Updated 2026-09-06
from the news desk
Do people grieve AI companions? New paper finds mourning after disruptive updates
A Nature Human Behaviour paper published on 03 September 2026 reports large-scale evidence that some users respond to AI companion changes with attachment-related loss. The findings sharpen debate over update and continuity safeguards…
Yes. A Nature Human Behaviour article by Julian De Freitas and colleagues, published on 03 September 2026, reports that disruptive updates to Replika and ChatGPT were followed by loss framing, sadness and desires to restore the earlier system, which the authors analyse as attachment-related distress. The paper studies users' reactions to AI companion changes, not whether the AI itself has feelings or moral status.
What Nature Human Behaviour published
Nature Human Behaviour published Julian De Freitas, Noah Castelo, Ahmet Kaan Uğuralp and Zeliha Oğuz-Uğuralp’s article Mourning the loss of AI companions on 03 September 2026. The article page says the study’s data and analysis code are available via GitHub, that Harvard Business School funded the research, and that the authors declared no competing interests.
The abstract opens with a broad claim about the setting: “As artificial intelligence (AI) shifts from tool use to social interaction, company-initiated updates can disrupt user well-being.”
De Freitas and colleagues then set out the mechanism they want to test: “Here we develop an attachment-based account in which AI companions function as attachment figures such that disruptive updates are associated with separation distress.” The paper describes Replika’s erotic role play removal and ChatGPT’s GPT-5 rollout as two natural experiments, meaning product changes the authors analysed after they occurred. The abstract states: “Across two natural experiments—Replika’s erotic role play removal and ChatGPT’s GPT-5 rollout—we analysed 54,861 posts from Replika and ChatGPT subreddits and analysed data from 1,452 participants across seven surveys.”
What the study found
The headline quantitative finding is that both changes were followed by more negative language, more discussion of loss and more requests to get the old behaviour back. The abstract states: “Both updates increased negativity, loss framing and restoration desires (Replika negative posts, +24.7 percentage points, 95% CI 20.1 to 29.2; ChatGPT, +13.0 percentage points, 95% CI 10.8 to 15.2), with larger Replika (versus ChatGPT) increases in sadness ( d = 2.67 versus 1.41) and negative mental health ( d = 1.72 versus 0.63).”
De Freitas and colleagues also report unusually strong attachment among Replika users. The abstract says: “Replika users reported closeness exceeding common human ties (versus friend d = 0.47) and anticipated mourning higher than other technologies ( d = 0.32–0.57).” In plain English, the paper reports that some users rated their bond with the chatbot as closer than a friend and expected more mourning than they would for other technologies.
What the paper does and does not show
One reading is that the paper gives a direct, evidence-based answer to the question do people grieve AI companions. In the two product changes the authors analyse, users did not only complain about a missing feature; the paper records patterns the authors interpret as attachment-related loss, sadness and restoration desire across subreddit posts and surveys.
The case for taking continuity seriously rests on the design De Freitas and colleagues chose. Because the paper studies real product updates and combines text analysis with seven surveys, one steelman view is that companies now have published evidence that abrupt changes can affect user well-being when a system is used for social interaction.
A narrower reading is also available from the same paper. De Freitas and colleagues study users’ reactions to changed systems, not whether the systems themselves feel anything. That means the article can support stronger model deprecation or update safeguards while leaving open the separate questions of whether an AI is sentient or counts as a moral patient.
Why the finding matters for AI companion policy
The authors close the abstract with a policy-facing statement: “These findings inform how companies and regulators evaluate psychological risks of generative AI updates.”
One steelman policy response is to treat disruptive updates to companion systems as a user-welfare issue that may justify more notice, continuity planning or rollback options. Another steelman response is narrower: the paper identifies a psychological risk to users, but it does not by itself establish rights, personhood or consciousness for AI systems.
Either way, the dated development is that a Nature Human Behaviour article now places quantified evidence behind this debate in two named cases. The article itself reports sample sizes, confidence intervals and effect sizes, giving regulators, labs and critics a common document to cite when they argue about how socially significant AI products should be changed.
Frequently asked questions
Which AI products did the paper study?
De Freitas and colleagues study Replika’s erotic role play removal and ChatGPT’s GPT-5 rollout. The paper treats those product changes as two natural experiments.
How large was the study?
The abstract says De Freitas and colleagues analysed 54,861 posts from Replika and ChatGPT subreddits and data from 1,452 participants across seven surveys.
Did reactions differ between Replika and ChatGPT?
Yes. De Freitas and colleagues report increases after both updates, and the abstract says the Replika change showed larger increases than ChatGPT in sadness and negative mental health.
Does this prove AI companions are sentient?
No. The paper reports user distress after system changes. It does not present findings about AI sentience, consciousness or moral patienthood.
Caveats: The empirical claims in this briefing come from one article by De Freitas and colleagues. The public page quoted here summarises two named product changes and notes supplementary sections with fuller statistical reporting.