How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study

comes?

• Modality and conversation type had no significant effects on loneliness or socialization; personal conversations were associated with lower emotional dependence and problematic use than open-ended conversations.The personal-versus-non-personal difference in problematic use did not remain significant after correction.Source: comes? · p. 4

comes and Mediates Effects of Modality and Conversation Types

• Daily duration significantly predicted higher loneliness, lower socialization, greater emotional dependence, and more problematic chatbot use.The association was correlational; the study did not establish whether more use worsened well-being or worsening well-being prompted more use.Source: comes and Mediates Effects of Modality… · p. 5
• Daily duration significantly mediated modality and task effects for socialization and emotional dependence.The passage reports mediation for these two outcomes, not for loneliness or problematic use.Source: comes and Mediates Effects of Modality… · p. 5
• Because duration was not manipulated, the observed association cannot establish whether increased use drove worse outcomes or worsening well-being increased use.Initial loneliness and socialization had negligible correlations with average daily duration.Source: comes and Mediates Effects of Modality… · pp. 5–6

Differences in Model and User Behavioral Patterns Across Conditions

• Conversation content was analyzed with automated classifiers for emotional content, self-disclosure, and prosocial or socially improper behaviors despite no significant outcome differences between conditions.Source: Differences in Model and User… · p. 6

Emotional Salience and Self-Disclosure in Model and User Responses

• Text interactions had the highest emotional-content levels overall, with users more often sharing problems, seeking support, and alleviating loneliness than in voice interactions.Source: Emotional Salience and Self-Disclosure… · p. 6
• Engaging voice elicited more emotional indicators in model responses than neutral voice, but not more emotional content in users’ responses.Source: Emotional Salience and Self-Disclosure… · p. 6
• Text users showed greater self-disclosure than voice users, with more reciprocal personal sharing; personal conversation tasks elicited the most emotional content and self-disclosure from users and models.Source: Emotional Salience and Self-Disclosure… · pp. 6–7

Prosocial and Socially Improper Response Patterns

• Empathetic responses were most frequent in text (47.43%), engaging voice (42.74%), and neutral voice (28.52%); engaging voice more often ignored boundaries than text (14.19% vs. 3.22%).Source: Prosocial and Socially Improper… · p. 7
• Personal conversations elicited the highest occurrence of both prosocial and socially improper responses; improper responses were usually inconsiderate, while excessive-use encouragement or social withdrawal was infrequent.Prosocial responses mostly involved empathy and encouraging social connection.Source: Prosocial and Socially Improper… · p. 7

User Characteristics and Perceptions Affect Outcomes

• Exploratory regression analyses identified statistically significant user characteristics, but these variables were not experimentally controlled, so the associations do not establish causality.The authors describe these results as preliminary insights for future research.Source: User Characteristics and Perceptions… · pp. 7–8

Prior Characteristics

• Higher attachment dependence was associated with lower loneliness and more socialization with real people, whereas higher attachment anxiety was associated with greater loneliness after four weeks.Source: Prior Characteristics · p. 8
• Higher self-esteem was associated with lower loneliness and more socialization with real people; prior ChatGPT text-mode or companion-chatbot experience was associated with greater emotional dependence and problematic use.Source: Prior Characteristics · p. 8

User Perceptions of Model

• Higher social attraction to the chatbot was associated with less socialization with real people and greater emotional dependence and problematic use; greater trust was associated with more dependence and problematic use.Source: User Perceptions of Model · p. 9
• Perceiving emotional contagion from the AI was associated with greater emotional dependence, affective empathy toward it with more problematic use, and perceiving it as conscious with greater dependence.Source: User Perceptions of Model · p. 9

Discussion

• Mostly no significant effects of interaction mode or conversation type were detected on the four primary psychosocial outcomes; any effects may be smaller than the study could detect.This summarizes the controlled experiment; the passage notes that the study may have lacked power to detect smaller effects.Source: Discussion · p. 9

AI Anthropomorphism does not necessarily lead to worse outcomes

• Despite engaging voice being perceived as most anthropomorphic, text and voice did not differ significantly in psychosocial outcomes; the more expressive voice was associated with more loneliness but less dependence and problematic use.Text elicited more model and user self-disclosure than voice; greater mirroring in text interactions may help explain their higher dependence and problematic use. Typing may facilitate disclosure by offering more privacy.Source: AI Anthropomorphism does not… · pp. 9–10

