This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: Artificial intelligence systems, particularly those powered by large language models and designed to communicate with users on a one-to-one basis, are increasingly used worldwide. Some of these systems, such as Replika, are also marketed as "AI companions" that can engage socially with humans and use emotional language.
Users who regularly interact with AI systems could gradually start to view them as a steady presence in their lives or even as attachment figures that bring them comfort, security or reassurance. While this possibility has been widely debated in recent years, the nature and psychological consequences of these bonds remain poorly understood. Researchers at Harvard Business School, Harvard University and the University of Alberta recently tried to better understand how users react to changes in how an AI communicates with them.
The findings of their surveys and natural experiments, published in Nature Human Behaviour, suggest that sudden company-initiated updates affecting an AI system's apparent "personality" can sometimes elicit anxiety, sadness, distress and other challenging feelings typically associated with breakups or grief. "My lab had been following Replika's Reddit communities and noticed that users were complaining that they had 'lost' their companions or that they felt these companions had been 'lobotomized' after an app update, so I decided to take a closer look," Julian De Freitas, first author of the paper, told Tech Xplore. "What struck me was that people were using the language of bereavement, not the language of customer complaints.
That pointed to a specific idea: that AI companions may function as attachment figures—entities people want to stay close to, turn to for reassurance when distressed, and rely on to function elsewhere in their lives." De Freitas and his colleagues hypothesized that if users become emotionally or psychologically attached to AI agents, a software update that influences how the agents interact with them may not just annoy them but also produce separation distress. The term separation distress is typically used to describe the sadness, anxiety, loneliness and emotional discomfort that people can feel when losing access to an important attachment figure. Earlier work in this area was largely theoretical.
The researchers instead wanted to test their hypothesis experimentally. To do this, they conducted two natural experiments and seven surveys. "Our natural experiment focused on two real updates that users experienced as changing the AI's personality: Replika's abrupt, unannounced removal of erotic role-play in February 2023, and OpenAI's pre-announced GPT-5 rollout in August 2025," explained De Freitas.
"We collected all 54,861 public posts on the relevant subreddits for the 30 days before and after each update, then used an LLM to code each one for sentiment, emotion, whether the user said the AI had changed or been lost, whether they described losing a friend or partner, whether they mentioned mental health concerns, and whether they wanted the old version back." After collecting this large pool of data, two of the authors hand-coded 350 posts to validate the classifications made by the LLM. Subsequently, they ran seven surveys involving a total of 1,452 respondents to assess how people perceived or anticipated the loss of an AI companion. Discover the latest in science, tech, and space with over 100,000 subscribers who rely on Phys.org for daily insights. d research that matter—daily or weekly.
Extract — continue reading at the source.
