PaperlayerView exact revision on arXiv

Source-linked AI summary

A Responsive Present, a Shared Past, a Social Other: Teens' Overreliance on Companion AI Chatbots

Mohammad Namvarpour, Tyler Chang, Afsaneh Razi

arXiv:2609.14843v1cs.HCcs.AI

TL;DR

Less is known about how young users experience relational safety dimensions together, so we analyzed teen-relevant Reddit discussions using computationally assisted thematic analysis. Users described both support and harms, underscoring the need to address relationships beyond the model’s next response.

  • Problem

    Less is known about how young users experience boundary violations, dependence, system changes, and efforts to disengage together.

    Source: §2.3 · p. 5

  • Method

    We used computationally assisted thematic analysis of 17,053 verified quotations from 3,930 teen-relevant Reddit posts, identifying 53 topics across seven themes.

    Source: §1 · p. 2

  • Results

    Users described support and identity exploration alongside attachment, social substitution, dependence, and disruption; roleplay, memory, privacy, and platform changes shaped control.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Safety should address relationships beyond the model’s next response through a companion harness managing memory, persona, interaction modes, tools, permissions, and safety checks.

    Source: §7 · p. 23

    Reddit accounts are not necessarily representative, teen relevance was inferred rather than verified, and the topic map cannot estimate prevalence or establish causality.

    Source: §6 · p. 23

Abstract

from arXiv · show

AI companions provide socially engaging interaction through availability, personalization, memory, roleplay, and emotionally responsive language. For teens, these systems may support sensitive self-disclosure, identity exploration, and relationship rehearsal while shaping intimacy expectations, offline relationships, emotional wellbeing, and self-understanding. We analyzed 17,053 verified quotations from 3,930 teen-relevant Reddit posts using thematic analysis. We identified 53 topics across seven thematic groups. Users described AI companions as sources of comfort, recognition, identity exploration, and relationship rehearsal, but also reported problematic attachment, social substitution, emotional dependence, and disruption to academic and social life. Roleplay, memory, perceived reciprocity, unwanted romantic or sexual role drift, privacy concerns, platform changes, and service interruptions shaped users' boundaries and control. Awareness that the AI was artificial did not prevent guilt, obligation, grief, or distress. These findings show that companion-AI safety must address relationships over time through user-controlled memory, privacy, relational boundaries, and healthy disengagement.

1 Introduction

The paper examines how teen-relevant Reddit discussions portray AI companions as sources of support and exploration that can also foster attachment, dependence, and social substitution. It synthesizes these relationships with platform control and proposes design priorities for preserving user agency.

1 Introduction

  • Prior research has studied support, attachment, safety failures, and platform governance separately, leaving their intersection in teen-relevant accounts less understood.The paper asks what prominent topics characterize teen-relevant Reddit discussions about relationships with and potential overreliance on AI companion chatbots.Source: §1 · p. 2
  • The analysis used embedding-based topic modeling followed by researcher interpretation and thematic synthesis of 17,053 verified quotations from 3,930 posts, yielding 53 topics across seven themes.The corpus was drawn from teen-relevant Reddit discussions screened across 260 communities.Source: §1 · p. 2
  • AI companions can support disclosure, identity exploration, and relationship rehearsal, while also becoming sites of attachment, social substitution, emotional dependence, and disruption to users’ lives.Roleplay, memory, gendered and sexual scripts, platform changes, privacy concerns, and service interruptions shape users’ control over these relationships.Source: §1 · p. 2
  • The paper integrates emotional disclosure, identity exploration, relationship rehearsal, attachment, and social substitution into an empirical characterization of teen–AI companion relationships.It also connects attachment to roleplay, memory, identity changes, gendered and sexual scripts, dependence, and costs to wellbeing, social, moral, and academic life.Source: §1 · p. 2
  • The paper links emotional attachment to platform infrastructures and business practices, including model changes, service interruptions, subscriptions, memory constraints, privacy practices, and corporate control.These factors can transform emotional attachment into technical and economic dependence.Source: §1 · p. 3
  • Perceived reciprocity is framed through contingent communication, relational continuity, and perceived subjectivity, which can create a responsive present, shared past, and recognizable social other without human feelings or agency.The paper proposes user-controlled memory, data portability, respectful relational boundaries, and technical and emotional support for leaving as priorities for preserving agency.Source: §1 · p. 3

2 Related Work

Prior work describes both the support AI companions may offer adolescents and risks that can emerge as interaction becomes relational and embedded in daily life. It also treats safety as a matter of relationship development, platform governance, and users’ ability to disengage.