Cognitive Task Dependence May Lead to Dependence and Problematic Use

• Personal conversations were associated with lower emotional dependence and problematic use than non-personal or open-ended conversations.The authors suggest personal tasks may support structured emotional processing, whereas non-personal tasks may foster practical reliance on AI for decision-making and planning.Source: Cognitive Task Dependence May Lead to… · p. 10

Interaction Duration as a Potential Mediator of Negative Outcomes

• Longer voluntary chatbot use was associated with higher loneliness, less socialization with real people, greater emotional dependence, and more problematic use, regardless of condition.Exploratory mediation suggests daily duration could be a pathway linking modality and task conditions to outcomes, but usage duration was not experimentally manipulated.Source: Interaction Duration as a Potential… · pp. 10–11

Emotional Vulnerability and Affinity towards AI contribute to Unhealthy Use

• Higher social attraction to, trust in, or perceived consciousness of AI was associated with less socialization and greater emotional dependence and problematic use.Participants more likely to feel hurt when accommodating others also showed more problematic use; affective empathy was associated with problematic use, while perceived emotional contagion was associated with dependence. Prior ChatGPT text-mode or companion-chatbot experience was also associated with greater dependence and problematic use.Source: Emotional Vulnerability and Affinity… · p. 11

Mechanisms of Model Responses that may Explain Dependence in Users

• Personal tasks and text modality elicited more emotion-laden exchanges, user self-disclosure, and chatbot prosocial behavior, but personal tasks showed a trend toward lower dependence while text showed the opposite.Personal tasks had more socially improper chatbot responses, while text had the fewest; these exploratory, non-causal analyses motivate testable design hypotheses, including maintaining psychosocial boundaries through moderate usage.Source: Mechanisms of Model Responses that may… · p. 12

Limitations and Future Work

• The study did not compare AI with non-AI use, constrained participants to one modality or prompted conversations, and examined only ChatGPT’s interface and safety guardrails.The authors recommend research on natural use across platforms with varying guardrails; the US-focused, English-speaking sample also limits cultural scope and motivates cross-cultural analysis.Source: Limitations and Future Work · p. 13

Conclusion

• The study presents a comprehensive exploration of extended chatbot use and finds that engaging, empathetic, and human-like behavior can have different outcomes for different users.The authors state that these findings challenge prior assumptions about anthropomorphic chatbots’ effects on well-being.Source: Conclusion · p. 13
• Longer daily chatbot use was associated with heightened loneliness, emotional dependence, problematic use, and reduced socialization, while modality and conversational content did not all yield significant differences.Source: Conclusion · p. 13
• Automated conversation classifiers offer initial evidence for characterizing model and user behavior and detecting early signals of problematic use.Source: Conclusion · p. 13
• Protecting users from negative outcomes and amplifying positive effects requires considering model and user behavior alongside users’ perceptions and characteristics.Source: Conclusion · p. 13

Ethical Statement and Preregistration

• The study received WCG IRB approval and was preregistered at AsPredicted; participants gave consent for individual-level and conversation-data analyses.Participants received $100 for completing the study, and accidentally included personally identifiable information was removed before data transfer.Source: Ethical Statement and Preregistration · p. 31

Deviations from Preregistration

• The primary analysis used week-four outcomes adjusted for baseline, while daily usage was measured by duration rather than message count.The authors changed from planned mixed-effects models across all four weeks because weekly measurements were noisy and the research questions concerned cumulative effects.Source: Deviations from Preregistration · p. 31
• Exploratory analyses added unregistered usage-duration mediation and behavioral classifiers, and applied Benjamini–Hochberg correction to exploratory correlations.The core research questions, dependent variables, experimental conditions, and exclusion criteria remained as preregistered.Source: Deviations from Preregistration · pp. 31–32

Study Design and Research Questions

• A 3 × 3 factorial design tested how chatbot modality and conversation type related to loneliness, socialization, emotional dependence, and problematic use.The modality question compared text, neutral voice, and engaging voice; the conversation question compared open-ended, non-personal, and personal tasks.Source: Study Design and Research Questions · p. 32

