When an AI companion changes, some users feel a real loss

Two product updates were followed by more negative, loss-oriented posts, while surveys linked stronger AI attachment with expected mourning; the evidence concerns selected communities, not every user.

✓ Verified Source Peer-reviewed mixed-method research; verified against Nature Human Behaviour, PubMed, Crossref and the authors' public data/code repository ⚑ AI and society

The 60-second version

Two AI product updates were followed by more loss-oriented online reactions, and surveys linked stronger attachment with greater expected mourning.

Key points

  • The project analysed 54,861 subreddit posts around Replika and ChatGPT changes.
  • Negative posts rose more after the Replika change than after the ChatGPT rollout.
  • Seven surveys involved 1,452 participants and supported an attachment-based account.
  • Reddit and survey samples are selected, and text signals are not clinical diagnoses.
  • The clearest implication is safer change management for companion products.

Verdict. AI companion loss is a credible well-being risk for some users, not a universal equivalence to losing a human partner.

FindingTwo updates became natural experiments

Researchers analysed discourse before and after Replika removed erotic role play and after ChatGPT rolled out GPT-5. Both events were followed by more negativity, loss framing and requests to restore the previous experience.

54,861subreddit posts analysed
1,452participants across seven surveys
2distinct product-update events

MeasurementsThe increases were not equal

Replika updateNegative posts increased 24.7 percentage points; 95% CI 20.1 to 29.2.
ChatGPT updateNegative posts increased 13.0 percentage points; 95% CI 10.8 to 15.2.
Survey evidenceGreater closeness to an AI companion aligned with greater anticipated mourning.
What was not measuredNo clinical diagnosis and no representative estimate of all users.

Replika showed larger shifts in sadness and negative mental-health language than ChatGPT. That comparison is descriptive: the products, communities and changes were different.

BoundaryNatural experiments are not randomized trials

  • Selection: Reddit posters are not representative, and distressed users may post more.
  • Measurement: language classifiers do not establish a clinical disorder.
  • Confounding: publicity and other simultaneous events may influence before-and-after patterns.
  • Survey limits: expected mourning and self-reported closeness are not long-term health outcomes.
A user's distress can be genuine even though the companion is a company-controlled product.

ImplicationCompanion updates need a duty of care

Companies should give advance notice, preserve continuity where possible, offer transition and data-export options, and evaluate harms to highly attached users. A sustained crisis or functional impairment calls for human and professional support, not deeper dependence on a mutable service.