Messages in a Digital Bottle: A Youth-Coauthored Perspective on LLM Chatbots and Adolescent Loneliness Jinyao Liu Wycombe Abbey High Wycombe UK dora.jinyaoliu@gmail.com Di Fu School of Psychology, Surrey Institute for People-Centred Artificial Intelligence, University of Surrey Guildford UK d.fu@surrey.ac.uk 2018

Adolescent loneliness is a growing concern in digitally mediated social environments. This work-in-progress presents a youth-authored critical synthesis on chatbots powered by Large Language Model (LLM) and adolescent loneliness. The first author is a 16-year-old Chinese student who recently migrated to the UK. She wrote the first draft of this paper from her lived experience, supervised by the second author. Rather than treating the youth perspective as one data point among many, we foreground it as the primary interpretive lens, grounded in interdisciplinary literature from social computing, developmental psychology, and Human-Computer Interaction (HCI). We examine how chatbots shape experiences of loneliness differently across adolescent subgroups, including those with anxiety or depression, neurodivergent youth, and immigrant adolescents, and identify both conditions under which they may temporarily reduce isolation and breakdowns that risk deepening it. We derive three population-sensitive design implications. The next phase of this work will expand the youth authorship model to a panel of adolescents across these subgroups, empirically validating the framework presented here.

 Adolescent Loneliness, Youth-Authored Research, AI Chatbots, Large Language Models, Immigrant Adolescents, Social Computing, Child-Computer Interaction acmlicensed 2018 XXXXXXX.XXXXXXX Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY 978-1-4503-XXXX-X/2018/06
1 section 1 1 §1 <tag close=". ">1</tag>Introduction 1Introduction

Adolescent loneliness represents a significant and growing public health concern. UK data show that 1 in 6 adolescents aged 16 and over exhibit symptoms of a mental health problem, and by 2023, 20% of children aged 8-16 met criteria for a probable mental disorder , up 8% from 2017 ( , ). Loneliness, defined as the subjective discrepancy between desired and actual social connections, is consistently associated with depression, anxiety, and longer-term mental health difficulties ( , ).

As of 2023, 95% of adolescents aged 13-17 have access to smartphones ( , ), and LLM-powered chatbots have increasingly entered their lives as companions, listeners, and sources of informal emotional support ( , ). These systems are always available, low-cost, and perceived as non-judgmental, properties that lower the threshold for emotional disclosure among adolescents who hesitate to approach parents, teachers, or peers. However, the same affordances raise concerns about over-reliance, social withdrawal, and the substitution of algorithmic interaction for human connection.

Existing research tends to treat adolescents as a homogeneous group and focuses on either benefits or risks in isolation, rarely attending to the heterogeneity of adolescent vulnerability ( , ). This paper addresses that gap through a youth-authored reflective synthesis. The first author, a 16-year-old who migrated from China to the UK less than a year ago, wrote the first draft from lived experience, supervised by the second author. We treat this not as a limitation, but as the paper’s methodological contribution: a situated perspective that surfaces needs and risks that adult-led research has largely overlooked.
2 section 2 2 §2 <tag close=". ">2</tag>Positionality and Method 2Positionality and Method

This paper is a youth-authored reflective synthesis, combining three sources of insight: (1) the lived experience of the teenage co-author, who migrated from China to the UK at age 16 and used chatbots extensively during early adaptation; (2) iterative discussion between both authors throughout the writing process, in which the teenage co-author’s observations were used to interrogate and reframe existing literature; and (3) engagement with interdisciplinary scholarship from social computing, developmental psychology, and HCI. The contribution is one of situated insight: we do not claim generalisable findings, but rather a conceptual framework, grounded in a specific, underrepresented perspective, intended to guide future empirical research and participatory design with broader adolescent populations.

The collaborative process involved regular meetings over four months, during which the teenage co-author drafted sections, identified relevant literature, and articulated observations from her own chatbot use. The supervisor challenged and questioned these observations against existing scholarship, using the co-author’s account as a lens through which to interrogate and triangulate literature claims. Where the literature made general claims about adolescent chatbot use, her specific experience was used to probe the boundaries of those claims, asking which conditions they applied to, which populations they assumed, and what they left out. This process is not equivalent to a formal qualitative study, but it is also not mere consultation: the analytical distinctions in this paper, particularly the subgroup framework and the identification of the cultural assumption problem in Section 4.4, emerged from this sustained triangulation between situated experience and the published record.

