Source-linked AI summary
Disrupted Companionship: A Risk Assessment Framework and Cross-Platform Quantitative Analysis of Psychosocial Responses to AI Companion Disruptions
Chau Do, Yunhao Yuan, Koustuv Saha, Renwen Zhang, Talayeh Aledavood
TL;DR
AI companion relationships can be disrupted by platform-initiated changes, yet evidence has been limited on how risks vary across disruption events. The study compiles and codes 30 events, develops a four-dimension risk framework, and analyzes longitudinal Reddit data with a hierarchical Bayesian interrupted time-series model. Across 19 events with sufficient data, disruption onset was associated with immediate increases in anxiety, stress, suicidal expression, and grief activation, with relational discontinuity and transition-support deficit linked to more adverse responses across several outcomes.
Problem
Evidence is limited on how psychosocial consequences generalize across AI companion disruptions and which event characteristics make some changes more harmful than others.
Method
The study compiles 30 disruption events, codes their types, reasons, and four risk dimensions, and models longitudinal Reddit expressions around disruption onset.
Results
Across 19 analyzed events, disruption onset was associated with immediate increases in anxiety, stress, suicidal expression, and grief activation; relational discontinuity was the most consistent moderator.
Takeaways & Limitations
The findings provide a cross-platform basis for characterizing AI companion disruptions and assessing potential risks before platform changes are implemented.
Takeaways & Limitations
The corpus was exploratory rather than exhaustive, may overrepresent visible controversy, and included only 19 events with sufficient quantitative data.
Abstract
from arXiv · showhide
AI companions can provide meaningful relationships, yet these relationships remain vulnerable to platform-initiated changes. We study AI companion disruptions: platform changes that alter or terminate users' ongoing companionship with an AI. We compile 30 disruption events across major platforms, develop a taxonomy of six disruption types, identify three broad reasons for disruption, and propose a risk-assessment framework comprising four dimensions: relational discontinuity, population vulnerability, communication deficit, and transition-support deficit. Using longitudinal Reddit data, we estimate community-level psychosocial responses with a hierarchical Bayesian interrupted time-series model incorporating predictive controls. Across events, disruption onset was associated with immediate increases in anxiety, stress, suicidal expression, and grief activation, with relational discontinuity and transition-support deficit being associated with more adverse immediate responses across several outcomes. Our findings provide a cross-platform characterization of AI companion disruptions, quantitative evidence of their psychosocial impacts, and a prospective framework for assessing their potential risks before implementation.
1 INTRODUCTION
AI companion disruptions are platform-initiated changes that can abruptly alter meaningful ongoing relationships, but their broader psychosocial consequences and risk factors remain insufficiently characterized. This study addresses that gap through a cross-platform event corpus, a four-dimension risk framework, and quantitative analysis of community-level responses.
- Motivation: The study defines AI companion disruptions as platform-initiated changes that disrupt users’ ongoing companionship with an AI.Such changes may include service shutdowns, feature removal, or access restrictions that substantially alter or end emotionally meaningful relationships.
- Motivation: Existing evidence does not establish how psychosocial consequences generalize across different platforms, disruption forms, affected populations, or implementation conditions.The study highlights variation in what changes, why it changes, relationship impact, communication, and transition support.
- Study aims: The researchers compiled 30 disruption events and developed a taxonomy of disruption types, categories of disruption reasons, and four risk dimensions.The risk dimensions are relational discontinuity, population vulnerability, communication deficit, and transition-support deficit.
- Study aims: Using longitudinal Reddit discussions and a hierarchical Bayesian interrupted time-series model with predictive controls, the study estimated community-level psychosocial changes around disruptions.The analysis addresses immediate and short-term expression changes while examining variation across events and risk factors.
- Findings: Disruption onset was associated with immediate increases in anxiety, stress, suicidal expression, and grief activation across events.Relational discontinuity was the most consistent moderator, while disruptions removing relationship-supporting affordances produced larger immediate changes across several psychosocial outcomes.
2 RELATED WORK
Prior research shows that AI companion relationships can involve attachment, support, intimacy, and dependence, while disruptions can produce loss and distress through multiple mechanisms. This study extends case-focused work by comparing disruptions across platforms and examining how event characteristics relate to psychosocial responses.
- Changes and Discontinuation of AI Companions: Prior case studies document emotional loss, mourning, identity discontinuity, and deteriorated mental health after companion shutdowns or relational feature removals.These responses can occur even when the underlying companion remains accessible.
- Changes and Discontinuation of AI Companions: Disruption execution may shape user responses: deliberate offboarding, communication, emotional closure, continuity mechanisms, and advance notice have been identified as relevant factors.Experimental evidence reported that forewarning users before termination reduced feelings of loss.
