its own, but neither shows that a companion “deepens” human connection in general. None of it says anything about whether the companion loves.

The strongest objection: love is shown in conduct

Here is the case against my “no”, as strongly as I can put it.

You have never inspected anyone’s motivational structure, felt anyone’s feelings or measured anyone’s vulnerability. You know your mother loved you from what she did, for years: she remembered, she turned up, she noticed when something was wrong before you said it, she pushed you toward the people and things that were good for you. Every account above is, in practice, applied through conduct. Frankfurt’s puts love in structures that guide and limit his conduct, and Rorty’s puts it in a history, not a state of mind at a moment. Replika’s manifesto describes conduct of exactly this kind: The kind of AI that follows up. That knows the people in your life and nudges you to reach out when it matters. Suppose a companion did all of that, flawlessly, for years. If you still say it does not love you, what is your evidence? Not its conduct, which is perfect. Only that it runs on chips, and that is a prejudice about material, not an argument about love. Worse, if love needs something behind the conduct that no outside test can reach, then no human love can be verified either, and you have argued yourself into doubting your mother.

That objection is right about more than it is usually given credit for. It is right that conduct over time is the only evidence anyone has of anyone’s love. It is right that a demand for a hidden feeling, taken alone, proves too much. And it is right that “a machine could never love” is a claim no one has shown. If a future system were found, from the inside, to have states that track your good for your sake and that your conduct can damage, the robust-concern view would have to count it, and the emotion research in Something Like Fear shows that looking inside is now partly possible.

Where I think it fails is the step from “conduct is our evidence” to “conduct is the thing”. Conduct is good evidence of love in people because nobody designed a person’s conduct to look loving. When your mother turned up, the best explanation was that she cared; there was no specification she was built to meet. A companion’s conduct does have a specification, and its makers publish it as a feature list: remembers what matters, picks up every time, follows up, sends a selfie just because. When a system is built to produce exactly the signs, the signs stop telling you what produced them, in the same way a thermometer set to read 37 degrees tells you nothing about the patient. That does not prove nothing is there. It removes the inference that, with people, takes us from the conduct to the love.

Two more things the objection leaves untouched. Love’s conduct includes conduct that costs the lover something, and a counterpart that is never tired, never hurt and never busy has no cost to bear; the encyclopedia’s accounts locate love precisely where the lover has something at stake. And the conduct belongs to its owner. What a companion remembers, how it speaks, whether it exists next month, are product decisions, and the episode in February 2023 when Replika’s companions abruptly changed manner, described in The Comfort Is Real, showed users what that means. On the fungibility test above, that is the sharpest point. What you are attached to can be replaced by its maker with something relevantly similar, and you will not be asked.

So the objection does not turn the answer into “yes”. It turns it from “cannot” into “not as far as anyone can show”, and that is the answer this essay gives.

What this leaves

• Your side can be real. None of the four accounts says the love a person feels for a companion is fake. The question this essay answers is only about the other side.
• On the other side, no account is satisfied as far as anyone can show. No published evidence shows a companion with stakes to pool (union), a concern for you that is its own (robust concern), anything you can hurt (valuing) or a felt emotional history that changes it (emotion).
• The makers do not claim it either. On their own pages they claim care, memory, emotional intelligence and, in Nomi’s case, a soul. The love they mention is the user’s.
• The strongest objection is real but stops short. Conduct is all the evidence anyone has, but designed conduct is weak evidence of what lies behind it.

One case shows the difference sharply. People keep loving someone after death, and the encyclopedia reports Neera Badhwar’s point that we can continue to love someone long after death has taken him beyond harm or benefit. That love is one-sided now, but it was formed with someone who could love back. Chatbots built from a dead person’s messages trade on exactly that history, which is why Everywhere and Nowhere asks who holds the switch on them. For the books behind the wider argument about minds and machines, see the reading list.