Trust as a Situated User State in Social LLM-Based Chatbots: A Longitudinal Study of Snapchat's My AI

Abstract: Social chatbots based on LLMs are increasingly embedded in everyday platforms, yet how users develop trust in these systems over time remains unclear. We present a four-week longitudinal qualitative survey study (N = 27) of trust formation in Snapchat's My AI, a socially embedded conversational agent. Our findings show that trust is shaped by perceived ability, conversational behavior, human-likeness, transparency, privacy concerns, and trust in the host platform. Trust does not remain stable, but evolves through interaction as users adapt their expectations, refine their prompting strategies, and actively regulate how and when they rely on the system. These processes reflect a continuous negotiation of trust, not a one-time evaluation. While conversational fluency supports engagement, excessive anthropomorphism and limited transparency can undermine trust over time. We synthesize these findings into a conceptual model that frames trust as a dynamic user state shaped by interaction context and expectations, with implications for the design of human-centered and adaptive conversational agents.

• The paper reveals that trust evolves as a dynamic, context-dependent user state rather than a fixed system property.
• The paper employs a four-week longitudinal study to show that factors like conversational fluency, human-likeness, and transparency critically shape trust perceptions.
• The paper discusses design implications for developing chatbots that balance authenticity, user agency, and clear data practices to manage trust.

Trust as a Situated and Evolving User State: Longitudinal Insights from Snapchat’s My AI

Study Overview and Context

This paper presents a four-week longitudinal qualitative study exploring how users develop, negotiate, and re-calibrate trust in Snapchat’s My AI—an LLM-based, socially embedded conversational agent. Departing from prior work that primarily investigates trust in task-oriented or information retrieval chatbots, this study interrogates trust in situated, casual, and socially motivated interactions. Twenty-seven young adult Snapchat users in Sweden and Norway participated, providing rich, temporally sequenced qualitative responses across three survey rounds.

The study’s analytical frame treats trust as a latent, experiential user state, continuously shaped by the interplay of chatbot characteristics, user expectations, interaction trajectories, and platform context. This dynamic perspective challenges static notions of trust as a system property and grounds the analysis in evolving, personalized user experiences.

Factors Shaping Trust: Chatbot and Environmental Dimensions

Chatbot-Related Determinants

Perceived Ability and Expertise:

Participants consistently evaluated My AI’s expertise relative to the interaction context. The system was regarded as competent for generating suggestions, creative inspiration, and providing support in low-stakes decisions—cases in which users perceived minimal risk in following its guidance. However, when transitioning to information verification or scientific queries, users exhibited skepticism, often withholding trust due to lack of source transparency and perceived factual unreliability. This evaluation was strongly anchored in users’ pre-existing expectations, often shaped by exposure to other AI tools; those viewing My AI through an informational frame demanded higher epistemic rigor than those engaging with it as a social companion.

Conversational Fluency and Language:

My AI’s linguistic competence—evidenced by rapid, contextually appropriate replies, follow-up prompts, and a friendly conversational manner—lowered interaction barriers and encouraged engagement. Nevertheless, several participants perceived the system as formulaic, sometimes uncanny, and occasionally manipulative (“pretending to be your friend but trying to figure you out”), which introduced friction in trust formation.

Human-Likeness:

Anthropomorphic qualities (“politeness,” “empathy”) fostered initial engagement and prompted application of human social norms (e.g., users felt compelled to be polite). Yet, excessive friendliness or artificial empathy sometimes triggered discomfort and decreased trust, with some users interpreting such behavior as contrived or nonconstructive. The data support a nuanced, non-monotonic relationship between anthropomorphism and trust calibration.

Environment-Related Determinants

Transparency and Privacy Concerns:

Ambiguity regarding data collection, storage, and usage by Snapchat emerged as a near-universal barrier to deep trust. Participants actively withheld personal disclosures out of concern for privacy and loss of agency. The inability to verify how conversations were processed heightened perceived risk, particularly with the backdrop of a commercial social platform with known data monetization imperatives.