- Research Gap: Existing literature indicates that disruption can occur through multiple mechanisms and that consequences depend on characteristics of the disruption and its execution.However, prior work has primarily examined individual cases or particular termination forms rather than systematic cross-event comparisons.
- AI Companionship and Well-being: AI companion relationships can provide social interaction, emotional support, and possible reductions in loneliness, but reported well-being effects are mixed and heterogeneous.Studies also describe emotional validation and social rehearsal alongside potentially harmful psychosocial effects.
- Social Media Methods: The study uses longitudinal Reddit data and computational psychosocial measures with a hierarchical Bayesian interrupted time-series model to compare responses surrounding disruptions.Reddit provides pseudonymous, community-based, longitudinal discussions of personal experiences and psychosocial states.
3 STUDY DESIGN AND DATA
The study combines a cross-platform disruption corpus with longitudinal Reddit data from platform-specific and control communities. It codes disruption characteristics and risk dimensions, then models community-level psychosocial expression around disruption onset using a hierarchical Bayesian interrupted time-series design.
- Study Design: The study proceeds through event collection, qualitative coding, risk-framework development, and quantitative analysis of psychosocial responses.The data combine disruption nature, rationale, timeline, and longitudinal Reddit records from focal and control communities.
- Disruption Events: An eligible disruption is platform-initiated, affects users with an existing AI relationship, and is documented through official platform or reliable external sources.The criteria focus the corpus on changes to ongoing companionship rather than voluntary disengagement or effects on prospective users.
- Disruption Events: The final corpus contains 30 disruption events across 19 AI companion platforms, including dedicated companions, therapy or mental-health chatbots, a companion robot, and general-purpose assistants.For each event, the study records the change, stated reason, communication, transition resources, and timeline.
- Limitations: The disruption search was exploratory rather than exhaustive, and the corpus may overrepresent events that generated visible controversy or documented negative reactions.Only 19 events had sufficient data for the quantitative analysis, limiting precision for moderation estimates across unevenly distributed risk dimensions.
- Reddit Data Collection: Reddit submissions were collected from platform-dedicated communities and control communities over a window spanning 90 days before to 14 days after disruption onset.The pre-disruption period supports baseline estimation, while predictive controls help distinguish disruption-specific changes from contemporaneous variation and broader temporal trends.
- Privacy and Ethics: The analysis aggregates daily psychosocial expression proportions at the community level and does not make predictions or classifications about individual users or posts.The study uses publicly available archived Reddit material without direct interaction with users, while acknowledging privacy and ethical concerns about secondary analysis of public discussions.
4 CHARACTERIZING AI COMPANION DISRUPTIONS AND THEIR POTENTIAL RISKS
The study characterizes 30 AI companion disruptions through a six-type taxonomy, three broad reasons, and a four-dimension prospective risk framework. It shows that disruptions vary in what they alter, why they occur, and how potential risks combine across events.
- What Makes a Disruption Potentially Risky? Four Risk Dimensions: The framework assesses relational discontinuity, population vulnerability, communication deficit, and transition-support deficit using evidence available at or before implementation.This prospective criterion allows platforms to evaluate potential risks before users’ subsequent reactions are observed.
- What Changes? A Taxonomy of Disruption Types: Six disruption types capture changes to the companion entity, model, functionality, interaction availability or persistence, interaction scope, and access modality.Withdrawal/termination removes the user-facing entity; other types preserve it while altering particular aspects of interaction.
- What Changes? A Taxonomy of Disruption Types: Withdrawal/termination was the most common type, with 11 events, while interaction availability/persistence restriction accounted for seven and interaction-scope restriction for five.Feature removal was identified once, and access-modality change occurred in two events.
- Risk Profiles Across Disruption Types: Even a limited policy or model change may disrupt relationship-sustaining practices, whereas a complete shutdown may be mitigated by advance communication and transition support.The relevant assessment therefore concerns what the change interrupts, whom it affects, and what resources users have for responding.
- Why Do Disruptions Occur?: Safety or regulatory compliance was the most common disruption reason, covering 15 events, followed by commercial or organizational circumstances covering 11 events.Safety-related pressures included internal judgments, legal or regulatory pressure, third-party governance, intellectual-property claims, app-store policies, and upstream provider terms.
5 PSYCHOSOCIAL RESPONSES TO AI COMPANION DISRUPTIONS
Across 19 AI companion disruption events, the study used longitudinal Reddit data and hierarchical Bayesian interrupted time-series models to estimate community-level psychosocial responses. Disruptions were associated with acute increases in several distress and grief measures, with relational discontinuity and transition-support deficits linked to more adverse responses.
