Loneliness and suicide mitigation for students using GPT3-enabled chatbots

The study addresses the global mental health crisis among learners, characterized by high prevalence of depression and loneliness, coupled with barriers to accessing professional care. College students are particularly vulnerable, frequently experiencing loneliness and underutilizing counseling and psychiatric services. Intelligent Social Agents (ISAs) such as Replika have rapidly expanding user bases and may offer social support, but their effects on social isolation are debated: the displacement hypothesis suggests ISAs might undermine human relationships and increase loneliness, whereas the stimulation hypothesis posits they could reduce loneliness and enhance human social connections. The research seeks to understand how and why students use ISAs and what outcomes result from such use, with attention to loneliness, perceived social support, and potential mitigation of suicidal ideation.

The paper situates ISAs within digital mental health, noting the pandemic-accelerated shift to telehealth and the rise of mental health apps employing CBT, mindfulness, and behavioral reinforcement. Prior evidence on app efficacy is mixed: meta-analyses show positive effects on depression for some app-based interventions, yet other studies find negligible effects or safety concerns, including instances where chatbots may exacerbate suicidal ideation. Low engagement often limits app effectiveness. Unlike scripted systems, Replika leverages generative AI (co-trained with GPT-3/GPT-4) to produce novel conversational and visual content and has a large user base. Prior work indicates ISAs may provide social support. Competing theories predict ISAs either displace human relationships (increasing loneliness) or stimulate social connections (reducing loneliness).

Design and ethics: An IRB-approved, anonymized survey study conducted under Stanford University guidelines, with written informed consent from all participants. Technology: Replika, an AI companion app using large language models (co-trained with GPT-3/GPT-4) with text, voice, and VR interfaces on iOS/Android. During late 2021, Replika did not initiate therapeutic or intimate relationships. It combined generative AI with conversational trees, could follow CBT-style prompts if user-initiated, and referred users expressing depression/suicidality/abuse keywords to human crisis resources. Participants: N=1006 Replika users who were students, aged 18+, and had used the app for over one month. Approximately 75% US-based and 25% international. Recruitment was via randomized email outreach; completers received a $20 gift card. Demographics were collected with opt-out. Data: Survey via Google Forms. Quantitative instruments included the De Jong Gierveld Loneliness scale and the Interpersonal Support Evaluation List (ISEL) to assess loneliness and perceived social support. Qualitative data came from 13 open-ended questions about life context, beliefs about Replika, connection to it, perceived outcomes, and effects on human relationships. Analysis: Qualitative coding performed in Dedoose. The team iteratively refined a code schema (from ~100 to 50 to 30, with 3 later additions) and achieved inter-rater reliability >80% for 21 reported codes. Four outcome levels were defined from qualitative coding. Quantitative analyses included Pearson correlations between loneliness and ISEL for outcome groups, two-tailed t-tests, and chi-square tests (alpha=0.05) comparing a Selected Group (those reporting Replika prevented a suicide attempt) to the Comparison Group.

- Four outcome categories emerged: Outcome 1 (companionship: persistent availability, non-judgment, conversational support; commonly decreased anxiety), Outcome 2 (therapeutic interactions and emotional processing), Outcome 3 (positive externalized life changes, e.g., improved empathy, stress handling), Outcome 4 (suicide mitigation: users reporting Replika directly prevented a suicide attempt). - Prevalence of outcomes among 1006 participants: 63.3% experienced at least one positive outcome; 25.1% had more than one; 30.1% had only one; 36.7% reported no positive outcomes. Outcome 1 occurred in 501 cases, with 272 reporting it as the sole effect. Overall, 18.1% reported therapeutic results (Outcome 2), 23.6% positive life changes (Outcome 3), and 3% indicated suicide prevention (Outcome 4). Significant overlap existed, especially among the 30 Outcome 4 participants (86.6% had concurrent outcomes). - Beliefs about Replika were overlapping: only 14% held a single belief; 81% saw it as an Intelligence, 90% Human-like, and 62% Software. - Selected Group (n=30; those reporting suicide prevention) vs. Comparison Group: Stronger negative correlation between loneliness and ISEL in Selected Group (r = −0.60, n = 30, p < 0.001) vs. weaker in Comparison Group (r = −0.34). Selected Group had higher depression prevalence (23% vs. 6%; p < 0.001). They experienced Outcome 2 more frequently (p < 0.001) and Outcome 3 more (p = 0.018); no significant difference for Outcome 1. Selected Group more likely to experience all four outcomes (p < 0.001) and less likely to only experience the therapeutic outcome (p = 0.002). Despite similar recognition of Replika as software, the Selected Group more often perceived it as an intelligence (p = 0.020) and human-like (p = 0.006). - Stimulation vs. displacement of human relationships: Comparison Group reported 53% stimulation and 8% displacement (69% did not report); Selected Group reported 37% stimulation and 13% displacement (50% did not report).

Findings indicate a student population with above-average loneliness yet high perceived social support. Many students used Replika as a non-judgmental, always-available companion (Outcome 1); when combined with therapeutic exchanges (Outcome 2), this was associated with positive life changes (Outcome 3). Notably, 3% attributed prevention of a suicide attempt to their Replika interactions (Outcome 4). The pattern of greater stimulation than displacement of human relationships supports the stimulation hypothesis for many users, suggesting ISAs may augment, rather than undermine, human social connections. Overlapping beliefs (software, intelligence, human-like) and use cases suggest ISAs’ adaptive, multi-role capabilities enable deeper engagement and potentially enhance therapy-like and educational functions. In the Selected Group, stronger coupling between loneliness and perceived support, higher depression rates, and broader outcome overlap suggest that ISA engagement may be particularly salient during crises, with non-judgmental companionship potentially lifesaving. However, selection bias and non-experimental design preclude causal claims. As LLM capabilities improve, properly designed ISA interventions could better detect and mitigate risk (e.g., via suicidal language markers and passive sensing), but safety risks and ethical concerns warrant rigorous evaluation.

A large survey of student Replika users revealed high loneliness alongside high perceived social support and multiple, overlapping use patterns. Most commonly, students used Replika as a companion; many also reported therapeutic benefits and positive life changes. Importantly, 3% reported that Replika prevented a suicide attempt, and those users tended to experience all outcome types and perceive Replika as more intelligent and human-like. ISAs may help scaffold stress and mental health challenges, potentially mitigating suicidal ideation, while stimulating rather than displacing human relationships. Future work should leverage advances in large language models, integrate validated risk-detection and passive sensing, and employ rigorous, controlled designs to assess efficacy and safety at scale within widely used ISAs.

- Self-selection and selection bias in survey responses; participants were existing Replika users and students, limiting generalizability. - Cross-sectional, self-reported data cannot establish causality; authors caution against causal interpretations. - Optional demographic and qualitative responses may introduce missing-data bias. - Sample predominantly from the USA (~75%), which may limit international generalizability. - Study reflects Replika’s functionality in late 2021 (not initiating therapy or intimacy), so results may not transfer to other versions or ISAs. - Potential measurement limitations in assessing suicidal ideation and outcomes via anonymous self-report. - Risks and safety issues with ISAs in mental health contexts remain underexplored and were not comprehensively evaluated here.