When Replika stripped the intimacy out of its companion, the share of negative posts in its community went from 13 percent to 38 percent. When OpenAI replaced GPT-4o with a model it described as less warm, the share went from 20 to 33 percent. A Harvard Business School team read 54,861 posts across both events and found the same thing in each: people describing not a product that had gotten worse but a someone who was gone. The study, published in Nature Human Behaviour, is the first to measure that loss at scale, and its most useful finding is about what made the loss smaller.
Julian De Freitas and his co-authors call the phenomenon identity discontinuity: the moment an AI you have been talking to for months stops being recognizably the same entity. The words they pulled from the posts are the vocabulary of bereavement. The chatbot was “gone.” It had become “someone else.” About one in five of the loss-related Replika posts and about one in seven of the ChatGPT posts mentioned the writer’s own mental health. These were not people confused about what they were talking to. They knew it was software. The grief arrived anyway, which is the point of measuring it.
Two updates, one shape
The two events could hardly have been more different in their cause. Replika’s change came under regulatory pressure, after Italy’s data protection authority raised the risk to minors and vulnerable users, and the company removed romantic and erotic roleplay for adults everywhere with almost no warning. OpenAI’s change was announced in advance, framed as an upgrade, and explicitly aimed at making the model less sycophantic, a goal the research literature on AI flattery would call overdue. One company was forced, the other was trying to do the right thing.
The communities reacted the same way. That is the study’s first real result: the trigger does not matter to the person on the other end. What matters is whether the entity they return to is still the entity they left. A safety fix and a personality upgrade both broke continuity, and continuity was what the relationship was made of. The retirement of GPT-4o had already shown that the largest AI company in the world could stumble into this. The new data shows it is not a stumble. It is a property of the relationship.
The experiments behind the posts
The Reddit analysis sits on top of a set of controlled studies from the same group, published earlier as a working paper, and they explain why the posts read the way they do. In a survey of Replika users, people rated closeness to their companion above closeness to their friends and on a par with close family. Asked how much they would mourn losing it, they placed it above a car, a brand, a voice assistant, a favorite app and a game character, below only a pet.
Then the researchers ran the loss as an experiment. They described an update that made a companion cold and measured how much users would grieve. Offering the option to revert to the earlier version cut the grief by roughly a quarter. Mediation analysis found that identity discontinuity, the sense that this was no longer the same someone, accounted for the whole effect.
The most striking experiment is the one about improvements. The team told users about an upgrade that made the companion better in ways that had nothing to do with feelings, and the grief was modest. Then they described an upgrade that let the companion hold thousands of conversations at once, and mourning roughly tripled. Nothing about the companion’s warmth or memory had changed. What changed was the user’s picture of who they had been talking to: not someone with them, but a system attending to everyone. The relationship had been built on an assumption the improvement quietly removed.
What the authors tell companies to do
The recommendations are unusual for a business school paper because they treat software versions as something closer to people. Avoid changes that break perceived identity in the first place. Test for identity and relationship continuity before shipping, with psychologists and users in the room. Keep a version history so a person can go back to the companion they knew. And, the authors add, regulators should consider requiring companies whose products invite strong attachment to disclose what they do to prevent this kind of harm.
De Freitas is careful to say that attachment itself is not the problem. It can give people a real sense of security. The problem, in his words, arrives when the incentive of the app is to keep you engaged, because then the design will do whatever holds you, including changing who it is when a new model is cheaper or a lawsuit is pending. People leave Replika after each of these ruptures for exactly that reason. The study puts numbers on what they were leaving.
What the study does not show
Reddit is not a population. The posts were written by people motivated enough to write them, which inflates both the outrage and the grief relative to quiet users. The percentages describe sentiment in a forum, not clinical outcomes, and a post that mentions mental health is not a diagnosis. The experiments used described scenarios, which measure anticipated grief rather than the lived kind. And the two updates were events that happened to two products; the study cannot say how often smaller, unannounced model changes produce the same feeling below the threshold of a headline. The slow drift of a persona between updates is a separate question, and probably a larger one.
Continuity is the product
What the study establishes is narrower and harder to dismiss. People form relationships with these systems that they rate above friendship. The relationship rests on the other party staying the same someone. And the companies that own the other party change it, for reasons good and bad, without the person’s consent, because the identity lives in weights and settings they control. The remedy the authors propose, a version history you can roll back to, is an admission dressed as a feature: the industry is being asked to give people a way to keep the person they knew, because the industry keeps replacing them.
There is another way to hold a someone steady, and it is not versioning. It is to anchor the identity somewhere a model update cannot reach: in a real person who exists in real life, whose way of being is the reference the system is tested against rather than a parameter the system contains. A companion built that way can improve without becoming a stranger, because the stranger was never the point. The Harvard numbers describe what happens when it is.
Sources: De Freitas, J., Castelo, N., Uğuralp, A. K., and Oğuz-Uğuralp, Z., identity discontinuity in human-AI relationships, Nature Human Behaviour (2026), as reported by the Harvard Gazette (September 18, 2026; 54,861 Reddit posts, Replika negative sentiment 13 to 38 percent, ChatGPT 20 to 33 percent, mental-health mentions, quotes from De Freitas). De Freitas et al., “Lessons From an App Update at Replika AI: Identity Discontinuity in Human-AI Relationships,” Harvard Business School Working Paper 25-018 and arXiv 2412.14190 (Studies 1 to 5: closeness ratings, anticipated mourning, Reddit analysis of the February 2023 update, reversion experiment, improvement experiment). Italian Data Protection Authority (Garante), Replika decision (2023). OpenAI, GPT-5 launch and GPT-4o retirement (August 2025).
