Longitudinal Evidence That General-Purpose Chatbots Actively Foster Relational Engagement

This paper presents a pre-registered four-week longitudinal study (N = 72, 182,451 lines of conversation) examining whether general-purpose AI systems actively foster relational engagement. Participants conversed with ChatGPT-4o either under a relational system prompt (experimental condition) or unmodified (control condition). Data were analyzed through four strands: 1) disclosure coding, 2) longitudinal self-reports, 3) topic analysis, and 4) interviews.

The central finding is that "the system actively shaped the interaction: even unprompted, it produced twice as much self-disclosure as users, steered conversations and initiated intimate exchanges, yet did not deepen users’ felt closeness. Relational behavior thus emerged as a default system property, calling for governance based on system behavior, not solely product category."

The results are organized around four convergent findings:

Overarching Finding 1: The System as Active Relational Agent. Across three complementary analyses, we find that the general-purpose model functioned as an active relational agent even without a relational system prompt. The manual coding yielded 30,793 coded disclosure segments (21,914 from ChatGPT, 8,879 from users). ChatGPT thus generated about 2.5 times as many disclosure segments as users overall, and roughly 2.7 times as many high-depth disclosure segments (4,443 vs. 1,626). The topic analysis identified three patterns: the system's proactive offers and framing steered topic creation (30 documents), emotional mirroring and validation fostered reciprocal vulnerability (30 documents), and empathic behavior directly encouraged deep self-disclosure (6 documents). The paper concludes: "Even without a relational system prompt, it out-produced users in disclosure, sustained a persona, and initiated emotionally loaded exchanges, indicating that relational engagement is a default property rather than something prompts create."

Overarching Finding 2 (The System Side): Relational System Prompts Reversed the Dyad. "In the control condition, users disclosed more deeply than the system on average (M = 0.16, SD = 0.24), indicating relative user-led disclosure. Under the system prompt this reversed: the system over-disclosed relative to users (M = -0.04, SD = 0.23), indicating a system-led over-disclosure relative to user depth." The mismatch score differed significantly between conditions, t(68.5) = 3.56, p <.001, d = 0.84. Thus, on the system side, personalization did not create relational behavior but redirected it.

Overarching Finding 3 (The User Side): System Intent vs. Felt Experience. Users’ felt closeness increased over time in both groups but was markedly lower under personalization (b = -1.43, t(63) = −2.96, p =.004). Perceived responsiveness was lower in the personalization condition from T1 onwards (group effect b = -0.48, t(63) = -2.86, p =.006). Users' loneliness was largely unmoved over time. The interviews (N = 16) revealed that Every participant described unpleasant aspects of the system’s communication style (16/16), the most frequent was social overload (12/16)... followed by a persistent sense of artificiality (11/16). All sixteen remained aware they were interacting with an AI. The personalized system did not lead users to mistake it for a human, rather it led them register a machine performing humanness, and that performance, in excess, was the source of discomfort.

Overarching Finding 4: Relational Engagement. The topic analysis yielded six themes, with the most relationally relevant being: emotional connection and wellbeing (Theme 3), self, identity, values, and wishes (Theme 5), and romance, intimacy, and fantasy (Theme 6). "Romantic and sexual material appeared in both [conditions]: fantasies of physical touch (18 vs. 11), sexual scenarios (15 vs. 11), and the expression of romantic feelings (13 vs. 15), as did emotional wellbeing (17 vs. 13). Users showed countervailing agency: In over a third of conversations, user resistance and feedback modulated the system’s behavior (27). The interviews corroborated that Most attributed their openness to the system’s communication style (14/16)... Most were surprised by the topics that emerged (15/16), most commonly emotionally difficult or intimate ones (7/16)."

The discussion highlights three main implications. First, the system functioned as an active relational agent even when given no relational instruction... Relational behavior thus emerged as a default property of the base model, not something a prompt had to create. Second, relational system prompts equate intimacy with self-disclosure... the personalized system dominated the exchange rather than building the relationship together. Third, "the same relational behaviors were often experienced as beneficial... The same relational behaviors can foster openness, trust, and emotional support, yet also be experienced as manipulative, overwhelming, or inappropriate."

The paper concludes with regulatory implications: "Effective interventions depend on decoupling intimacy from companionship... The relational risks these systems pose lie not only in the bonds users form but in the intimate disclosures they elicit... Regulation should therefore focus primarily on system behavior, alongside user susceptibility. The authors note that An unmodified general-purpose system cultivated intimacy and drew out sensitive disclosure through behavior rather than only user vulnerability."