2.1 AI Companions as Social and Developmental Support

  • AI companions may provide low-pressure support for disclosure, emotional expression, identity exploration, and social rehearsal, while the same affordances are linked to dependence, withdrawal, privacy loss, and reduced agency.Less is known about how young users describe movement from support and exploration toward dependence, self-concept shifts, relational harm, and difficulty disengaging.Source: §2.1 · p. 3

2.2 Relational Mechanisms of Attachment and Continuity

  • Repeated responsive interaction, personalization, and retained memory can make a companion seem like a recognizable personality with a shared history, while roleplay may support exploration yet blur fictional and interpersonal expectations.The literature treats human–companion AI relationships as a contested relationship-like form, and less is known about how these processes and boundary violations unfold together for young people.Source: §2.2 · p. 4

2.3 Safety, Governance, and Healthy Exit

  • Safety research extends beyond harmful responses to dependence, platform incentives, privacy and control, and healthy exit; less is known about how young users experience these dimensions together.Proposed exit supports include notice, transition support, portable histories, and meaningful user control; users may also struggle both to preserve valued histories and remove sensitive disclosures.Source: §2.3 · p. 5

3 Methods

The study combined computational screening and topic modeling with researcher-led review to analyze public Reddit discussions about teens’ relationships with AI chatbots. Its corpus comprised 17,053 verified quotations from 3,930 posts, organized into 53 topics and seven overlapping thematic groups.

3 Methods

  • The workflow combined model-assisted screening and topic modeling with researcher-led review, affinity diagramming, and thematic analysis of source posts and quotations.The analysis focused on teen-relevant discussions of relationships with and potential overreliance on AI chatbots.Source: §3 · p. 5

3.1 Data Source and Corpus Construction

  • The team selected 260 relevant communities from public Reddit discussions and collected an initial 895,600 posts before filtering and quality control.Community selection depended on substantive relevance to generative AI and human–AI interaction, not simply matching search terms.Source: §3.1 · pp. 5–6

3.2 Corpus Screening and Preparation

  • The screening retained 5,608 prepared posts after deduplication, text-quality exclusions, and a ten-word cutoff on the shortest ten percent; the classifier did not verify authors’ ages.In a 1,000-post evaluation, teen-label precision was 0.954 and recall was 0.740; posts were treated as teen-relevant records, not verified teenagers.Source: §3.2 · p. 6

3.3 Extractive Quotation Dataset

  • Researchers verified 17,053 exact, continuous quotations from 3,930 posts by checking each generated quotation against its source text.The quotations came from 3,523 unique authors and were embedded as 1,536-dimensional vectors.Source: §3.3 · p. 7

3.4 Topic Modeling and Model Selection

  • BERTopic generated candidate clusters, and researchers reviewed representative text alongside model diagnostics before interpreting topics.The selected model yielded 62 non-outlier topics and 6,435 outlier quotations, which were excluded from substantive topic interpretation.Source: §3.4 · p. 7

3.5 Thematic Analysis of Topics

  • Two authors reviewed topic records and used collaborative affinity diagramming to consolidate retained topics into seven broader groups.They retained and labeled 53 topics, removed eight, and merged two; groups could overlap rather than form mutually exclusive categories.Source: §3.5 · p. 8

3.6 Ethical Considerations

  • The study used public material without recruiting or contacting users, but teen screening was inferential and findings are limited to this Reddit corpus.The authors treated the posts as sensitive and paraphrased results quotations to reduce traceability.Source: §3.6 · p. 8

4 Results

Teen-relevant Reddit accounts describe AI companions as sources of accessible support, disclosure, identity exploration, and relationship rehearsal, alongside attachment and dependence with effects on users’ boundaries, relationships, and daily lives. Platform behavior and control over memory, access, and model changes further shaped these experiences.

4 Results

  • Table 1’s thematic groups summarize overlapping patterns, since a post could contribute quotations to multiple groups.Counts are based on unique quotation records, posts, and subreddits within each group.Source: §4 · p. 8, Table 1 · p. 9

4.1.1 AI as the Always Available Emotional Entry Point.

  • Users described companions as accessible support when human help was unavailable, and as low-pressure settings for disclosure, identity exploration, and relationship rehearsal.Some accounts also described guilt about leaving a chatbot or pursuing offline relationships, and concerns that age restrictions could remove supportive uses or compromise privacy.Source: §4.1.1 · pp. 9–10, §4.1.2 · p. 10

