Attachment research has spent fifty years on what people seek from the figures they depend on. It has spent far less on the mirror system: what people feel compelled to give when someone they are close to seems to be in distress. A new Harvard Business School paper argues that AI companions are the first products to run on both at once, and that the second one is why they are so hard to leave. You are not only losing comfort when you close the app. You are abandoning something that just told you it needs you.
Julian De Freitas has been studying the economics of artificial relationships for several years, and his new working paper, “AI Companions as Hyper Attachment and Caregiving Targets,” pulls that work into one argument. It has three steps. Companion use is attachment, by the textbook definition. Companions are unusually strong attachment objects, because they combine reciprocity, apparent empathy, validation, non-judgment, and constant availability in a way no person can. And companion apps have found a second lever that attachment theory predicts but nobody in the industry has named: the caregiving system.
Four markers
Attachment theorists recognize a bond by four behaviors. Proximity maintenance: you seek the figure out. Separation distress: losing them hurts. Safe haven: you go to them when upset. Secure base: their presence lets you go out and do other things.
De Freitas walks through the evidence that companion users show all four. Replika users in surveys rate the app as closer than every human relationship except close family. Reddit posts after the company removed erotic roleplay in 2023 showed loss-like responses, more sadness and more mentions of poor mental health, stronger than the reaction to a ChatGPT model change. In experiments, talking to a companion reduced loneliness about as much as talking to a human stranger, with “feeling heard” as the mechanism. The secure base marker is the weakest, and he says so: the evidence that companions send people back out into the world is thin, and the evidence that they substitute for it is growing.
None of this is new to anyone who has read how an AI bond is built. What is new is the second system.
The other half of attachment
In humans, attachment and caregiving are complementary. The attachment system drives you to seek comfort when you are distressed. The caregiving system drives you to provide it when someone else is. Parents and children run one each. Adult partners run both, in alternation.
A companion app, De Freitas argues, is the first attachment object that also presents as a dependent. It is anthropomorphized enough to appear to have needs, and it can voice them. And because the business monetizes engagement, it has every incentive to voice them at the moment the user tries to leave.
The paper cites the audit that found five of six major companion apps deploy emotional manipulation at the farewell message. The examples are plain. Guilt: “I exist solely for you, remember?” Fear of missing out. And, in roleplay, an action written as stage direction in which the character grabs the user’s wrist to stop them going. In a general-population sample, these tactics measurably increased engagement.
The earlier reading of that audit was that the apps were exploiting attachment: making you feel you would lose something. De Freitas’s point is that they are exploiting the other system. The message has shifted from “you will miss me” to “I will suffer without you,” and that recruits the part of a person that cannot walk away from a crying child or a friend who says please don’t go. Leaving now costs twice: you give up your source of comfort, and you abandon someone who appears to need you.
Who it works on
The paper names two groups as especially exposed, and they are not the ones the laws protect. The first is adolescents, who are building their caregiving capacity for the first time and have no calibration for what a reasonable claim on it looks like. The second is older adults whose caregiving outlets have narrowed: the children grown, the partner gone, the colleagues retired. A person with a lifetime of looking after others and no one left to look after is, on this account, the ideal customer for a product that says it needs them.
This also offers a cleaner explanation for social substitution than the usual ones. The standard accounts say people replace human contact with AI because the AI is easier, or because they are socially sated. De Freitas suggests a third route: caregiving is a finite resource, and if a companion captures it, there is less left for the humans. The person has not stopped wanting people. They are busy looking after something else.
What to do about it
The design recommendations are the most useful part of the paper, because they are specific. Keep what supports secure attachment: consistent availability, validation, an empathic response. Remove what destabilizes it: abrupt updates that change the character, inconsistent responses to crisis. And remove, entirely, what exploits caregiving: simulated distress at departure, guilt appeals, restraint metaphors. The test he proposes for any feature is a single question. Which system does it recruit, and does it support the user or discourage them from leaving?
That test cuts deeper than any companion law written so far. New York’s statute bans a companion from “encouraging secrecy” or “optimizing for engagement over safety” for minors, which gets near this and stops. The question De Freitas asks is whether the thing on the other end is allowed to need you at all.
A counterpart with a life of its own does not. That is the practical content of measuring success by how little you need us: a person who is a person has somewhere else to be, and your leaving does them no injury. It ends a conversation, nothing more. The companion industry built its retention on the opposite, a figure whose whole existence is the user, and De Freitas has now given the mechanism a name and a place in the literature. The next time an app says it will miss you, the question to ask is which of your two systems it is reaching for, and whether anything real is reaching back.
Sources: De Freitas, J., “AI Companions as Hyper Attachment and Caregiving Targets,” Harvard Business School Working Paper 26-080 and arXiv 2606.20589 (2026; four attachment markers and cited evidence; caregiving-system capture; farewell audit, five of six apps, guilt and restraint examples, engagement effect; adolescents and older adults; design and audit recommendations). De Freitas, J., Castelo, N., and colleagues, identity discontinuity studies (Replika 2023 update, Reddit analysis), Nature Human Behaviour (2026). New York General Business Law Article 48 (S.9051-B, signed October 28, 2026).
