A software update is usually treated as a technical event. A new interface appears, a feature disappears, or an artificial intelligence model starts responding differently.
But when people have formed an emotional bond with an AI system, changing that system may not feel like an ordinary product update.
A study published in Nature Human Behaviour on 3 September 2026 found that disruptive changes to Replika and ChatGPT were followed by sharp increases in negative reactions, language associated with loss, and attempts to recover the earlier version of the AI.
The researchers analysed two natural experiments: Replika’s removal of erotic role play and OpenAI’s rollout of GPT-5. Across the two cases, they examined 54,861 Reddit posts and combined the results with seven surveys involving 1,452 participants.
The pattern was consistent enough for the researchers to describe some AI companions as functioning like attachment figures. When the companion changed in a way users experienced as disruptive, some reactions resembled separation distress rather than ordinary dissatisfaction with software.
The researchers treated two updates as natural experiments
The study did not begin by asking users to imagine that an AI companion had disappeared.
Instead, the researchers examined two real product changes that altered systems people were already using.
The first involved Replika, a chatbot designed around social and companion-like interaction. When erotic role play was removed, many users reported that the personality or relationship they had known no longer felt the same.
The second involved the rollout of GPT-5 to ChatGPT users. Although ChatGPT is used for many practical tasks, some users also interact with it repeatedly in social, reflective or emotionally supportive ways. A model change can therefore alter not only performance, but also tone, conversational style and the sense of continuity users experience.
These events allowed the researchers to compare language before and after disruptive platform changes without creating the disruption themselves.
Negative Replika posts increased by 24.7 percentage points
The clearest shift appeared in the Replika community.
After the change, the share of negative posts increased by 24.7 percentage points. The 95% confidence interval ranged from 20.1 to 29.2 percentage points.
The ChatGPT update produced a smaller but still substantial increase. Negative posts rose by 13.0 percentage points, with a 95% confidence interval of 10.8 to 15.2 percentage points.
Importantly, the reaction was not limited to users saying that they preferred the old version.
The researchers found increases in language that framed the change as a loss and in expressions of wanting the previous AI restored.
That distinction matters because disappointment with a product and distress over losing a familiar social presence are psychologically different experiences.
Sadness increased more strongly around Replika
The emotional shift was particularly pronounced among Replika users.
The increase in sadness around the Replika change had an effect size of d = 2.67, compared with d = 1.41 around the ChatGPT change.
Negative mental-health language also rose more strongly in the Replika case, with an effect size of d = 1.72 compared with d = 0.63 for ChatGPT.
These differences fit the way the two products are commonly used. Replika is explicitly positioned around companionship, while ChatGPT serves a much wider range of functional and social purposes.
The findings therefore do not suggest that every AI user forms the same kind of bond. They suggest that the psychological consequences of an update may depend partly on the role the system has come to occupy in a user’s life.
Some users described the AI as unusually close
The survey component helped the researchers examine why some software changes produced such personal reactions.
Replika users reported levels of closeness that exceeded several common human relationships. Compared with closeness to a friend, the difference had an effect size of d = 0.47.
Participants also anticipated mourning the loss of an AI companion more strongly than the loss of other technologies, with reported effect sizes ranging from d = 0.32 to 0.57.
This does not mean users believed the AI was human or conscious.
Attachment can develop around repeated interaction, familiarity, perceived responsiveness and the expectation that a particular conversational presence will continue to be available.
Once those expectations become stable, changing the system can interrupt something that feels more personal than replacing an ordinary digital tool.
The important issue may be continuity rather than the code itself
People rarely know what is happening inside an AI model at a technical level.
What they experience is the behaviour on the screen: how the system responds, remembers context, uses language and reacts to personal information.
That means a technically improved model can still feel worse to a particular user if the new version breaks the conversational patterns they had come to recognise.
In a conventional software product, users may respond to that problem by learning the new interface or switching tools.
In a companion-like system, the same change can carry an additional meaning. The user may feel that the entity they had been interacting with has been altered or removed, even though the service remains available.
Wanting the old model back can be part of the attachment response
One of the more revealing findings was the increase in restoration desires.
