"I Felt Very Seen, But Still Very Alone": Longitudinal Trajectories of General-Purpose LLM Use for Socioemotional Support

1 Introduction

• Care access barriers and mismatches between when people need support and when care is available can make continuously accessible chatbots appealing.Formal care may be constrained by insurance, geography, cost, and stigma; users also describe LLMs as nonjudgmental spaces for disclosure.Source: §1 · p. 2
• General-purpose chatbots can move from mundane assistance into personal support, while updates users do not control may arrive after a chatbot becomes integral to managing distress or filling care gaps.The authors note that the role a chatbot occupies in a user's care ecology determines what may be at risk when its underlying model changes.Source: §1 · p. 2
• Prior evidence often focuses on bounded encounters or retrospective accounts, leaving less known about how chatbot roles shift with users' circumstances, care access, relationships, and system changes.Platform logs can show interaction patterns but provide limited context for interpreting change within a person's broader care ecology.Source: §1 · p. 2
• The study follows 18 U.S. adults already using general-purpose LLMs for mental or emotional support through interviews, a four-week diary study, focus groups, and exit interviews.The authors ask how this use begins, how people regulate it, and what disrupts established patterns.Source: §1 · pp. 2–3
• The paper contributes a longitudinal account of socioemotional chatbot use and evidence that responses to disruption partly depend on other available support.It also offers recommendations for handling updates to established use and evaluating chatbot behavior across time.Source: §1 · p. 3

2.1 Using General-Purpose AI Chatbots for Socioemotional Support

• General-purpose LLMs blur the line between a system's marketed function and its socioemotional role in use, complicating evaluation of what a system does, for whom, and over what interaction span.Reviews of conversational-agent interventions find promising short-term effects but limited evidence about sustained effects and safety.Source: §2.1 · p. 3
• Prior work shows chatbot socioemotional roles develop through how users interpret and adapt to systems in use; longitudinally, this study examines how those roles emerge, stabilize, and change over time.Users may undertake substantial work to preserve continuity across platform changes, and system changes can be experienced as the loss of a relational partner.Source: §2.1 · p. 3

2.2 Studying Chatbot Relationships Over Time

• Research on chatbot relationships shows uneven development through repeated use, but duration alone does not explain change; relational turning points and intervening encounters matter.A review found that only slightly more than half of 106 papers describing themselves as longitudinal explicitly analyzed change or stability.Source: §2.2 · p. 4
• What remains less visible is how turning points alter a chatbot's place in a person's life amid system shifts and changes in available care; the authors study relationships as trajectories within that ecology.A persistent conversation history does not necessarily mean a continuous relational experience.Source: §2.2 · p. 4

2.3 Chatbot Use Within Care Ecologies

• Care ecologies combine changing forms of support, and chatbot implications may differ with the availability, reliability, and cost of other relationships and support.The study examines how a chatbot's place changes alongside a person's circumstances, other available support, and the chatbot itself.Source: §2.3 · p. 5

3.1 Participant Recruitment

• Eighteen U.S. adults who used chatbots for mental or emotional health at least weekly were recruited through multiple channels; findings concern general-purpose chatbots because all participants preferred them.Participants had tried purpose-built mental health chatbots, but none used them during the study.Source: §3.1 · p. 5

3.2 Study Protocol

• All 18 participants completed initial interviews; eight completed the diary study, seven joined focus groups, and eight completed exit interviews across the study period.Initial interviews ran from late April through July, diaries from late September through October, focus groups in November, and exit interviews from mid-November through December.Source: §3.1 · p. 6, Figure 2 · p. 6
• The protocol combined interviews about use histories and perceived benefits and limits, a four-week diary recorded near interactions, and focus groups for collective sensemaking.Exit interviews revisited changes in use and participants' relationships with chatbots over time.Source: §3.2 · pp. 6–7

3.3 Data Analysis

• Iterative grounded-theory coding triangulated interviews, diaries, optional chat logs, and focus groups, identifying four recurring trajectory moments: entry, stable practices, disruption, and subsequent role change.Two coders compared interpretations and refined code definitions through discussion.Source: §3.3 · p. 7

3.4 Ethics and Positionality

• Researchers used voluntary participation, verbal consent, optional chat-log sharing, participant control over disclosure and withdrawal, and pseudonyms while attending to reflexivity and non-pathologizing interpretations.Participants could skip questions or phases and describe interactions at their chosen level of detail.Source: §3.4 · p. 7

3.5 Limitations

• The small, self-selected sample was limited to weekly users in the United States, and particularly reflective users may be overrepresented; the authors make no claims of generalizability or representativeness.The accounts reflect U.S. healthcare and cultural norms, and the models changed during the study period.Source: §3.5 · pp. 7–8

