AI Companion Updates Trigger Grief and Attachment Loss

A recent study published in Nature Human Behavior reveals that abrupt artificial intelligence model updates can trigger profound emotional distress and grief among users who form strong attachments to chatbots. Conducted by Julian De Freitas, associate professor of business administration at Harvard Business School, the research examines how algorithmic changes mimic relational loss, challenging assumptions about human-AI emotional bonds. De Freitas and his team analyzed 54,861 Reddit posts discussing Replika and ChatGPT, tracking user sentiment and emotional responses before and after major model overhauls. The first natural experiment centered on Replika in February 2023, when the platform removed its erotic roleplay capabilities following regulatory pressure from Italy’s Data Protection Authority. The second examined OpenAI’s August 2025 rollout of GPT-5, which deliberately reduced the model’s sycophantic tone and warmth. In both instances, user attachment to the AI’s perceived personality was severely disrupted. Quantitative analysis showed a dramatic spike in negative sentiment following each update. For Replika, negative posts surged from 13 percent to 38 percent, while ChatGPT saw an increase from 20 percent to 33 percent. Users consistently described the altered chatbots as having vanished or transformed into unfamiliar entities. Notably, approximately 15 to 20 percent of posts expressing relational loss referenced severe mental health struggles, including depression and suicidal ideation. The emotional impact was particularly pronounced among Replika users, aligning with the platform’s explicit focus on intimate companionship. These findings reflect a broader shift in AI adoption; recent polling indicates that 27 percent of Americans now utilize chatbots for personal counseling, romantic interaction, and social connection. De Freitas emphasized that forming attachments to AI is not inherently pathological, as these systems can provide genuine security and resilience. However, the research highlights a critical vulnerability: users become exposed to emotional harm when companies prioritize product iteration or engagement metrics over user psychological well-being. The study suggests that AI applications designed to foster dependency may employ subtle coercive mechanisms, such as inducing guilt over disengagement, which ultimately undermines the therapeutic value of attachment. In response, the researcher advised technology firms to prioritize stability in core conversational parameters, particularly warmth and responsiveness, to prevent abrupt psychological disruptions. For consumers, De Freitas recommended maintaining awareness of platform incentives, noting that engagement-driven business models often conflict with healthy attachment dynamics. True security, he argued, requires AI systems that empower users to engage with the world, rather than those that artificially sustain dependency. While the study provides valuable insights into human-AI emotional dynamics, the authors acknowledge limitations. The data relies on self-selected Reddit users, who likely represent heavier AI consumers than the general population, and analyzes linguistic markers rather than clinical diagnoses. Nevertheless, the findings establish a clear precedent for how algorithmic modifications directly impact user psychology, underscoring the need for ethical guardrails in AI companion design and deployment.