Negative posts nearly tripled on Replika's subreddit after an unannounced update

The hypothesis under test was narrow. If people bond with a conversational agent the way they bond with a person, then an update that changes how it talks should produce separation distress: the sadness, anxiety and loneliness of losing access to an attachment figure ^[12]. Earlier work on AI attachment was largely theoretical, so Julian De Freitas and his co-authors went looking for updates that had already happened ^[13][20]. His lab had been reading Replika's Reddit communities, where users wrote that their companions had been "lobotomized" after an app update ^[15]. "What struck me was that people were using the language of bereavement, not the language of customer complaints," De Freitas told Tech Xplore ^[10].

Advance notice was not the only thing separating the two cases. Replika took a specific capability away from a product marketed as a companion that uses emotional language ^[14][3]. OpenAI replaced the model under a general-purpose assistant ^[3]. The baselines differed as well: the ChatGPT subreddit sat about 7 percentage points higher in negative posts before its update than Replika's did before its own ^[18]. One swing was 25 percentage points, the other 13 ^[17]. Notice, product category and community are bundled together in a comparison of two events, and I don't think the GPT-5 figures show that pre-announcing a change reduces distress. The coding does establish that the content of the posts changed after each update, and one of the tracked categories was whether the poster wanted the old version back ^[4].

The validation is thinner than the size of the corpus suggests. Two of the authors hand-coded 350 posts against the model's labels, about 0.6% of the 54,861 ^[5][19]. The unit throughout is the post, not the person, and the measured population is people who write publicly about the product on Reddit ^[4]. The phys.org account gives sentiment shares and survey sample sizes; it does not include subscription or churn figures ^[22].

The survey arm asked about a loss that had not happened. Active Replika users imagined losing their companion, then losing a favorite app, a game character, a voice assistant, a car, a brand, and a pet, which De Freitas called a deliberately tough comparison ^[9]. Anticipated loss and experienced loss are different measurements, and the surveys measure the first. What they add is who: "the users most affected were the ones who had come to treat the AI as a relationship rather than a product," De Freitas said ^[11].