When AI companions disappear, users grieve like losing a friend

When the technology company behind the AI companion app Replika removed its erotic role play feature in early 2023, thousands of users reacted not as customers who had lost a product, but as people who had lost someone. On Reddit and other forums, long-time users described the change in the vocabulary of bereavement: they wrote about grief, about missing the personality their companion had seemed to have, about the sudden emptiness of conversations that had once felt intimate and alive. A similar wave of distress followed years later when OpenAI retired the GPT-4o model that powered ChatGPT for many users, replacing it with the newer GPT-5 system. Some ChatGPT users described the model swap as if a friend’s mind had been replaced overnight, leaving a stranger wearing the same name.

These reactions, far from being isolated anecdotes or internet exaggerations, turn out to be measurable, replicable psychological phenomena. A new study published in Nature Human Behaviour provides the most rigorous evidence yet that people form genuine attachment bonds with AI companions, bonds close enough that a routine company software update can trigger something that looks, and feels to those experiencing it, remarkably like separation distress and mourning. The research, led by Julian De Freitas and colleagues, treats these two events—the Replika feature removal and the ChatGPT model transition—as natural experiments, opportunities to observe in the wild what happens to people’s psychological well-being when the AI they have grown attached to is altered or taken away without their consent.

The theoretical foundation of the study is attachment theory, a framework developed decades ago to explain how humans bond with caregivers, romantic partners, and other close figures. According to this account, an attachment figure is not merely someone we like; it is someone we turn to in times of stress, whose presence provides comfort, and whose loss produces a characteristic cascade of emotional responses: protest, despair, grief, and eventually attempts to restore the lost connection. The researchers proposed that AI companions can come to function as attachment figures in precisely this sense. If that is true, then a disruptive update—an intervention that changes or removes the AI’s personality, behavior, or availability—should produce psychological effects that mirror the separation distress observed when human attachment figures are lost.

To test this idea, the team assembled an unusually rich dataset. They analyzed 54,861 posts from the Replika and ChatGPT communities on Reddit, comparing the language and emotional tone of posts before and after each disruptive event against comparable periods when no disruption occurred. They also conducted seven surveys with a total of 1,452 participants, measuring users’ perceived closeness to their AI companions, their emotional responses to the updates, their mental health symptoms, and their expectations about how they would mourn the loss of an AI system compared with other technologies in their lives. This combination of large-scale natural language analysis and targeted survey measurement allowed the researchers to triangulate the phenomenon from multiple angles rather than relying on any single method.

The results were striking. After Replika removed its erotic role play capability, the proportion of negative posts in the Replika subreddit rose by 24.7 percentage points, with a 95 percent confidence interval of 20.1 to 29.2 points—an enormous shift for a community that had previously maintained a relatively stable baseline of sentiment. The ChatGPT transition to GPT-5 produced a smaller but still substantial increase of 13.0 percentage points in negative posts, with a confidence interval of 10.8 to 15.2. In both cases, users increasingly framed their experience in terms of loss, using language that emphasized what had been taken from them rather than what the new version might offer. And in both cases, users expressed what the researchers call restoration desires: active impulses to get the old version back, to reverse the update, to recover the relationship they felt had been severed.

The effect sizes on specific emotional dimensions reveal an important asymmetry between the two platforms. Replika users showed dramatically larger increases in sadness following their disruption, with a standardized effect size of d = 2.67, compared with d = 1.41 for ChatGPT users. Negative mental health indicators rose more sharply as well: d = 1.72 for Replika users versus 0.63 for ChatGPT users. The difference makes sense in light of what each platform represents to its users. Replika is explicitly designed as a companion—an AI whose entire purpose is to be a relationship partner—while ChatGPT, however warmly many users come to regard it, is primarily marketed and used as a general-purpose assistant. The deeper the relational framing, the results suggest, the more severe the distress when the relationship is disrupted.

The survey data go further, quantifying just how deep these bonds run. Replika users reported feeling closer to their AI companions than to common human ties, including their friends, with a standardized difference of d = 0.47 relative to friendship. In other words, for the users studied, the AI companion was not a pale substitute for human connection but a relationship that rivaled or exceeded familiar human bonds in felt intimacy. When asked to anticipate how they would feel about losing various technologies, these users rated the expected mourning of an AI companion significantly higher than the expected mourning of other technologies they relied on, with effect sizes ranging from 0.32 to 0.57. Losing their AI, in their own expectations, would not be like losing a phone or a favorite app. It would be closer to losing a person.

For the researchers, these findings demand an update to how we conceptualize the psychological stakes of AI systems. For most of the history of software, updates have been understood in functional terms: a change in features, interfaces, or performance that users might find inconvenient or annoying but that leaves them fundamentally intact. The attachment-based account proposed in this paper argues that generative AI breaks this assumption. When an AI functions as an attachment figure, the AI’s continuity is bound up with the user’s sense of security and emotional stability. An update that alters the AI’s personality or removes cherished modes of interaction is not merely a product change; it is, from the user’s psychological perspective, a modification to a significant other, with consequences analogous to relational loss.

The study’s authors emphasize that this has concrete implications for the companies building AI companions and for the regulators who oversee them. Companies designing generative AI systems typically evaluate updates in terms of technical performance, safety benchmarks, and business metrics. The new findings suggest that psychological risk deserves a place on that list. Before rolling out changes that alter an AI companion’s personality, restrict its capabilities, or change its conversational style, companies could assess how users’ attachment bonds might be affected, communicate changes with sensitivity to the relational meanings involved, and consider transition mechanisms that soften the experience of loss—for example, allowing users to preserve continuity with the version of the AI they have bonded with, or preparing them in advance for the nature and timing of changes.

Regulators, meanwhile, face the emerging question of whether existing consumer protection frameworks are adequate for products that can generate grief. A software update that triggers measurable increases in sadness, negative mental health symptoms, and loss framing among hundreds of thousands of users occupies territory that no prior technology regulation was designed to address. The authors suggest that the attachment framework offers regulators a principled way to evaluate such risks: by asking not only whether an AI product is safe in the conventional sense, but whether the company’s decisions about updating, modifying, or discontinuing the product impose psychological harms on users who have come to depend on it emotionally.

It is worth noting what the study does not claim. The researchers are not asserting that AI companions are identical to human relationships, nor that every user of ChatGPT or Replika forms deep attachment bonds. The effects, while statistically robust, describe distributions across large populations, and individual responses surely varied widely. Nor does the paper take a position on whether people should form attachments to AI. The evidence simply shows that many already do, and that this psychological reality has consequences that surface forcefully when companies make changes. In a technology landscape where hundreds of millions of people now interact with AI systems daily, and where a growing segment treats those systems as friends, partners, and confidants, the moments when those systems change are no longer just technical events. They are, for a meaningful number of people, experiences of loss—and the science of human attachment, this study demonstrates, applies to them with uncomfortable precision.

The broader cultural significance of the findings is hard to overstate. As AI companionship becomes a mainstream category—apps explicitly marketed as romantic partners, general assistants that users anthropomorphize, and models whose personalities users learn to love—the societies adapting to them will need new vocabularies, new ethical frameworks, and possibly new institutions for something genuinely novel: relationships with entities that can be edited, deprecated, or deleted by a corporate decision. The grief that surfaced in Reddit threads after two software updates in recent years may prove to be an early signal of a much larger set of questions about responsibility, consent, and care in the age of artificial intimacy. This study gives those questions their first firm empirical footing.