The growing adoption of interactions with Large Language Models (LLMs) has become a societal issue that requires a deeper understanding. Trust enables engagement and adoption, and is critical for safe human-AI collaboration. It is shaped not only by LLM capabilities but also by interaction intent and users’ pre-existing AI attitudes, familiarity, and literacy. Previous research shows that text-based chatbots can become addictive for certain interaction types. While prior usage and AI literacy positively correlate with emotional dependence, attitudes can influence trust even before actually engaging with a system. This study examines how different types of interaction (task-oriented vs. reflexive) affect user trust in AI, while taking into account their characteristics. Participants engaged with ChatGPT through prompts that had been pilot-tested in a previous research. For each interaction type, assessments were administered at two phases (pre, post). The study employed a 2 (Interaction type) × 2 (Phase) within-subjects design. In addition, baseline measures of user characteristics (age, gender, AI attitudes, familiarity, literacy) were collected. A total of 110 participants from diverse nationalities and occupations completed the study. Despite their non-expert backgrounds, participants reported high AI literacy, especially in critical appraisal, and most used LLMs for both professional and personal purposes. Linear Mixed Models revealed that pre-existing attitudes towards AI and age significantly predict trust. Interaction type and the interaction between assessment phase and interaction type were strongly significant across all models, while order of interacting was not. Overall, AI trust was higher for task-oriented interactions. However, reflexive interactions led to a significant post-exposure increase, suggesting that direct engagement can enhance trust in specific contexts. Attitudes, shaped by prevailing societal discourse, appear central to trust formation. AI literacy did not predict trust directly but may still foster more responsible and critical AI use, calling for further exploration. The findings highlight that trust is not an outcome determined by technological performance but dynamically shaped at the intersection of AI design, interaction intent, and user context. It demands educational initiatives and public discourse to foster responsible and critical engagement with LLMs and other AI agents.

--

Large Language Models,Trust,Artificial Intelligence,Literacy,attitudes

• Psicologia - Ciências Sociais