A Responsive Present, a Shared Past, a Social Other: Teens' Overreliance on Companion AI Chatbots
Mohammad Namvarpour, Tyler Chang, Afsaneh Razi
cs.HC · cs.AI
Quick verdict
Thematic analysis of 17,053 Reddit quotes maps teen AI-companion overreliance; pipeline transparent, but teen status model-inferred and sampling harm-biased.
Summary
The paper presents a computationally assisted thematic analysis of teen-relevant Reddit discussions about AI companion chatbots, asking which topics characterize relationships with and potential overreliance on these systems. From 260 subreddits the authors screened 895,600 posts; keyword filtering, GPT-4o-mini teen-relevance classification, and verbatim substring-verified quotation extraction yielded 17,053 quotations from 3,930 posts. Embeddings (text-embedding-3-small) fed a BERTopic model selected through a 392-run grid search, producing 62 non-outlier topics (37.74% outlier rate); two authors jointly consolidated these into 53 topics across seven themes: always-available emotional entry point, adolescent vulnerability, roleplay reshaping identity and memory, gendered and sexual scripts, dependence reorganizing relationships and reality, moral/social/academic costs, and platform design converting attachment into technical and economic dependence. Conceptually, the paper defines perceived reciprocity via three affordances (contingent communication, relational continuity, perceived subjectivity), maps design guidance onto the CARE framework, and sketches a companion-harness architecture for memory control, boundaries, and healthy exit. Key limitations, partly acknowledged: teen status is inferred by GPT-4o-mini without human validation; the keyword screen loads on dependence and harm vocabulary, risking circularity; the corpus skews heavily toward Character.AI and NSFW communities; coding was joint, without inter-rater reliability; and reported quotations are author-paraphrased. Findings are descriptive and not prevalence estimates.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Teen-relevant Reddit accounts portray companion AI as an always-available emotional entry point for disclosure, identity exploration, and relationship rehearsal.
- The same relationships produce attachment, social substitution, emotional dependence, and moral, social, and academic costs, and awareness that the AI is artificial does not prevent guilt, obligation, grief, or distress.
- Perceived reciprocity — constituted by contingent communication, relational continuity, and perceived subjectivity — is the central relational process that makes companion AI relationship-like.
- Platform decisions (model changes, subscriptions, memory and data policies, filters) convert emotional attachment into technical and economic dependence.
- Companion-AI safety must be relational and architectural: user-controlled memory, portability, enforceable boundaries, and supported exit via a companion harness governed by CARE principles.
Strengths
Unusually transparent, auditable pipeline
The paper publishes the full subreddit list and per-community counts (Table 4), all 11 keyword groups (Table 5), both operational LLM prompts (A.4-A.5), the exact BERTopic configuration (Table 6), and the top-10 grid-search rows with metrics (Table 7) — rare documentation depth for computational qualitative HCI work.
Fabrication-resistant quotation extraction
Every LLM-extracted quotation (GPT-4o-mini, deterministic decoding) was accepted only after exact continuous-substring verification against the source post (§3.3, A.5), guarding the corpus against model hallucination; 17,053 quotes passed from 3,930 posts.
Model selection separated from interpretation
Conceptual and design integration beyond description
Weaknesses
The 'teens' construct is an unvalidated classifier label
Selection circularity and corpus skew
No inter-rater reliability in thematic consolidation
Paraphrased evidence undermines the 'verified quotations' chain
Large excluded outlier pool is uncharacterized
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.14843, cs.HC) asks one descriptive research question: what prominent topics characterize teen-relevant Reddit discussions about relationships with, and potential overreliance on, AI companion chatbots? It delivers (1) an empirical map of 53 topics in seven thematic groups (Table 1), (2) a conceptual model of *perceived reciprocity* built from three affordances — contingent communication (a responsive present), relational continuity (a shared past), and perceived subjectivity (a social other) (Fig. 2), (3) a design translation into the CARE framework (Table 2), and (4) a proposed companion-harness runtime architecture (Fig. 3). Claim strength is calibrated: the authors repeatedly disclaim prevalence, causality, and population generalization (§3.6, §6), and the title's 'overreliance' is anchored in user-reported dependence narratives rather than clinical criteria. The risk lies less in stated claims than in what the design can support.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does this study actually verify that the Reddit users are teenagers?
No. Teen status was inferred by the GPT-4o-mini model from post content (age statements, school references, parental control cues) using cautious rules. The authors state this is an inference, not verified demographic information, and limit their findings to this Reddit corpus rather than to teenagers as a population.
How big is the dataset and where does it come from?
The authors scraped 895,600 public posts from 260 subreddits via the Arctic Shift archive. After keyword filtering, model-based teen-relevance screening, and verbatim quotation extraction with substring verification, the analyzed corpus was 17,053 quotations from 3,930 posts. Only 59 subreddits contributed quotations, with Character.AI communities supplying 9,172 (about 54%).
What is the paper's main conceptual contribution?
A model of perceived reciprocity built from three affordances: contingent communication (a responsive present), relational continuity (a shared past), and perceived subjectivity (a recognizable social other). It also maps design guidance onto the CARE framework and proposes a companion-harness runtime architecture for user-controlled memory, relational boundaries, and healthy exit.
Does the paper claim AI companions harm teens?
No causal or prevalence claim is made. It reports how users describe both benefits (emotional support, identity exploration, relationship rehearsal) and harms (attachment, social substitution, unwanted sexual role drift, distress at service interruptions), and argues that safety design should address relationships over time rather than only individual responses.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text is truncated: portions of §3.3, Results §4.2.2-4.3, Discussion §5.2-5.3, and part of Appendix A.3 are omitted; assessment of those parts relies on Table 1, surviving fragments, and the appendices.
- Figures 1-3 are available only as embedded captions; the visual detail of the companion-harness architecture (Fig. 3) could not be evaluated.
- Results quotations are author-paraphrased by design (§3.6), so the raw evidentiary voice of the corpus cannot be audited from the text.
- Some reference entries (roughly refs 37-60) are cut off in extraction; citation completeness was not fully checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.