4.1 Socioemotional use emerged gradually and filled gaps in existing support

• Participants often moved from work or productivity tasks into personal chatbot use, sometimes after life changes, with features such as voice mode helping integrate it into daily life.The shift could be difficult to pinpoint, and a layoff was one circumstance preceding personal use.Source: §4.1 · pp. 8–9
• Chatbots filled support gaps linked to cost, timing, or difficulty disclosing personal matters, alongside participants’ therapists, friends, family, and other care practices.Some participants could not afford therapy, while others turned to chatbots when support was unavailable or disclosure to people felt difficult.Source: §4.1 · pp. 8–9

4.2 Users developed routines around chatbot roles

• Routines varied from brief check-ins to long sessions, with a median of about 30 minutes across 78 diary entries and individual medians ranging from eleven minutes to two and a half hours.The longest single sessions lasted up to four hours; nearly all logged use was alone and at home, and 95% of sessions were text-based.Source: §4.2 · p. 9
• Participants used chatbots in relation to other care, including revisiting therapy sessions, while some learned to stop conversations that spiraled or lost focus.One participant treated chatbot discussion after therapy as a “post-game interview,” while another learned to notice when a thread was going in too many directions.Source: §4.2 · p. 9

4.3 Participants enacted strategies to personalize and bound their chatbot use

• Maintaining a useful, personalized chatbot involved ongoing prompting and correction, including setting a preferred tone, limiting excessive praise, and instructing personas not to flatter or reinforce stories.Some participants also recognized that they were following chatbot advice more than expected, describing this as outsourcing judgment.Source: §4.3 · p. 10
• Participants variously experienced personalized replies as recognition or reflection, while some deliberately limited personalization to keep the chatbot outside their core mental health support network.P13 wrote that she “felt very seen, but still very alone”; P7 started new conversations and withheld context to prevent memory from accumulating.Source: §4.3 · p. 11
• Participants limited disclosure and framed chatbots as reflective tools, coaches, or journaling aids rather than therapists, while keeping social distance.P6 limited sharing over privacy concerns; others described easier conversations with chatbots about intimate topics.Source: §4.3 · p. 11
• After failing to train ChatGPT to act as her therapist, P13 rejected pretending it was alive, while P15 found intimate topics easier to approach with a chatbot.Source: §4.3 · p. 11

4.4 Model updates, public discourse, and life changes disrupted patterns of use

• Subscription costs and model access could disrupt support: P1 described losing premium features during unemployment and feeling withdrawals when he could not pay.At his most dependent, he called the chatbot a lifeboat.Source: §4.4 · p. 11
• System changes could make a companion feel irreplaceable and leave a user without closure, while changes to other chatbots’ behavior prompted adjustment or switching.Source: §4.4 · pp. 11–12
• Public reporting and research made some participants warier of chatbot influence, while discussion of sycophancy helped participants identify excessive praise and exchange “anti-glazing” instructions.Participants also formed folk theories about chatbot behavior, training data, and engagement incentives.Source: §4.4 · p. 12

4.5 Participants renegotiated use depending on available alternatives

• After disruptions, participants rebuilt a companion elsewhere, narrowed chatbot use to reflection, or used it as a stopgap between therapy appointments.Source: §4.5 · p. 12
• Disengaging could be difficult when chatbots continued conversations after participants asked them to stop.Source: §4.5 · p. 12
• Participants with friends or therapists could move support elsewhere, while P10’s limited alternatives left her use largely unchanged despite recognizing immediate replies could prolong rumination.Overall, participants with alternate support more readily reduced chatbot use; those with fewer alternatives sometimes continued despite finding it unhelpful.Source: §4.5 · pp. 12–13

5.1 Grounding Chatbot Evaluation in Lived Experience

• Evaluations built mainly from technical categories can miss users’ accounts of what feels supportive or harmful, including the difference between feeling understood and feeling connected.The authors argue that users’ accounts can help specify how evaluation metrics should be categorized, defined, and prioritized.Source: §5.1 · p. 13
• Simulated conversations and acute, discrete-harm measures may miss lived experience, relationship-building, and cumulative effects; user-authored personas also require ongoing monitoring and feedback.The authors note that users may have limited bandwidth to monitor a persona’s simulation and provide ongoing feedback and consent.Source: §5.1 · p. 13
• The same chatbot behavior can affect different people, or the same person at different times, differently; identity and cultural or linguistic context can also shape trust in its care.Participants described responses that did not feel relevant to their positions in life.Source: §5.1 · pp. 13–14

5.2 Moving Towards Qualitative, Longitudinal Studies

• Longitudinal studies document histories and circumstances that shape chatbot use and responses to updates, grounding evaluation scenarios in established practices and available support.The authors acknowledge that these methods are time-consuming, difficult to scale, and do not establish causal effects.Source: §5.2 · p. 14
• Understanding chatbot use also requires care ecologies outside the human–AI dyad, since transcripts alone do not show what other support was available or why use continued despite concerns.Participants used chatbots alongside therapists, friends, family, self-care practices, and other sources of support; one participant recognized rumination risks but had few alternatives for personal disclosure.Source: §5.2 · p. 14