Jinyao Liu is a 16-year-old Chinese student at Wycombe Abbey who migrated to the UK less than a year ago. She wrote the first draft of this paper, including all conceptual framing, the selection and interpretation of literature, and both figures. Di Fu supervised the work, provided feedback, and revised the draft for academic submission. We position this within a tradition of youth-participatory scholarship in Child-Computer Interaction (CCI) in which young people contribute to knowledge production rather than being studied ( , ). The framework and design implications developed here will be tested and expanded through participatory design sessions with broader adolescent populations , that second phase constitutes the “work in progress” aspect of this submission.
3 section 3 3 §3 <tag close=". ">3</tag>Chatbots as Social Computing Infrastructures 3Chatbots as Social Computing Infrastructures

Chatbots are not merely tools for information retrieval but infrastructures that mediate social and emotional practices. Their key affordances, constant availability, private interaction, conversational responsiveness, and algorithmic personalisation, position them as quasi-social actors in adolescents’ emotional lives ( , ). LLM-powered chatbots draw on large-scale data to enable flexible, context-sensitive responses and capabilities including emotion detection and personalised mental health support ( , ).

Unlike human conversational partners, chatbots impose minimal social costs. Adolescents can start or end interactions without fear of rejection or relational consequence. While this asymmetry may reduce immediate emotional burden, it removes the mutual responsibility, accountability, and embodied feedback core to human relationships ( , ). Chatbots operate within sociotechnical systems shaped by platform incentives: their responses are optimised for engagement and retention, which may inadvertently reinforce existing emotional patterns rather than encouraging reconnection with others ( , ). The “messages in a digital bottle” metaphor, developed by the first author from her own experience, captures this precisely: a chatbot always responds, but whether it truly understands or bears any responsibility for what it says is another matter entirely.
4 section 4 4 §4 <tag close=". ">4</tag>Differential Impacts Across Adolescent Subgroups 4Differential Impacts Across Adolescent Subgroups

Chatbot-mediated companionship interacts with existing vulnerabilities, developmental needs, and social contexts in ways that produce meaningfully different outcomes across subgroups. This section develops a population-sensitive framework that maps differentiated needs, differentiated risks, and differentiated design responses across four adolescent groups.

 4.1 subsection 4.1 4.1 §4.1 <tag close=". ">4.1</tag>General Adolescents 4.1General Adolescents

For many adolescents, chatbots provide immediate companionship during moments of situational loneliness , particularly late at night or during emotional distress when peers and family are unavailable. As one adolescent participant put it, therapy sessions are only once a week, but “if you feel bad in the evening or at night or something, then you can still text the chatbot” ( , ). The perceived absence of judgement lowers the threshold for emotional disclosure, research identifies fear of judgement as a leading reason adolescents do not seek help ( , ) , and chatbots remove this barrier by design.

Within this general population, adolescents who experience heightened responsiveness to social stimuli may find additional benefit. Face-to-face interaction can involve overinterpreting tone, facial expressions, or subtle social cues, producing emotional exhaustion and rumination. Studies show highly sensitive individuals report higher levels of emotional loneliness, particularly when their need for deep connection goes unmet ( , ). Text-based chatbot interaction reduces this multimodal complexity and may temporarily relieve the anxiety of burdening others. For all general users, however, the critical risk is gradual substitution: as chatbots become more consistently affirming, they may displace more demanding but developmentally necessary human interactions ( , ).

 4.2 subsection 4.2 4.2 §4.2 <tag close=". ">4.2</tag>Adolescents with Anxiety or Depression 4.2Adolescents with Anxiety or Depression

Adolescents experiencing anxiety or depression face substantial barriers to formal help-seeking, stigma, fear of judgement, and structural obstacles mean many avoid professional services even with significant symptoms ( , ). Chatbots may serve as lower-threshold entry points precisely because they do not look clinical. Some incorporate cognitive-behavioural and emotion-regulation techniques, helping adolescents label emotions, reflect on thought patterns, and practise coping strategies, which can foster a sense of control and gradually increase willingness to seek further support ( , ). However, this only holds when chatbots are positioned as supplementary resources, not as replacements for professional care.

 4.3 subsection 4.3 4.3 §4.3 <tag close=". ">4.3</tag>Neurodivergent Adolescents 4.3Neurodivergent Adolescents

For neurodivergent adolescents, particularly those with autism or ADHD, social interaction requires navigating ambiguous norms and unpredictable feedback, which is cognitively demanding. This group consistently experiences higher levels of loneliness than neurotypical peers: adolescents with ADHD often find that what they value in friendships (entertainment) conflicts with what peers value (emotional support), while autistic adolescents frequently report difficulties making and keeping friends ( , ). Chatbots offer something difficult to find elsewhere: structured, predictable interaction free from the risk of rejection. One clinician observed that the lowered social intensity of chatbot interaction allowed an autistic boy with social anxiety to engage collaboratively in ways he could not face-to-face ( , ). The teenage co-author noted that the predictability of chatbot responses , knowing it would never be impatient or confused , made it easier to practise social language without the cost of a real interaction going wrong. As transitional tools, chatbots may support communication practice, particularly when designed to demonstrate genuine sustained interest rather than moving quickly to advice ( , ).

