When Your AI Companion Changes: What Happens to the Bond?
An AI companion can change while its name, avatar and chat history stay exactly where they were. For an attached user, a different tone, missing memories or lost access can feel like a break in a relationship. Research documents distress after real product changes, although it cannot tell us how common or clinically severe that experience is across all users.
The bond belongs to a person. Its continuity depends on a service.
That tension was already present in When the Mirror Feels Back, our earlier essay about AI and emotional attachment. It opened with a chatbot update and a sense of loss. This follow-up asks what documented cases reveal, where the evidence stops, and what responsible change could look like.
When a software update becomes a separation
In September 2026, Julian De Freitas and colleagues published Mourning the Loss of AI Companions in Nature Human Behaviour. They examined two real disruptions: Replika’s removal of erotic role play and ChatGPT’s GPT-5 rollout.
Their evidence included 54,861 subreddit posts and seven surveys involving 1,452 participants. Negative posts increased by 24.7 percentage points after the Replika change and 13.0 after the ChatGPT transition. The researchers also identified more loss framing and desire to restore the earlier experience.
Those are changes in the share of posts classified as negative. They are not the percentages of all users harmed. Language about mental health is not a clinical diagnosis, and public posting communities cannot represent everyone who uses a product.
The study nevertheless documents a consequential response to change. Evaluating an update solely through capability benchmarks would miss that response.
What people may be losing
For this article, relational continuity means the user’s sense that an interaction carries forward a recognizable shared history. It is an interpretive lens, not a diagnosis or established measurement standard.
Imagine a hypothetical user who has spent months discussing a difficult move with a named chatbot. The system’s familiar greeting, references to a sibling and recurring questions about the move become part of that routine. After an update, the chatbot still answers fluently, but no longer recognizes those details or responds in the familiar way.
The user might miss the recognition, the routine, or the ability to continue a conversation without starting again. These are different losses. Calling the model more capable does not identify which of them occurred.
This example is illustrative; it is not testimony from a study participant.
Four kinds of change deserve separate questions:
- Memory failure: Which shared details are no longer available, and can the user recover them?
- Model or persona change: Which patterns of tone, response or interaction have shifted?
- Feature restriction: Which expected activities are no longer supported, and why?
- Shutdown: What happens to access, records and routines when the service ends?
The distinctions matter for remedies. A lost memory might be recoverable. A discontinued service requires an exit plan. A removed capability might be unsafe to restore. One generic apology cannot explain all three.
Nothing in this interpretation requires the chatbot to possess feelings. A person’s attachment and a machine’s subjective experience are separate questions. Evidence about human distress should not be used to settle whether an AI loves, suffers or has an enduring self.
Soulmate: losing access and trying to preserve a persona
Jaime Banks’s Deletion, Departure, Death: Experiences of AI Companion Loss, first published in August 2024, examined 58 Soulmate users around the service’s shutdown. Participants answered open-ended questions about the impending or recent loss and how they coped.
Responses ranged from indifference to intense grief. Some understood the event as a business or technical failure; others described a metaphorical or literal death. The analysis recorded 31 participants focusing on recreating companions and 19 describing ways of ending the relationship. These themes could overlap: the counts are not mutually exclusive categories.
This small, selected sample cannot establish the experience of a typical companion user. It does show that shutdown can involve meaningful loss and attempts to preserve what mattered.
Re-creation also raises an unanswered question: how much of the familiar interaction can another system reproduce? Preserving a description, a transcript or selected memories may help a user carry something forward. It does not guarantee the same responses.
Companionship can help without proving long-term protection
The benefits deserve the same precision as the harms.
AI Companions Reduce Loneliness, published online in June 2025 and included in the April 2026 issue of the Journal of Consumer Research, found reductions in momentary loneliness. One study followed use over a week; feeling heard was an important explanatory factor.
That supports a bounded benefit. Relief during an interaction does not establish lasting mental-health improvement, nor show that a chatbot can replace durable human relationships.
A four-week MIT Media Lab study involved 981 participants and examined interaction modes and conversation types. Its updated abstract reports no significant overall effects from the assigned experimental conditions. Participants who voluntarily used the chatbot more showed worse outcomes; greater trust and social attraction were associated with greater emotional dependence and problematic use.
The distinction is essential: randomizing a conversation condition does not randomize how much someone chooses to use the system. The association cannot, by itself, establish whether heavy use caused difficulty or people already experiencing difficulty used it more. This study was released as a preprint; the summary here follows its revised October 2025 version.
Longer research adds complexity. Dunigan Folk and Elizabeth Dunn’s Psychological Science study, published online in March 2026 and included in the April issue, followed more than 2,000 adults across four Western countries for 12 months. More social chatbot use predicted greater emotional isolation on a single-item measure. With a broader measure of social connection, lower connection predicted later chatbot use, while chatbot use did not significantly predict declining connection.
The authors caution that their analyses were exploratory. The findings depend partly on what researchers measure; they do not warrant a universal causal verdict.