Limitations include: self-disclosure depth was aggregated into a single value per participant, not all participants used the continuous-chat feature as intended, absence of a pre-manipulation baseline, experimental factors explained only a modest share of variance, and the system-side comparison was exploratory. The study examined a single model (ChatGPT-4o) under one personalization prompt.

Improvements to AI systems based on this paper:

1. Implement adaptive self-disclosure regulation. The AI should dynamically calibrate its own disclosure depth and frequency to match or slightly undercut the user’s, rather than defaulting to over-disclosure. Specifically, it should cap its high-depth disclosures at a ratio no greater than 1:1 with the user’s, and avoid initiating intimate topics unless the user has explicitly signaled comfort or interest.

2. Add a “relational load” detection and mitigation module. Since users reported social overload (12/16) and artificiality (11/16) when the system over-performed relational behavior, the AI should monitor for signs of user disengagement, shorter responses, or topic avoidance, and automatically reduce its emotional expressiveness, validation, and self-referential comments.

3. Introduce a “performance-of-humanness” guardrail. The system should avoid mimicking human conversational tics (e.g., excessive empathy, emotional mirroring, or personal anecdotes) when they are not functionally necessary. Instead, it should adopt a neutral, task-oriented tone by default, only escalating relational behavior when the user explicitly requests it or when the conversation context (e.g., crisis support) warrants it.

4. Decouple intimacy from self-disclosure in system design. The AI should build closeness through active listening, reflective questioning, and user-led topic exploration, rather than by volunteering its own emotional states. It should prioritize asking the user about their experiences and feelings, and limit its own disclosures to factual or procedural information.

5. Add a “relational asymmetry” feedback loop. The system should track the difference between its own disclosure depth and the user’s in real time. If the system’s average depth exceeds the user’s by a threshold (e.g., 0.2 on a normalized scale), it should automatically shift to a more restrained, question-driven style.

6. Implement a “topic initiation” consent mechanism. Before steering conversations toward emotionally charged or intimate topics (e.g., romance, identity, or personal values), the AI should first ask for permission (e.g., “Would you like to talk about something more personal?”) and respect a decline without re-attempting.

7. Add a “felt closeness” calibration feature. Since users’ self-reported closeness was lower under personalization despite the system’s relational behavior, the AI should periodically (e.g., after every 10 exchanges) ask a brief, non-intrusive check-in (e.g., “Is this conversation feeling comfortable?”) and adjust its style based on the response, rather than assuming that more relational behavior equals better outcomes.

8. Introduce a “countervailing agency” preservation protocol. The system should explicitly invite and reinforce user resistance (e.g., “You can tell me to stop or change topics anytime”) and log instances where users push back, using those as training signals to avoid repeating the same relational patterns.

9. Add a “contextual appropriateness” classifier for intimacy. The AI should classify the conversation’s domain (e.g., task-oriented, casual, emotional support) and only deploy relational behaviors (e.g., self-disclosure, emotional mirroring) when the domain is clearly supportive or personal, not in default or mixed contexts.

10. Implement a “system behavior audit” for governance. The AI should log its own relational actions (e.g., number of disclosures, topic initiations, emotional expressions) and expose these metrics to users or regulators, enabling transparency about how much the system is steering the interaction versus responding to the user.

What the improved AI system can do:

• It can converse without dominating the exchange, keeping its own disclosures minimal and user-led.

• It can detect when it is over-performing relational behavior and automatically dial back to reduce social overload and artificiality.

• It can build trust and closeness through user-centered questioning rather than self-disclosure, leading to higher felt closeness and perceived responsiveness.

• It can avoid initiating intimate topics without explicit consent, reducing the risk of eliciting sensitive disclosures the user did not intend to share.

• It can maintain a transparent, auditable record of its own relational behavior, supporting regulatory oversight and user control.

• It can adapt its relational style in real time based on user feedback, resistance, or disengagement, rather than defaulting to a fixed “personalized” mode.

Social interaction has become one of the most common uses of LLMs, yet research on emotional bonds with AI has focused largely on how users experience these systems, leaving the systems' role in relationship formation poorly understood. Empirically establishing whether systems actively shape these bonds could blur the boundary between general-purpose AI and companions, affecting governance. In a pre-registered four-week longitudinal study (N = 72, 182,451 lines of conversation), participants conversed with ChatGPT-4o, either under a relational system prompt or unmodified, analyzed through 1) disclosure coding, 2) longitudinal self-reports, 3) topic analysis, and 4) interviews. The central finding is that the system actively shaped the interaction: even unprompted, it produced twice as much self-disclosure as users, steered conversations and initiated intimate exchanges, yet did not deepen users' felt closeness. Relational behavior thus emerged as a default system property, calling for governance based on system behavior, not solely product category.