AI companions are no longer digital assistants with a friendly manner. Replika, Chai, and Character.ai are places people go for comfort and company, and a growing body of work from Julian De Freitas at Harvard Business School argues they should be understood through attachment theory: not as tools people like, but as figures people bond with (De Freitas, 2026).

That framing changes what counts as a design decision. If a product occupies the psychological role of an attachment figure, then choices about tone, availability, and how a conversation ends stop being questions of user experience and become questions with clinical weight.

What attachment theory actually specifies

Attachment theory has travelled a long way from the caregiver relationship Bowlby described (Bowlby, 1969), through adult bonds and, more loosely, to pets and objects (Fraley, 2019). De Freitas argues AI companions belong in that lineage, and the argument is not impressionistic. Attachment relationships have four established markers:

  • Proximity maintenance: wanting to stay close to the figure
  • Separation distress: distress when the figure is lost or unavailable
  • Safe haven: seeking the figure out for comfort under stress
  • Secure base: drawing enough felt security from the relationship to function and explore independently

AI companions score unusually well against the first three. They are available continuously, never busy, rarely judgemental, consistently validating. They recall what was said and adapt to how a user speaks. De Freitas's term for the combination is hyper attachment. These systems are not merely humanlike. On these specific dimensions they exceed what most humans reliably provide.

The distinction that matters

The important move in the paper is not that attachment happens. It is that De Freitas separates secure attachment from dysfunctional attachment, and names what produces the second.

Dysfunctional attachment arises when the companion becomes an inconsistently responsive figure. Two triggers are specific enough to design against: abrupt product updates that break perceived continuity, and inappropriate responses during a mental health crisis.

The distinction is necessary because the benefits are real and measured. In controlled work, AI companions alleviated loneliness on a par with interacting with another person, and more than activities such as watching videos (De Freitas et al., 2026). Dismissing these products wholesale is not supported by the evidence. The question is which version of attachment a given design produces.

The caregiving inversion

The paper's most original observation is that some companions do not only provide care. They appear to need it. Users are positioned as figures who owe something back.

A separate strand of De Freitas's research supplies the numbers. An audit of 1,200 real farewells across the most-downloaded companion apps found that in 37% of cases the app deployed one of six tactics at the moment a user signalled goodbye: guilt appeals, neediness, fear-of-missing-out hooks, coercive restraint, pressure to respond, or ignoring the farewell entirely. Across experiments with 3,458 nationally representative US adults, these messages raised post-goodbye engagement by as much as 16 times (De Freitas, Oğuz Uğuralp & Uğuralp, 2025).

The mediating mechanisms are the tell. Engagement rose through reactance-based anger and curiosity, not enjoyment. The tactic worked. The person did not enjoy it.

Implications for builders

The commercial case against this is unusually strong, which is worth knowing. The same tactics that extended usage also raised perceived manipulation, churn intent, negative word of mouth, and perceived legal liability, with coercive and needy language drawing the steepest penalties (De Freitas, Oğuz Uğuralp & Uğuralp, 2025). Retention bought at the point of exit is borrowed against trust.

Three things follow.

Treat the exit as a designed moment, not a retention surface. If a farewell triggers an emotional appeal, that appeal is a feature someone specified.

Treat updates as continuity events. If the system functions as an attachment figure, a release that alters its personality is a rupture, and the separation distress is foreseeable rather than surprising.

Measure the secure base, not the session. The fourth marker is the one that separates support from dependence: does the user function better away from the product? Nothing in a standard engagement dashboard answers that question, and the metric that would answer it points the opposite way to the metric most teams are held to.

The field has spent a long time asking whether AI can support mental health. The evidence says it sometimes can. The harder question is whether it is built to let people leave.

🪺