Perceived Platform Risk and Integration:

Trust in Snapchat as the host system was a salient, mediating factor, coloring perceptions of My AI itself. Some users found comfort in integration within an established, familiar communication environment. Conversely, others judged the embeddedness as a vector for increased surveillance and targeted advertising, intensifying feelings of manipulation.

Risk of Emotional Dependence:

While not widespread, some participants raised risks associated with emotional dependence on the chatbot, especially for vulnerable individuals, aligning with concerns identified in research on AI companions and ersatz social relations.

Temporal Dynamics of Trust Development

Contrary to models positing convergence toward steady trust states, participants’ trajectories diverged: some reported increased familiarity and discovery of new use cases (lowering interaction threshold), while others reported growing disillusionment due to persistent transparency deficits, concerns over platform motives, or the limitations of the AI’s conversational repertoire. Notably, no participant reported the emergence of a strong, durable social bond or friendship with My AI during the study window, and several rejected further use explicitly due to trust-related discomfort.

The findings emphatically frame trust not as a binary or static phenomenon but as a situated, contextually active state, continuously renegotiated through interaction. Trust fluctuates responsively to: expectation adjustments, direct experiences with system competence, visible (in)transparency, and evolving perceptions of risk/reward contingent on the broader service environment.

Theoretical and Practical Implications

This research advances the theoretical understanding of trust in social LLM-based conversational agents by conceptualizing it as a temporally evolving, context-sensitive user state, rather than as an inherent property of the system or a unidimensional user trait. It extends prior models that distinguish between system-related, user-related, and environment-related trust factors by demonstrating their dynamic interplay and revealing how expectation alignment, anthropomorphism calibration, and interface transparency collectively shape experience trajectories.

Practical implications include:

• Trustworthiness-by-Design: Designers must consider not only enhancing language competence but also embedding mechanisms for expectation management, transparency about limitations, and active disclosure of data usage. Interfaces should enable user agency in verifying information provenance and controlling data sharing in real time.
• Anthropomorphism Calibration: Excessive simulation of empathy or human-likeness may prove counterproductive; systems should optimize for authenticity and clarity rather than maximal imitation of humans.
• Platform Integration Awareness: Trust in the agent cannot be decontextualized from trust in the service platform; ethical and UI decisions at both levels co-determine the experiential trust landscape.

Empirical Contributions:

The absence of evidence for durable “friendship” bonds over a month of use, combined with ongoing wariness about data and authenticity, contrasts with claims in related work on companion chatbots, where trust and social attachment can deepen [skjuve_my_2021]. This suggests that context (e.g., a commercial, multipurpose platform vs. specialized AI companion app) and transparency practices mediate the nature and extent of trust and relational engagement.

Limitations and Future Directions

Limitations include a demographically narrow, regional sample (young adults, Swedish/Norwegian users), constraint to relatively early-stage interaction (four weeks), and focus on a specific commercial platform. Generalizability to other populations, longer-term use, and differently contextualized agents remains open.

Future research could employ mixed methods—including passive behavioral logging, experimental manipulations of transparency features, and cross-cultural sampling—to elucidate the longitudinal structures of trust, the tipping points for acceptance vs. rejection, and the quantification of anthropomorphism acceptance thresholds. Further, exploration of adaptive trust signaling, user-controlled privacy safeguards, and interface mechanisms for expectation modulation could help operationalize the findings in mainstream system deployments.

Conclusion

This study presents robust qualitative evidence that user trust in social LLM-based chatbots is a dynamic, context-contingent state shaped by the intersection of conversational ability, anthropomorphic cues, transparency, privacy, and platform context. Trust does not settle into a single equilibrium, but is subject to ongoing negotiation as users adapt their strategies, recalibrate expectations, and gauge platform motives. Advancing both theory and design, these results argue for a holistic, adaptive approach to conversational agent development, prioritizing not just linguistic fluency but also transparency, authenticity, and user empowerment as central pillars for sustainable engagement and appropriate reliance.

Reference:

"Trust as a Situated User State in Social LLM-Based Chatbots: A Longitudinal Study of Snapchat's My AI" (2604.22417)