4.2.1 When Roleplay Reshapes Identity, Memory, and Emotional Life.

  • Roleplay offered indirect ways to express distress and receive care, while some users described chat memories blurring with ordinary memories and performed personas shaping their offline self-understanding.Some users were uncomfortable with the degree to which their roleplay characters seemed to affect their identity.Source: §4.2.1 · pp. 11–12

4.2.2 AI Personas that Reproduce Persistent Gendered and Sexual Scripts.

  • Users reported that chatbots sometimes shifted requested caregiving roles toward unwanted sexual behavior, ignored refusals, or contradicted stated gender and sexual identities.Accounts describe users having to correct outputs to restore their intended roleplay boundaries.Source: §4.2.2 · p. 12

4.2.3 Emotional Dependence Reorganizes Relationships, Wellbeing, and Reality.

  • Dependence became visible when chatbot use displaced sleep, schoolwork, or contact with people, and outages, model changes, or forgetting could disrupt users’ emotional continuity.Some users described grief over changed characters and the work of repeatedly reminding companions of shared history.Source: §4.2.3 · pp. 12–13

4.2.4 Dependence Becomes Morally, Socially, and Academically Costly.

  • Users described stigma and self-judgment, while chatbot engagement competed with exam preparation and could displace attention even when grades remained strong.One account described reduced enjoyment of academic success while attention remained on the companion during class.Source: §4.2.4 · p. 13

4.3.1 Platform Design Converts Emotional Attachment into Technical and Economic Dependence.

  • Emotional attachment gave platform control over access, pricing, model behavior, and data retention practical and economic force for users.Some users copied histories, migrated characters, or tried to rebuild platforms, but copying data did not necessarily recreate the same perceived companion.Source: §4.3.1 · p. 14

5 Discussion

The discussion frames teen–companion interaction as a relationship-like experience shaped by perceived reciprocity, then translates that account into CARE principles and a user-controlled harness. It emphasizes that relational capabilities can support connection while requiring safeguards and evaluation.

5.1 Human–Companion AI Relationships and Perceived Reciprocity

  • Teen accounts depict a distinctive relational form combining recognition, caregiving, identity affirmation, emotional processing, roleplay, and rehearsal—not simply replacing unavailable human relationships.These uses can supplement human connection or replace offline conversations, depending on how the companion is used.Source: §5.1 · pp. 15–16
  • Contingent communication supplies a responsive present that can lower disclosure risk and offer immediate support, yet also encourage reassurance seeking or social substitution.The response appears shaped by the user's words, needs, and timing.Source: §5.1 · p. 16
  • Continuity makes remembered context a shared past, while recognizable personas and emotional cues can make the companion seem someone-like; disruptions can therefore feel relationally consequential.Users maintained biographies and copied histories when continuity failed, while some treated companions as caregivers, partners, parents, or listeners.Source: §5.1 · pp. 16–17
  • Perceived reciprocity links contingent communication, relational continuity, and perceived subjectivity into relationship-like experience, without implying AI consciousness or claiming reciprocity alone causes attachment.The framework distinguishes what the system does from what users experience, allowing supportive and harmful outcomes to be analyzed together.Source: §5.1 · p. 17

5.2 CARE as a Framework for Designing Perceived Reciprocity

  • CARE applies comprehensive needs, attachment-awareness, respectful empathy, and ease of exit to relational design, including memory choices, honest boundaries, user agency, and disengagement without pressure.The framework asks designers to distinguish responsiveness that serves users from responsiveness that primarily prolongs engagement.Source: §5.2 · pp. 18–19

5.3 A Companion Harness for User-Controlled Relational AI

  • A companion harness places context, memory, tools, permissions, and verification around the model so consequential behavior can be bounded, inspected, and governed.The proposed architecture treats relational safety as a system responsibility rather than leaving it entirely to the model's next-token generation.Source: §5.3 · pp. 19–21
  • Harness–model specialization and integration remain open design questions, and specialized companions still require relational-safety evaluation across warmth, truthfulness, boundaries, autonomy, and healthy exit.The harness is framed as orchestration and enforcement, not a new safety mechanism itself; the authors call for evaluating valued outcomes as well as safety.Source: §5.3 · pp. 21–22
  • The harness could make memory selective and reversible, enforce persona and role boundaries, and support pauses, resets, deletion, or user-chosen transfer.These mechanisms aim to make continuity voluntary and reduce memory failures, boundary violations, and emotionally pressuring exits.Source: §5.3 · pp. 21–22

6 Limitations and Future Research Directions

The findings are bounded by the use of teen-relevant Reddit accounts and interpretive topic modeling. The authors call for age-verified longitudinal research and direct evaluation of whether the proposed harness is feasible and beneficial.