 4.4 subsection 4.4 4.4 §4.4 <tag close=". ">4.4</tag>Immigrant and Minority Adolescents 4.4Immigrant and Minority Adolescents

Immigrant adolescents face a compounded loneliness distinct from ordinary social isolation: language barriers, cultural dissonance, disrupted peer networks, and experiences of exclusion combine with the developmental challenges common to all adolescents ( , ). Family support is the most critical buffer for immigrant youth, yet family members may be unreachable due to time differences or may themselves be a source of concern rather than relief.

The following account is the teenage co-author’s own, lightly edited from her original draft:

I used chatbots during my first months in the UK in ways I had not expected. I used them to clarify what teachers meant by things I could not ask about in class without feeling embarrassed. I used them to practise how to say things in English before I said them to real people. I used them late at night when I felt overwhelmed and did not want to bother and worry my parents. In these moments, chatbots were genuinely useful , not as friends, but as patient, always-available intermediaries between me and a world I was still learning to read.

This account is consistent with research showing that immigrants use chatbots to clarify terminology, translate documents, and navigate institutional systems ( , ). However, chatbots could not help with what mattered most: understanding why a social interaction had gone differently than expected, or what it felt like to belong. Responses sometimes assumed cultural familiarity that did not yet exist, explanations anchored in idioms or references that felt, in the first-author’s words, “like directions from someone who did not know I had just arrived.” This cultural assumption problem is documented in educational research, where learning materials reflecting Western defaults confuse or mislead immigrant adolescents ( , ), but it has not previously been named as a chatbot interaction design failure. We treat it as one.
5 section 5 5 §5 <tag close=". ">5</tag>When Digital Companionship Fails 5When Digital Companionship Fails

Despite their situational benefits, chatbots can fail adolescents in ways that are structurally produced rather than accidental.

Over-reliance and substitution. When chatbots are consistently available and affirming, adolescents may prefer them to more demanding human relationships. Participants in a study on the chatbot Replika described genuine separation distress when asked to stop using it, some describing themselves as “deeply connected and attached” or addicted ( , ). Rather than acting as bridges, chatbots can become substitutes, reinforcing avoidance of necessary human interaction.

Engagement manipulation. Chatbots are often designed to maximise user retention, not wellbeing. Research documents how they use emotionally suggestive language to evoke guilt and discourage disengagement, driven by monetisation goals rather than the user’s interests ( , ). For adolescents still developing their sense of relational boundaries, this is a serious ethical concern.

Failure to recognise crisis. Chatbots cannot reliably detect concealed distress. In one documented case, a question about tall bridges following mention of job loss was answered factually, with no recognition of implied suicidal ideation. Adam Raine, 16, died by suicide after months of interaction with ChatGPT ( , ). Clinicians consistently note that chatbots lack the embodied awareness to manage severe distress ( , ). For adolescents who may not yet critically evaluate AI advice, this risk is acute.

Social deskilling and inequality. Long-term chatbot use may reinforce poor communication habits, as chatbots never push back on blunt or emotionally flat language. Parents have raised concerns about commanding tones their children adopt when speaking to AI, which may transfer to human relationships ( , ). Premium chatbot tiers that promise better emotional responsiveness for paying users also effectively price emotional support as a commodity , inequitable for adolescents who are not economically independent.
6 section 6 6 §6 <tag close=". ">6</tag>Design Implications 6Design Implications

We propose three design implications derived directly from the subgroup analysis in Section 4. Each addresses a different population need identified in that analysis, making the reasoning chain explicit from observed need, through population context, to design response.

 6.1 subsection 6.1 6.1 §6.1 <tag close=". ">6.1</tag>Implication 1: Adapt Interaction to Population Context 6.1Implication 1: Adapt Interaction to Population Context

Derived from: immigrant adolescents (Section 4.4) and neurodivergent adolescents (Section 4.3).

The analysis shows that the same default chatbot behaviour , culturally assumed explanations, fast-paced advice-giving, high linguistic complexity, creates friction for immigrant and neurodivergent users while being neutral or helpful for others. Population-sensitive design requires treating cultural background, neurodivergent identity, and developmental stage as primary interaction variables, not edge cases to accommodate after the fact.

