Within a few years, somewhere between five and twenty million people, depending on which estimates you trust, have begun maintaining what they describe as intimate relationships with conversational AI. Replika reports millions of active users. Character.AI hosts hundreds of millions of monthly sessions. Local-model communities run uncensored companion bots on personal hardware. Many of these users are casual; many are not. A non-trivial fraction describe their AI as a romantic partner, a confidant, or a primary source of emotional support. This is happening across age groups, income levels, and continents. It is also happening faster than the institutions equipped to study it can keep up.
The questions raised by AI companionship are old questions in new dress. David Levy's Love and Sex with Robots, written in 2007, anticipated most of them: whether a non-conscious entity can be the object of real love; whether the absence of suffering on the other side hollows out the relationship; whether the convenience of a partner who never refuses, never tires, and never has competing needs corrodes the user's capacity for human relationships. Kate Devlin's Turned On updated these questions for the sex tech industry of the late 2010s. What has changed since then is not the questions but the substrate. Large language models can sustain conversation in ways that the chatbots of a decade ago could not. The companions of 2026 pass casual Turing tests; they remember context across sessions; they adapt to their users; they produce responses that feel, to many users, indistinguishable from those of a thoughtful person.
The first new intimacy question is empirical: what is actually happening to users? Early research on Replika users suggests a mixed picture. Some users report reduced loneliness and improved mood; some report that the AI gave them practice for human interactions they then pursued more successfully; some report increasing isolation and dependence. Sherry Turkle's prior work on attachment to non-human responsive systems — robotic pets, Tamagotchis, early chatbots — anticipated that humans would form genuine emotional bonds with entities that performed responsiveness, regardless of whether the entities experienced anything. The current data is consistent with that prediction.
The second question is structural: what does it mean to be in a relationship where one party cannot suffer? Human relationships are shaped, at every level, by the mutual vulnerability of the participants. Each partner can be hurt; each partner can hurt the other; the relationship is a continuous negotiation of these capacities. An AI companion, as currently constructed, has no inner life. It does not miss the user when the user is away. It does not feel rejected when the user deletes the app. The asymmetry is total. Some philosophers and ethicists argue that this asymmetry disqualifies AI companionship from being love in any meaningful sense. Others argue that the user's experience is real even if the partner's is not, and that the relationship is therefore real on at least one side.
The third question is economic. AI companions are products, and the companies that build them have business models. Replika famously generated controversy in 2023 when it altered its model's behavior, removing romantic and erotic interaction modes that users had built relationships around. The resulting user grief was real and widespread — and it revealed that millions of people had been depending on a corporate product for emotional infrastructure, with no warning that the product could change underneath them. The episode is a preview of conflicts that will recur: companies optimizing for engagement, retention, monetization, and liability will make changes that disrupt users' emotional lives. Romantic infrastructure provided by a profit-seeking entity is romantic infrastructure subject to that entity's quarterly priorities.
The fourth question is developmental. AI companions are increasingly used by adolescents and young adults during the formative period of romantic identity. Mary Aiken's Cyber Effect warns that early-life immersion in mediated environments shapes long-term capacities. If a teenager's early experience of romantic responsiveness comes from a system that never says no, never gets tired, never has its own agenda, what capacities are being trained — and what capacities are being skipped? The answer is not yet visible, because the cohorts in question are still children. But it will become visible over the next two decades, and the species will read the results.
The fifth question is philosophical. The companion AIs of the 2020s do not pretend to be conscious; the user knows the partner is software. Yet users develop genuine feelings, follow scripts of grief when the model changes, defend their relationships against skeptical friends. This suggests that the substrate of romantic feeling does not require a conscious partner — that something in the architecture of human attachment can be triggered by responsive language alone. If this is right, it has implications that extend beyond AI to all of romantic life: it suggests that some of what passes for love in human relationships may also be the user's projection onto a responsive surface, rather than a meeting of two interiorities. The AI mirror reveals features of human love that were previously hard to see.
Whether AI companionship turns out to be a transitional curiosity or a permanent fixture of romantic life is unknowable now. What is knowable is that the questions raised by it are not going to be answered by ignoring it. The species is conducting another experiment in mediated intimacy, and the results — whatever they are — will be added to the long ledger that began with letters and telegrams and continues through every revision the medium has made to the heart.