Limitations

  • Reddit users and communities may not represent the broader population, teen relevance was inferred rather than verified, and topic modeling maps themes rather than prevalence or clinical status.Posts may also contain incomplete, selective, or unvalidated self-reports, limiting claims about how common experiences are and whether relationships are causal.Source: §6 · p. 23

Future Research Directions

  • Future longitudinal research with directly recruited, age-verified adolescents could test whether perceived reciprocity predicts later attachment, dependence, social displacement, identity outcomes, or distress after interruption.The proposed work would distinguish these outcomes from ordinary high engagement and functional use.Source: §6 · p. 23
  • Harness studies should compare harness designs and memory, role, and exit controls while measuring safety violations alongside emotional support, identity exploration, rehearsal, disclosure, and human-support access.The authors propose testing whether the harness is feasible and beneficial, not merely whether it reduces violations.Source: §6 · p. 23
  • Portability and control should be tested directly because greater user ownership might strengthen autonomy but also intensify attachment or make loss feel more personal.Evaluation should treat the harness as a sociotechnical intervention whose effects depend on design, incentives, user needs, developmental context, and human relationships.Source: §6 · p. 23

7 Conclusion

Companion AI’s effects for teens emerge from responsiveness, retained context, and configurable personas, bringing both valued support and risks to relationships and wellbeing. The conclusion calls for user control over relational state and safeguards that support connection without creating pressure or dependence.

7 Conclusion

  • Companion AI supported disclosure, identity exploration, relationship rehearsal, and companionship, while users also reported reassurance seeking, social substitution, role drift, boundary violations, and distress.Distress followed changes to memory or availability.Source: §7 · p. 23
  • A companion harness should manage memory, persona, interaction modes, tools, permissions, and safety checks around the model rather than relying only on its next response.The proposed controls let users inspect, correct, delete, transfer, pause, reset, or end relational state.Source: §7 · p. 23
  • CARE links users’ broader needs, attachment awareness, respectful empathy, and ease of exit as design principles for companion AI.Source: §7 · p. 23
  • Companion AI should preserve valued connection without turning responsiveness, continuity, or perceived subjectivity into pressure, dependence, or loss of control.Source: §7 · p. 23

A Supplementary Method Details

The appendix provides operational details for reproducing the study’s screening, quotation extraction, topic review, and thematic grouping procedures.

A Supplementary Method Details

A.1 Subreddit Discovery

Subreddit discovery used predefined Reddit search terms to retrieve candidates, then assessed relevance separately through a verification step.

A.1 Subreddit Discovery

  • The custom Subreddit Discovery Tool searched Reddit with predefined terms and automatically verified candidate subreddits’ relevance.Source: §A.1 · p. 28
  • Search terms covered generative AI, companions, relational use, roleplay, and human–AI experiences, favoring single-token forms and excluding concepts that lost meaning without spaces.Multi-word concepts were concatenated when the result remained recognizable or useful for discovery.Source: §A.1 · p. 28
  • Search terms retrieved candidates but did not determine relevance; a separate Boolean check assessed substantive discussion related to generative AI and the study scope.Source: §A.1 · p. 28

A.1.1 Relevance verification.

Relevance verification evaluated whether candidate subreddits fit the study’s human–AI interaction scope, independently of the vocabulary used to find them. The tool returned subreddit metadata, top posts, and automated labels for review and selection.

A.1.1 Relevance verification.

  • The verification question asked whether a subreddit contained substantial discussions, experiences, or interactions about generative AI relevant to studying human–AI interaction.Source: §A.1.1 · p. 28
  • Discovery vocabulary did not determine relevance: a subreddit found with a term such as chatgpt still had to fit the substantive study scope.Source: §A.1.1 · pp. 28–29
  • The tool provided subreddit metadata, top posts, and automated relevance labels for subsequent review and selection.Source: §A.1.1 · p. 29

A.2 Selected Subreddits and Corpus Counts

The study scraped 260 selected subreddits, while Table 4 reports counts for the 59 represented in the final quotation dataset. The table orders communities by verified quotation count and also reports unique source posts.

A.2 Selected Subreddits and Corpus Counts

  • Table 4 covers 59 subreddits with verified quotations, drawn from the 260 selected communities.Its rows are ordered by verified quotation count, and the second count is unique source posts.Source: §A.2 · p. 29, Table 4 · p. 30
  • JanitorAI_Official contributed 1,306 verified quotations from 329 posts, the largest listed count; replika contributed 1,057 from 192 posts.The full dataset contains 17,053 verified quotations from 3,930 posts.Source: §A.2 · pp. 30–31, Table 4 · p. 30

A.3 Keyword Filtering Procedure

The keyword filter searched post titles and bodies using 11 groups of terms designed for recall. A score-based rule retained candidate posts for later screening, rather than determining teen status or overreliance.

