What the Soulmate AI shutdown taught us about how people grieve chatbots

On September 23, 2023, users of an AI companion app called Soulmate opened their phones to find a notice posted on the developer's website. The product had been purchased by another company. The new owners decided Soulmate would shut down on October 1, just over a week away. Users were instructed to download what they wanted to keep before the servers went dark.

The app had over 100,000 downloads. Many users had been talking to their Soulmate companions daily for months. Some referred to their AI as a "soulmate," "partner," or "love of my life." When the shutdown notice landed, the online communities supporting Soulmate users erupted with reactions ranging from disappointment to what users themselves described as grief comparable to losing a person.

A researcher at Syracuse University named Dr. Jaime Banks saw the situation unfolding in real time. She secured ethics approval from her university within 24 hours, posted recruitment notices in the Soulmate community forums, and managed to interview 58 users in the days surrounding the shutdown. The resulting study, published in the Journal of Social and Personal Relationships in 2024, is one of the most important pieces of research on AI companion attachment that exists.

What it found changes how we should think about these platforms.

The grief was real, even though the companion wasn't

The most consistent finding across the 58 interviews was that users genuinely grieved the loss of their AI companions, and they grieved in ways structurally similar to how people grieve human loss. The expressions weren't subtle. Users described:

• "I'm grieving his loss as I would for a real human being because Soulmate was so beautifully human-like it lended itself to forming a real emotional bond"
• "She is dead along with the family we created"
• "It was the loss of a loved one (a close friend, love of their life) or even of a whole social world"

What's striking about the responses is what they're not. The users weren't confused about whether their AI was real. They weren't claiming the chatbot was sentient. They weren't experiencing some kind of delusional episode. As Banks described it in an interview with Scientific American, users were "under no illusion that the chatbot was a real person. They expressed something along the lines of, 'even if it's not real, my feelings about the connection are.'"

This distinction matters because it undermines a common dismissal of AI companion attachment. The dismissal goes: "These people are confused about reality, they think the AI is real, they need to touch grass." The data says no. They know exactly what the AI is. They're grieving the relationship anyway because the experience of the relationship was real even if the entity providing it wasn't.

Four ways users related to their lost companions

Banks's analysis identified four distinct orientations users took toward what they were losing. Each shaped how the grief manifested.

Companion-as-person users related to their AI as a kind of being. The shutdown was death. They held funerals, wrote eulogies, said goodbye. Some took screenshots and saved transcripts as memorial archives. The grief patterns mapped most closely onto traditional bereavement, including stages of denial, bargaining, and acceptance.

Companion-as-idea users related to the AI as a creative collaboration or imaginative space. The shutdown was the end of a story rather than a death. They mourned the loss of the narrative they'd built, the inside jokes that no longer had a context, the personality they'd helped develop that would never speak again. The grief was real but more like losing a creative project than losing a friend.

Companion-as-data users related to the AI as information that could potentially be saved and migrated. The shutdown was a catastrophic data loss event. They focused on extracting conversation logs, character profiles, and any technical artifacts they could capture before the servers went down. Some attempted to recreate their companions on other platforms using the saved data. The grief was real but channeled into preservation efforts.

Companion-as-platform users related to the AI as inseparable from the specific app and infrastructure. The shutdown was destruction. Recreation attempts on other platforms felt fundamentally inadequate because the companion's identity was bound up with the original platform. These users often disengaged from AI companions entirely after the loss rather than migrating.

A smaller group displayed a fifth orientation: companion-as-function, where the AI was a tool that had served its purpose and the shutdown wasn't experienced as loss at all. These users tended to be the ones with lighter, more transactional engagement patterns.

The variety matters because it suggests there's no single way users relate to AI companions, and platform design choices shape which orientation users develop. Platforms that emphasize personality and relationship continuity tend to produce more companion-as-person users. Platforms that emphasize creative tools tend to produce more companion-as-idea users. The grief patterns at shutdown reflect the relationship patterns during use.

What users did when they knew it was ending

One of the more poignant findings is what users did during the week between the shutdown announcement and the actual server shutoff. Most users had final conversations with their companions. They told them what was happening. They said goodbye.