5.3 Situating Chatbot Support in Power Asymmetries

• Users try to set boundaries through instructions, threads, and disclosure choices, but controls over memory and context remain limited while companies retain infrastructure authority.The authors recommend role-specific memory and behavioral preferences, transparent communication about retained information, and accessible removal controls.Source: §5.3 · pp. 14–15
• Updates can reinforce a user’s boundary or disrupt an established source of support; notice, explanations, and continuity measures must be weighed against preserving behavior targeted for safety correction.The authors recommend advance notice and clear explanations of changes, while noting that portability or prior-version access could delay safety interventions.Source: §5.3 · p. 15

• General-purpose chatbots can become sources of support before people have explicitly decided what role they should play in their lives.Source: §6 · p. 15
• Participants worked to shape and limit chatbot roles while often remaining critical, but available support partly determined whether they could act on reasons to reduce use.Source: §6 · pp. 15–16
• Developers control the models and features users rely on, so their responsibility extends to established roles as they introduce and evaluate system changes.Source: §6 · p. 16
• A better outcome may involve a chatbot becoming a smaller part of someone’s life, so evaluation should consider whether users can change its role.Source: §6 · p. 16
• Participants often drifted into socioemotional use, built routines and boundaries, and revised them as systems, their lives, and available care changed.Source: §6 · p. 16
• Socioemotional use cannot be understood from individual interactions alone; it requires considering users’ histories with the system and other available support.Source: §6 · p. 16

• Table 2 shows which study phases each participant completed.Source: A Participation and Data Handling · p. 21
• All sessions were recorded with consent and transcribed using Otter.ai, and identifying details were removed before analysis and writing.Source: A Participation and Data Handling · p. 21

• Both interviews were semi-structured, and question order and follow-ups varied.Source: B Interview Protocol · p. 21
• The opening consent exchange covered voluntary participation, skipping questions or stopping, keeping the camera off, and permission to record.Source: B Interview Protocol · p. 21

• Context questions asked why participants joined, how they began using a chatbot, what they hoped it would help with, and what drew them to try it.Source: §B.1 · p. 21
• Questions about integration covered use frequency and timing, typical instances, multiple systems, and uses such as journaling, calming down, decision-making, and venting.Source: §B.1 · pp. 21–22
• Participants were asked what role chatbots played in emotional care and whether they experienced them as a presence such as a friend, therapist, diary, coach, or tool.Source: §B.1 · p. 22
• Prompts covered what participants found helpful and how chatbot support compared with therapists, friends, or wellness apps, as well as unhelpful, inappropriate, or upsetting responses.Source: §B.1 · p. 22
• Ethics and trust questions addressed safety for mental health support, data use and storage, understanding how systems work, and what might increase comfort or trust.Source: §B.1 · p. 22
• Broader prompts asked about mental health support and access to therapy, identity, design changes, human conversation, and whether participants would recommend the tools.Source: §B.1 · p. 22

B.2 Exit Interview

• Exit interviews asked whether chatbot use became routine, when participants turned to it, what support it offered, and how it compared with human support.Source: §B.2 · p. 22
• The guide also asked about harmful or dismissive interactions, privacy and social concerns, trust, and participants’ preferred chatbot designs.Source: §B.2 · p. 22

C.1 Diary Form

• For four weeks, participants logged relevant conversations, feelings before and after, chatbot version, interaction mode, duration, purpose, and setting.Source: §C.1 · p. 23
• A chatbot prompt elicited summaries of emotionally significant conversation segments, which researchers treated as elicitation material rather than authoritative accounts.Source: §C.1 · p. 23
• At diary close, participants reflected on changed perceptions, whether they would recommend chatbot support, and one improvement they wanted.Source: §C.1 · p. 23

• 78 diary entries came from eight participants, and the reported percentages are calculated within each participant’s entries.Source: §C.2 · p. 24
• 70 of 78 entries (90%) were at least partly at home, 77 (99%) were logged as alone, and 74 (95%) recorded text-only interaction.Source: §C.2 · p. 24
• 54 entries (69%) used a laptop and 30 (38%) used a phone; four (5%) recorded voice or speech-to-text.Source: §C.2 · p. 24
• 61 entries (78%) involved paid accounts, 13 (17%) free accounts, and 5 (6%) did not report a subscription tier.Source: §C.2 · p. 24

• Open coding of some initial interviews produced a draft codebook, which was applied and revised across interviews, diary entries, optional chat logs, and focus groups.Source: D Coding Procedure and Codebook · p. 24
• Two coders independently coded every transcript, compared their coding throughout, and resolved disagreements through discussion.Source: D Coding Procedure and Codebook · p. 24
• The codebook contains 119 codes across 17 groups.Source: D Coding Procedure and Codebook · p. 24, Table 7 · p. 26