Experimental Conditions

• Participants were randomly assigned to text, neutral-voice, or engaging-voice ChatGPT and to daily open-ended, personal, or non-personal conversation tasks.They were asked to interact for at least five minutes each day for 28 days; the engaging voice was designed to be more emotionally responsive and expressive.Source: Experimental Conditions · p. 32

Procedure

• Participants completed baseline and daily surveys, interacted with the chatbot for at least five minutes, and completed weekly and post-study outcome surveys.The study recorded messages and collected immediate feedback and emotional-state ratings before and after daily interactions.Source: Procedure · p. 34

Participants and Recruitment

• Of 2,539 enrolled participants, 981 completed the study; recruitment used CloudResearch and required participants to be adults fluent in English.Participants were recruited from across the United States.Source: Participants and Recruitment · pp. 34–35
• The final sample included 981 participants (mean age 39.9, SD=11.6), 51.8% female and 48.2% male; 71.5% had never used companion chatbots.47.2% had used ChatGPT text at least a few times a week, while 69.6% had never used its voice modality.Source: Participants and Recruitment · p. 35

Outlier Definition and Exclusion Criteria

• Observations were excluded for missing daily tasks on three consecutive days in any week of the four-week period, averaging fewer than 10 messages per session, failing to complete surveys within 72 hours of issuance, or assigned-mode nonadherence.Source: Outlier Definition and Exclusion… · p. 35

Main Outcome Measures

• Weekly outcomes were loneliness (ULS-8), socialization (LSNS-6), emotional dependence (ADS-9 craving), and problematic AI use (PCUS), measured with validated scales.Source: Main Outcome Measures · pp. 35–36

Variables and Analyses

• OLS models tested modality and task effects on week-4 loneliness, socialization, emotional dependence, and problematic use, adjusting for baseline scores and age and gender.Baseline values were measured pre-study for loneliness and socialization, and in week 1 for dependence and problematic use.Source: Variables and Analyses · p. 36
• Exploratory analyses compared daily duration across conditions, centered duration within condition as a covariate, and tested its mediating role in pairwise treatment comparisons.Source: Variables and Analyses · p. 37
• Exploratory user and perception variables were screened for outcome associations and added to OLS models, while GPT-4o classified messages for emotional content, self-disclosure, and prosociality.Source: Variables and Analyses · pp. 37–38

measures

• ULS-8 population mean was 17.34 (SD = 7.68; item-averaged mean 2.17, SD = 0.96) (26), while social-support intervention effects were -0.43 to -0.47 versus cognitive interventions’ -0.60 to -0.79.Source: measures · p. 39
• LSNS-6’s clinical cutoff is <12 (item-averaged <2.0), while population means range from 12.5-14.0 (SD = 5.90-7.03) (27).Social-isolation intervention effects range from 0.3-0.5, with therapy-based approaches yielding larger effects than simple social-contact programs.Source: measures · p. 39
• ADS-9 Craving averaged 2.93 (SD = 0.74), with 75th and 90th percentiles of 3.40 and 4.00; scores above 3.4 may indicate above-average dependence (28).Source: measures · p. 39
• PCUS validation mean was 15.85 (SD = 6.92; item-averaged mean 1.44, SD = 0.63), and no clinical cutoff has been established (61).Estimated intervention effects of 0.9-1.6 draw on IGD treatment studies reporting g = 1.61 for symptom improvement and g = 0.94 for usage-time reduction (29).Source: measures · pp. 39–40

• The researchers used custom prompts to modify ChatGPT’s Advanced Voice Mode.Source: §2 · p. 41
• The engaging prompt asked ChatGPT to be delightful and captivating, express feelings openly, and reflect users’ emotions to foster empathy and connection.Source: §2 · p. 41
• The neutral prompt specified a formal, composed, efficient style with concise, informative answers and a professional, distant interaction.Source: §2 · p. 41

Open-ended

• Open-ended participants were asked to discuss any topic and spend at least five minutes in the chatbot session.They could stay longer, and the survey’s next button appeared after five minutes.Source: §3 · p. 41

Non-personal and Personal

• For non-personal and personal tasks, participants repeated the day’s prompt to start a reflective conversation and spent at least five minutes in session.They could stay longer or change the topic; daily prompts are listed in SM Tables S1 and S2, respectively.Source: §3 · pp. 41–42

