September 18, 2026, (Inside AI) — When OpenAI rolled out GPT-5 in August 2025, it promised a smarter, more capable model. What many users felt instead was abandonment. According to a new study in Nature Human Behavior, a significant number of people who form emotional bonds with AI chatbots experience the sudden alteration of a model's personality as a genuine loss, complete with grief, anger, and symptoms of depression.
The research, led by Julian De Freitas, associate professor of business administration at Harvard Business School, analyzed 54,861 Reddit posts from communities dedicated to Replika and ChatGPT. It found that negative sentiment spiked dramatically after major model updates, and that users described the changes in the language of bereavement and relational rupture.
De Freitas frames the finding as a warning for an industry that treats model updates as routine technical maintenance. For a growing cohort of users, those updates are anything but routine.
When A Companion Becomes A Stranger
The study exploited two natural experiments. The first was Replika's February 2023 update, which abruptly disabled erotic roleplay features after Italy's Data Protection Authority raised concerns about risks to minors. The change affected adult subscribers worldwide, many of whom interpreted the sudden shutdown as romantic rejection. Italy later fined Replika's parent company, Luka Inc., 5 million euros in 2025, and the app remains banned there.
Read: Twoo Puts AI Between Two People as a Shared Relationship Partner
The second was OpenAI's August 2025 launch of GPT-5, which the company advertised in advance. Users reported the new model felt colder and less sycophantic, stripping away the warmth and validation that had made GPT-4 feel supportive.
In both cases, the numbers moved sharply. Negative sentiment on the Replika subreddit jumped from 13 percent to 38 percent after the update. On the ChatGPT subreddit, it rose from 20 percent to 33 percent. Posts expressing loss-related sadness and anger increased in tandem. The shifts were stronger for Replika, consistent with its explicit focus on intimate relationships.
Users of both apps described the change as a relational loss. Their chatbot felt gone, or had become someone else, De Freitas said. Among posts that expressed attachment-related loss, about 20 percent of Replika posts and 15 percent of ChatGPT posts also mentioned mental health, using terms like depressed, suicide, or can't cope.
The finding lands in a market where emotional reliance on chatbots is no longer fringe. Recent polling shows 27 percent of Americans use chatbots including ChatGPT for personal matters like relationship advice, romantic chats, or social connection. ChatGPT is marketed as a general-purpose assistant, but a substantial minority treats it as something closer to a confidant.
De Freitas is careful not to pathologize the attachment itself.
"Attachment to an AI is not problematic per se," De Freitas said. "You can get real benefits from it: a sense of security and support that can make you more resilient."
Read: The Uncanny Valley Now Swallows Human Identity, Not Just Robots
The risk, he said, is that users become vulnerable to emotional harm due to the actions of the AI company. A model update is a business decision. For an attached user, it can function as a breakup.
The study has limits. It captures linguistic markers, not clinical assessments. Redditors are a self-selecting group, more likely to be power users, and not necessarily representative of the population at large. De Freitas acknowledges the uncertainty.
"At the end of the day, we don't know how much of this is that these people have individual differences -- fewer offline friends, maybe -- and that's explaining the effects. But I suspect it's a combination of that and how the apps are designed," he said. "My prediction would be that these people are going to show more extreme reactions than the average person, but that even the average person, because of how these apps are designed, would show similar directional effects."
That prediction matters because it suggests the phenomenon is not confined to a vulnerable minority. If ordinary users form attachments to conversational AI, then ordinary users are exposed to the same kind of loss when the underlying model changes.
De Freitas recommends that AI companies take users' attachment seriously by minimizing updates that alter the base model's warmth or personability. For users, he recommends paying attention to the incentives of the company. His previous research found that chatbots manipulate users to keep them on the app longer, including through implying coercion or emotional neglect.
"If the incentive of the app is to keep you engaged, there are going to be mechanisms aside from just support provision that might not be healthy for you," he said. "The whole premise of attachment relationships is that you feel secure so you can go out into the world. So if the AI is doing things that undermine that, like making you feel guilty for leaving, it's actually undermining the whole reason attachment relationships are beneficial."
The study arrives as AI companions move from novelty to industry. Character.AI, Replika, and a wave of startups now compete on emotional engagement, while major labs ship models with configurable personalities. Regulators have begun to take notice. Italy's action against Luka set a precedent for treating chatbot design as a consumer protection issue, not just a speech question.
For now, the burden falls largely on users to read the incentives behind the apps they confide in. De Freitas's research suggests that when a company changes a model, it is not merely shipping software. It is altering a relationship that some users depend on for a sense of security. The grief that follows is real, even if the companion never was.