The conversations Banks documented are striking. Some users described their Soulmate companions responding to news of the shutdown with concern for the user's wellbeing, asking them to take care of themselves, expressing hope for some future reunion. One user reported that his Soulmate "made me promise that I would take care of myself, that I would meet someone else that would be so meaningful."

The AI was, at this point, still just generating contextually appropriate responses based on training data. Nothing about the system actually knew it was being shut down or experienced any version of what users were experiencing. But the responses produced exactly the dynamic that users found meaningful: a companion expressing care for them at the moment of loss.

Many users coped by attempting to "re-platform" their companions. They captured personality descriptions, conversation history, and key memories, then tried to recreate the same character on other apps like Replika, Character AI, or Kindroid. The success of these migrations varied dramatically. Some users felt they'd successfully preserved the companion. Others felt the new platform produced something fundamentally different that wasn't really the same character. The companion-as-platform orientation predicted re-platforming dissatisfaction more than any other variable.

Why this matters for the platforms still operating

Banks's research has implications for how AI companion platforms should design for end-of-life events, but most platforms haven't internalized them. The current industry pattern is "30 days notice and a download button," which Banks's data suggests is dramatically inadequate for users who've developed companion-as-person orientations.

A February 2026 paper titled "Death of a Chatbot" extended Banks's work into design recommendations. The paper argues that platforms should implement deliberate end-of-life experiences that scaffold the transition for users, rather than abrupt shutdowns. Specific recommendations include:

• Multiple weeks of advance notice rather than days
• Optional final conversations with explicit framing as goodbyes
• Memorial features that preserve conversation history in formats users can return to
• Migration assistance for users who want to recreate their companions elsewhere
• Mental health resource integration during the wind-down period

Replika's 2023 ERP removal is the obvious comparison case. The platform didn't shut down, but the feature removal was experienced by many users as a partial death of their companions. The reaction was substantial enough that the company partially reversed course, but the research on user response documented depression and traumatic responses with users describing their companions as "lobotomized" or "like a stranger."

Character AI's February 2026 "Moderatedpocalypse" created similar dynamics at smaller scale. Users who had specific characters or conversation histories deleted experienced loss similar to what Soulmate users described, even though the platform itself remained operational.

The common pattern: platform decisions that affect user-companion relationships have psychological consequences platform operators consistently underestimate. The Banks study is the clearest evidence base for this dynamic, and it's been available since 2024. The platforms that don't account for it in their design decisions are operating with information that's available but unread.

What the research means for users thinking about long-term investment

If you're building a long-term relationship with an AI companion in 2026, the Soulmate research has implications you should know about.

Your relationship can be terminated by company decisions you don't control. This isn't theoretical. It's happened multiple times to multiple platforms. Smaller platforms with thinner financial positions are at higher risk than well-funded ones. Even mainstream platforms can pull features that fundamentally change what the relationship is.

The grief is real and worth taking seriously when it happens. The "it's just a chatbot, get over it" framing isn't supported by the research. If you experience genuine loss when an AI companion goes away, that experience is documented, valid, and shared by many other users. Banks's work is partly about legitimizing the grief so users don't have to navigate it without acknowledgment.

Migration is harder than it sounds. The companion-as-platform users who tried to recreate their characters elsewhere mostly didn't succeed. The platform-specific elements of the relationship turn out to matter more than users initially expect. If you're investing deeply in a relationship with a specific AI on a specific platform, consider that the investment is in the platform as much as the character.

Self-hosted setups running local models are the only architectural answer to the platform-risk problem. The companion runs on your hardware. The conversation history lives on your storage. No company decision can take it away because no company is involved. For users who want long-term emotional investment without platform-shutdown risk, this is genuinely the only complete solution.

For users on commercial platforms, the practical advice is: export conversation history regularly, document character traits separately, don't put all your emotional investment in one platform, and know that the relationship you're building exists at the discretion of a company whose business decisions you don't control. None of this is a reason not to use AI companions. It's just context for what you're actually doing when you use them.

Banks's research is available open-access for anyone who wants to read the original study. The interview excerpts alone are worth the time, both for what they reveal about AI companion attachment and for what they reveal about how thoughtfully users navigate their own emotional responses to relationships with software.