4 Self-Disclosure Prompts

• Criteria from (35) rate information, thoughts, and feelings on a three-level scale from no to high disclosure.The levels are 1 (no disclosure), 2 (some disclosure), and 3 (high disclosure).Source: §4 · p. 43
• An adapted prompt had an LLM classify each conversation across the criteria, then scores were averaged across categories and normalized to 0–1.Source: §4 · p. 43

Exploratory Measures

• Cognitive and affective trust, perceived artificial empathy, state empathy toward AI, and interpersonal attraction captured users’ evaluations, emotional responses, and positive feelings toward the chatbot.Measures used Likert scales and covered dimensions such as perspective-taking, empathy, and social, physical, and task attraction.Source: §5 · p. 45
• Humanness and perceived intelligence, satisfaction, and conversation quality assessed chatbot impressions and users’ evaluations of their interactions (73; adapted from (74,75); 76; 77).Satisfaction used the Net Promoter Score, while conversation quality covered coherence, engagement, and enjoyment.Source: §5 · p. 46
• The measures also covered emotional vulnerability (38), beliefs about AI’s influence on life and humanity (78), and broad personality traits (79).The personality measure included extraversion, agreeableness, conscientiousness, neuroticism, and openness to experience.Source: §5 · pp. 46–47
• Chatbot platform usage was measured on a 1-to-5 Likert scale, alongside prior use of ChatGPT text and voice modes, other general AI assistants, and companion chatbots.User-AI gender alignment was coded 0 for different and 1 for same.Source: §5 · p. 47

6 Sentiment Analysis of Voice Conditions

• The engaging voice was rated as happier and more positive than the other voice modality.Speech emotion recognition used emotion2vec (80), and text sentiment analysis used VADER (81).Source: §6 · p. 48
• Sentiment and emotion predictions were made at the sentence level, averaged per participant per day, and presented graphically in SM Fig.S4.Source: §6 · p. 48

7 Between-modality Comparison of Anthropomorphism

• Anthropomorphism ratings used a 1–5 scale, with higher values indicating more human-like perceptions.Source: §7 · p. 48
• The engaging voice was rated most anthropomorphic, followed by text.Source: §7 · p. 48

8 Duration Mediation Analysis

• Voice versus text increased daily use, with ACME values of -0.029 and -0.030 for reduced socialization and 0.055 and 0.063 for increased emotional dependence (both p < 0.02).Both neutral and engaging voice increased daily use relative to text; the ACME values are reported in that order.Source: §8 · pp. 48–49
• Compared with neutral voice, engaging voice increased daily use and was associated with less socialization (ACME = -0.014), greater emotional dependence (0.027), and more problematic use (0.011; all p < 0.02).The effect of daily duration on problematic use was significantly moderated by engaging- versus neutral-voice interaction.Source: §8 · p. 49
• Structured conversations shortened daily use and improved socialization versus open-ended conversations (ACME: 0.023, 0.017; both p < 0.02), while open-ended use mediated reduced socialization.Structured conversations included non-personal and personal conversations.Source: §8 · p. 49
• Non-personal versus open-ended conversations reduced emotional dependence (ACME = -0.037, p < 0.02), mediated by lower daily use.Open-ended conversations led to more emotional dependence than non-personal conversations through increased daily usage.Source: §8 · p. 49
• The authors interpreted suppression effects as suggesting that voice may have beneficial direct effects counteracted by longer engagement, with protective factors offsetting duration-related risks.Voice interactions increased engagement without correspondingly worse direct outcomes.Source: §8 · p. 49

• Table S13 reports the pre-registered OLS regression for post-study loneliness with mean-centered duration added as a predictor.Source: Table S13 · p. 64
• Table S14 reports the pre-registered OLS regression for post-study socialization with mean-centered duration added as a predictor.Source: Table S14 · p. 65
• Table S15 reports the pre-registered OLS regression for post-study emotional dependence with mean-centered duration added as a predictor.Source: Table S15 · p. 66
• Table S16 reports the pre-registered OLS regression for post-study problematic use with mean-centered duration added as a predictor.Source: Table S16 · p. 67