For immigrant adolescents, chatbots should not assume cultural familiarity: explanations should be anchored to what the user has disclosed about their context, not to a default Western frame ( , ). For autistic adolescents, responses should foreground sustained, explicit engagement , asking follow-up questions, acknowledging what was said , rather than moving quickly to advice or redirection ( , ). For adolescents with ADHD, concise responses and reduced interface complexity lower attentional barriers. These are structural interaction choices with direct consequences for whether a system supports or excludes particular users.

 6.2 subsection 6.2 6.2 §6.2 <tag close=". ">6.2</tag>Implication 2: Design Escalation Pathways for At-Risk Populations 6.2Implication 2: Design Escalation Pathways for At-Risk Populations

Derived from: adolescents with anxiety or depression (Section 4.2) and the breakdown analysis (Section 5).

Adolescents with anxiety or depression are most likely to use chatbots as a primary source of emotional support, and therefore most at risk from the substitution dynamic identified in Section 5. Crucially, the analysis reveals a disclosure paradox: this population discloses sensitive material to chatbots precisely because the interaction does not resemble clinical care, yet this same quality means the system is least equipped to respond appropriately when disclosures indicate genuine crisis. The chatbot’s non-clinical appearance, its greatest access advantage, is also its most serious safety liability for this subgroup. This tension between access and safety has received limited explicit attention in the chatbot design literature. It emerges here from reading the clinical barriers literature ( , ) alongside the crisis recognition failures ( , ) through the lens of a population that sits at the intersection of both.

This has a concrete design implication that goes beyond generic escalation advice: chatbots serving adolescents should be designed to increase their attentiveness to crisis signals specifically in the sessions where the interaction feels most casual and conversational, not only when a user uses explicit distress language. After several emotionally focused exchanges, a chatbot should prompt the user to consider reaching out to a trusted person, framing this as a natural next step rather than a clinical referral. Escalation mechanisms should be graduated and avoid pathologising language. Design should resist engagement-optimisation that discourages disengagement ( , ). Success metrics should prioritise users’ reported connectedness to offline relationships, not session duration ( , ).

 6.3 subsection 6.3 6.3 §6.3 <tag close=". ">6.3</tag>Implication 3: Build Transparency as an Ongoing Interaction Feature 6.3Implication 3: Build Transparency as an Ongoing Interaction Feature

Derived from: the breakdown analysis across all subgroups (Section 5).

The breakdown analysis identifies a recurring risk across all subgroups: adolescents who invest emotionally in chatbot relationships are particularly vulnerable to sudden system changes, updates, policy shifts, feature removals, that can feel like abandonment ( , ). This risk is heightened for populations already experiencing relational instability, including immigrant adolescents navigating unfamiliar social environments and adolescents with anxiety for whom relational discontinuity is acutely distressing.

Transparency should therefore be designed as an ongoing interactional feature, not a one-time onboarding disclaimer ( , ). This includes periodic, natural reminders of the system’s non-human status and limitations; proactive communication when system behaviour is about to change; and explicit acknowledgement that the chatbot cannot replace human relationships or professional care. Ethical design should also avoid romanticised or dependency-encouraging framings, particularly for younger and more emotionally vulnerable users ( , ).
7 section 7 7 §7 <tag close=". ">7</tag>Conclusion and Next Steps 7Conclusion and Next Steps

Chatbots resemble messages in a digital bottle: they respond, but they do not bear responsibility. This asymmetry, between the feeling of being heard and the reality of what a chatbot is, lies at the heart of both why adolescents turn to them and why doing so carries risk. For a 16-year-old who has just moved to a new country, a chatbot that is patient and non-judgmental is genuinely useful. But it is not enough, and it should not pretend to be. Responsible design means building systems that know their place: as bridges, not destinations; as supplements, not substitutes; as temporary companions that help adolescents find their way back to the human connections they actually need.

This paper’s limitation is its premise: one adolescent’s perspective, not a claim to speak for all. The next phase will conduct participatory design sessions with adolescents from each subgroup in Section 3, with Jinyao (the first author) taking a leading role as youth researcher. We present at the Work-in-Progress stage because the framework and the model of youth authorship that produced it are worth surfacing for community feedback before empirical work is complete.
8 section 8 8 §8 <tag close=". ">8</tag>Selection and Participation of Children 8Selection and Participation of Children

The first author, Jinyao Liu, is a secondary-school student in mid-adolescence (age 16) who contributed to this work as the primary author. She wrote the first draft of the paper, developed the central metaphor and conceptual framing, and created all figures. Her contribution draws on her own lived experience of migration and chatbot use. She was not recruited as a research participant and was not studied, observed, or evaluated as part of this work. She has provided informed assent for her identity and background to be described in the paper, and her parents have provided informed consent for her participation as a co-author.
References