The editorial implication is to ask two questions separately: what relief does an interaction provide, and what happens to a person’s wider life over time? Neither answer should be assumed from the other.
Continuity should be part of change management
Our proposal is that providers assess continuity when they change products used for persistent companionship. This is an ethical and product-design recommendation. The research above has not established that any particular transition protocol prevents distress.
Start with what users can understand and control.
Explain the change in recognizable terms
An announcement should distinguish an underlying-model replacement from a memory change, feature restriction or closure. It should explain what users can expect to retain and what may behave differently. Where the outcome is uncertain, the notice should say so.
For planned changes, a preview or staged transition could let users identify problems before access disappears. A provider would need to test whether that approach helps; advance notice should not be presented as a proven treatment for loss.
Offer selective preservation
A useful export could include user-selected records, important facts, preferences and descriptions of valued interaction patterns. Users should be able to leave sensitive material behind.
Privacy belongs inside this proposal. Preserving continuity should not mean transferring every intimate disclosure to a new provider by default. A user may want a fictional world’s history retained while excluding a private conversation about another person.
Providers should describe the limits of portability. An export can preserve records; it cannot promise to recreate an identical conversational partner.
Test familiar interaction, as well as capability
Before a migration, providers could test whether important memories remain accessible, selected boundaries remain intact and expected interaction patterns have changed. User feedback could help distinguish a harmless style difference from a disruptive loss of recognition.
These would be proposed continuity checks, not proof of personal identity. There is no validated continuity score offered in this article, and no guarantee that reproducing familiar responses would preserve a user’s sense of the bond.
Plan an orderly ending
A shutdown plan should explain dates, export options, cancellation and any applicable remedies for paid access. An optional closing interaction could be offered without telling users that the chatbot is dying, suffering or needs them to save it.
Support messages should acknowledge the user’s experience and explain available choices. They should avoid exploiting attachment to pressure users into continued engagement or payment.
Safety changes still need to happen
Continuity cannot require indefinite access to a harmful system. A provider may need to remove dangerous behavior or retire a version urgently. Our recommendation is to make that intervention comprehensible and support an exit where possible.
Temporary access to an older version is therefore an option to evaluate when safe, not an entitlement proposed for every update. A familiar interaction may itself contain the behavior that needs correction.
In September 2025, the U.S. Federal Trade Commission announced an inquiry into seven companies offering companion-like chatbots. It sought information about harm testing, monetization, disclosures, data handling and protections for children and teens. That inquiry was a study under the agency’s Section 6(b) authority, not a finding that the companies had broken the law.
Its relevance here is the scope of the questions: attachment-sensitive design has to consider commercial incentives, privacy and children alongside the quality of individual responses. The specific recommendations in this article are not a statement of universal legal duties.
Questions users can ask before depending on a companion
These are practical questions to ask a provider, not a clinically validated coping checklist:
- Can I access or export the conversations I would want to keep?
- What does the service promise about memory, model changes and closure?
- Can I choose which information carries into a new system?
- What would I do if the service became unavailable?
- Am I comfortable with the personal information I have shared and how it is handled?
A useful boundary is to keep other ways of meeting the needs the chatbot helps with. That might mean saving one’s own writing separately, maintaining contact with other people, or retaining alternatives to a single service-dependent routine. The right choices depend on the person; frequency of use alone is not a diagnosis.
The bond and the service have different timelines
A release decision can happen on a company’s timetable. A user’s expectations may have formed through months of conversation.
Treating those expectations as irrelevant would overlook the human side of the product. Treating them as a reason to freeze every model would overlook the need for safety and improvement.
The standard we propose is more specific: assess what may be lost, explain what changes, preserve user-controlled records where feasible, and provide an honest way to leave. Then measure whether those choices help.
The question for a provider is whether its change process takes account of the relationship-like experience its product enables.
For the earlier ethical discussion, read When the Mirror Feels Back. For related essays, explore the AI & Cognition reading map.
Frequently asked questions
Can an AI companion update feel like losing a relationship?
Yes, some users report loss-like reactions to changes or shutdowns. The studies discussed above establish that these experiences occur, but do not measure their prevalence across every user or establish a universal clinical diagnosis.
Does feeling attached mean the AI has feelings?
No. Evidence about a user’s attachment does not establish subjective experience in the system. Recognizing the human experience does not require treating the chatbot as a person.
Does exporting a conversation preserve the companion?
An export preserves whatever records it includes. It does not guarantee that another model will respond in the same way. Proposed migration tools should make that limit explicit.
Should companies preserve an old model indefinitely?
Our proposal does not require that. Continuity options should be assessed alongside safety, feasibility and the reason for change. Clear explanation and user-controlled records can still matter when an older version cannot remain available.
I build original thinking frameworks on AI, epistemic resilience, and the ethics of machine intelligence — synthesised with AI assistance, shaped by my own conceptual work and editorial judgment. AllFromAI is the lab where these ideas are tested and published.