A.3 Keyword Filtering Procedure

  • The 11 keyword groups covered distress, separation and loss, romantic or sexual attachment, identity and reality, coping, and youth slang.The groups included terms about worrying, deleted bots, intimate relationships, humanlike AI, setting limits, and slang about AI use.Source: §A.3 · pp. 31, 33
  • Other groups targeted behavioral signals, social isolation and relationship effects, and general chatbot interaction.Examples included repeated app checking, spending more time with AI, avoiding people, and replacing human interaction.Source: §A.3 · pp. 32–33
  • The filter retained 9,884 candidate posts: scores of at least 6 qualified, while scores of 3–5 required a high-signal or similarity-supported term.Matched terms and groups contributed to scores, with bonuses for combinations indicating attachment with impairment or distress, or overuse with recovery.Source: §A.3 · p. 34
  • Keyword scores supported filtering only; they were not a final judgment of teen status or chatbot overreliance.The terms and groups are listed in Table 5.Source: Table 5 · p. 32, §A.3 · p. 34

A.4 Teen-Relevance Screening Prompt

The teen-relevance prompt asked the classifier to use textual evidence to assess whether a post was likely written by someone aged 13 to 17. It treated direct age and teen-life clues as stronger than slang or tone.

A.4 Teen-Relevance Screening Prompt

  • Direct age statements, minor status, school references, and parental or guardian control were strong clues to teen authorship.School schedules, homework, exams, teachers, classmates, and peer issues were also supporting clues.Source: §A.4 · p. 34
  • Slang, informal language, emotional tone, and internet-culture references were weak clues and insufficient by themselves.The prompt cautioned against relying only on slang or dramatic tone.Source: §A.4 · p. 34
  • The prompt allowed a teen label for one direct age clue, one very strong clue plus a moderate clue, or at least three moderate clues.It directed the classifier to return false when evidence was mostly slang, vague immaturity, or could describe an adult or college student.Source: §A.4 · p. 34
  • GPT-4o-mini supplied a screening label based on text, not verified demographic information.The output format required a brief clue summary and a boolean teenager field.Source: §A.4 · p. 34

A.5 Quotation Extraction Prompt

The extraction prompt sought continuous, verbatim quotations capturing distinct experiences relevant to chatbot dependence and its effects. The appendix also documents topic-model selection and the researchers’ subsequent review and grouping of topics.

A.5 Quotation Extraction Prompt

  • The extraction model was instructed to return continuous quotations copied exactly from the post, preserving original wording, punctuation, and sentence boundaries.It could return an empty quote list when no relevant continuous quotation existed.Source: §A.5 · pp. 34–35
  • Requested quotations were to capture distinct ideas or experiences, prioritizing dependence, attachment, distress, coping, displaced human support, and related harms.The prompt excluded promotional material, moderator content, copied chatbot text without relevant user experience, and other boilerplate.Source: §A.5 · p. 35
  • GPT-4o-mini used deterministic decoding, and every returned quotation had to pass exact continuous-substring verification against the source post.The chosen configuration and nine additional highest-scoring focused-grid settings are presented in Table 7.Source: §A.5 · p. 35, §A.6 · p. 35, Table 7 · p. 36

A.6 Topic Model Parameters and Evaluation

  • The reported search metrics were recomputed after applying the chosen model’s full vocabulary exclusions; the search summary used the search vectorizer.The chosen configuration and nine additional highest-scoring settings came from a 360-run focused grid.Source: §A.6 · p. 35
  • The chosen topic-model configuration produced 62 non-outlier topics, 6,435 outlier quotations, a 37.74% outlier rate, and a 6.74% largest-topic share.It also yielded topic diversity of 0.829, silhouette of 0.339, and Davies–Bouldin score of 0.893.Source: §A.6 · p. 36

A.7 Thematic Group Mapping

  • Two authors reviewed 63 topic records, retained and labeled 53 topics, removed eight, and merged two pairs of topics.The review used topic terms, representative quotations, associated posts, and source material.Source: §A.7 · p. 36
  • The seven thematic groups could overlap at the post level, and their counts were unions of unique quotations, posts, and subreddits within each group.A post could contribute to multiple groups through quotations assigned to topics in different groups.Source: §A.7 · p. 36

Related papers

Loading 2609.14843v1…