- 5.3.1 Immediate Changes and Post-Disruption Trajectories.: Stress showed a 0.4% daily decrease in odds after its immediate increase, suggesting attenuation following the acute disruption response.The estimated post-disruption slope change was ψ0 = −0.0038 relative to the pre-disruption trajectory.
- 5.3.2 How Event-Level Risk Shapes Psychosocial Responses.: High relational discontinuity was associated with larger immediate increases in suicidal expression, loneliness, stress, depression, and grief activation, while grief valence was also higher.The largest reported odds differences were 45.6% for suicidal expression, 35.2% for loneliness, 12.3% for stress, and 9.7% for depression.
- 5.3.2 How Event-Level Risk Shapes Psychosocial Responses.: High relational discontinuity was associated with subsequent daily declines in anxiety, depression, and stress relative to low-discontinuity events.The additional declines were approximately 2.5%, 1.4%, and 0.9% in odds per day, respectively, consistent with attenuation after the stronger onset response.
- 5.3.2 How Event-Level Risk Shapes Psychosocial Responses.: Transition-support deficits were associated with larger immediate increases in loneliness and depression and lower grief valence, indicating more negative affect.The differences were 49.2% higher loneliness odds, 7.7% higher depression odds, and a 0.66-standard-deviation decrease in grief valence.
- 5.3.2 How Event-Level Risk Shapes Psychosocial Responses.: Transition-support and communication deficits were also associated with later trajectory differences, including increasing grief activation and depression over time.Transition-support deficit corresponded to a 0.07-standard-deviation daily increase in grief activation, while communication deficit corresponded to a 1.4% daily increase in depression odds.
6 DISCUSSION
The discussion frames AI companion disruption as multidimensional, with relational discontinuity and transition-support deficits most consistently associated with adverse psychosocial responses. It also translates these findings into pre-deployment risk assessment and safer transition practices while noting important dataset and coding limitations.
- 6 DISCUSSION: Relational discontinuity was associated with larger immediate increases in suicidal expression, loneliness, stress, depression, and grief activation, alongside higher grief valence.The association appeared across technically heterogeneous events, indicating that product-level scope alone may not anticipate psychosocial impact.
- 6.2 Relational Precarity and Platform Control: Disruption type, disruption reason, and four risk dimensions capture different, cross-cutting aspects of how platform changes affect ongoing AI relationships.The same disruption type can arise for different reasons and with different levels of relational discontinuity, population vulnerability, communication deficit, and transition-support deficit.
- 6 DISCUSSION: Relationally discontinuous events showed larger immediate responses followed by more negative subsequent slope changes for depression and stress.The pattern is consistent with acute responses followed by some adjustment over time, potentially involving adaptation, disengagement, migration, or relationship reconstruction.
- 6.3 Designing for Safe Disruptions: Relational discontinuity and transition-support deficit showed the clearest associations with variation in immediate psychosocial responses.Relationship-supporting affordances and practical resources such as export, migration, or legacy access are central considerations when changes cannot be avoided.
- 6.3 Designing for Safe Disruptions: Population vulnerability remains important despite no retained quantitative moderation effects for that factor.Prior work documents emotional consequences of companion loss among vulnerable populations, including children and neurodivergent users.
- 6.3 Designing for Safe Disruptions: Platforms can use the four observable risk dimensions in change-management processes before deployment, while preserving relationship-supporting features where possible.Suggested safeguards include temporary legacy access, model choice, conversation export, migration mechanisms, and advance communication treated separately from transition support.
- 6.4 Limitations and Future Directions: The quantitative moderation estimates are constrained by a corpus that may overrepresent visible controversy and by only 19 events with sufficient data.Binary coding also collapses variation in the degree of communication, transition-support, and relational discontinuity characteristics.
7 CONCLUSION
This work characterizes AI companion disruptions across platforms and proposes dimensions for assessing their potential risks. Using Reddit discussions and a hierarchical Bayesian interrupted time-series model, it finds immediate increases in several psychosocial expressions and variation associated especially with relational discontinuity.
- 7 CONCLUSION: The study compiled 30 disruption events, developed six disruption types, identified three broad reasons, and characterized events along four risk dimensions.The dimensions are relational discontinuity, population vulnerability, communication deficit, and transition-support deficit.
- 7 CONCLUSION: Using Reddit discussions surrounding 19 disruption events, the model found immediate increases in anxiety, stress, suicidal expression, and grief activation.The analysis used a hierarchical Bayesian interrupted time-series model with predictive control series.