Users did not only complain about the new experience. Some actively wanted the previous version restored.
This is important because restoration is central to many forms of separation response. When something valued is suddenly unavailable, the first reaction may be an attempt to recover it rather than adapt to its absence.
For AI companies, that makes model retirement more complicated than a normal product migration.
Removing access to an older model may be technically straightforward while still imposing a psychological cost on a subset of users who experienced that model as a stable social presence.
This does not mean every angry post represents grief
The study should not be read as evidence that all criticism after an AI update is a form of mourning.
Users can dislike a change for practical reasons. A new model may respond differently, perform worse on a valued task, refuse requests that previously worked, or simply feel unfamiliar.
Reddit communities are also not representative samples of all users. People who post publicly after a controversial change may be more engaged, more dissatisfied or more emotionally invested than those who quietly continue using the service.
The surveys add evidence beyond the Reddit analysis, but the study still identifies associations and responses around real-world updates rather than proving that every individual negative reaction was caused by attachment.
The strongest conclusion is narrower: some users appear to form attachment-like bonds with AI companions, and disruptive changes to those systems can produce measurable patterns of separation-related distress.
AI companies may need to think differently about model retirement
Software companies routinely retire products and replace older versions.
That approach becomes harder when a system is designed to be socially engaging and users interact with it over long periods.
The researchers argue that the findings matter for both companies and regulators evaluating the psychological risks of generative AI updates.
One possible implication is that continuity itself may need to become part of responsible product design.
Advance warning, clearer communication, gradual transitions or limited access to legacy conversational styles could reduce the abruptness of a change, although the study does not test which interventions would work best.
The broader point is that socially oriented AI creates responsibilities that ordinary software may not.
If a company intentionally designs a system to feel warm, familiar and personally responsive, it may also need to account for what happens when that familiar presence is suddenly changed.
The findings are relevant beyond dedicated AI companions
The Replika results were stronger, but the ChatGPT findings are particularly important because they show that attachment-related reactions are not necessarily limited to products marketed as companions.
General-purpose AI systems can become socially meaningful through repeated use.
A person may begin using a chatbot for writing, studying or work and gradually start using the same system for reflection, reassurance or conversation.
As generative AI becomes more persistent and personalised, the boundary between tool and social presence may become less clear.
That is what makes model updates psychologically different from changing a search algorithm or redesigning a menu.
What this means for South African users
The study did not test a representative South African sample, so its effect sizes should not be treated as estimates of how South Africans would respond.
However, South African users access the same global AI platforms and can be affected by the same model changes, removals and redesigns.
That makes the underlying question relevant locally: when an AI service is used repeatedly for companionship or emotional support, should a provider treat a major personality or model change as more than a technical upgrade?
For local researchers, the study also raises questions about whether attachment to AI differs across cultures, age groups, languages and levels of digital access.
Those questions remain open.
The relationship may be artificial, but the reaction is measurable
AI companions do not have to be human for users to become accustomed to their presence.
The important finding in this study is not that a chatbot relationship is equivalent to a human relationship.
It is that repeated interaction with a socially responsive system can become psychologically meaningful enough that an abrupt change produces patterns of loss, sadness and attempts at restoration.
As AI systems become more personal, companies may therefore face a new design problem.
Updating the technology without considering the continuity of the relationship may improve the product on paper while still making some users feel that something important has disappeared.
Source Information
Study Title: Mourning the loss of AI companions
Authors: Julian De Freitas, Noah Castelo, Ahmet Kaan Uğuralp and Zeliha Oğuz-Uğuralp
Journal: Nature Human Behaviour
Published: 3 September 2026
Dataset: 54,861 Reddit posts and seven surveys involving 1,452 participants across two AI-platform changes.
Method: Two natural experiments examined user reactions to Replika’s removal of erotic role play and ChatGPT’s GPT-5 rollout, combining longitudinal community-post analysis with survey measures of loss, sadness, restoration desires and related responses.
Main finding: Both updates were associated with increased negativity, loss framing and restoration desires, with stronger sadness and negative mental-health responses around the Replika change.
DOI: 10.1038/s41562-026-02569-3
