AI usage patterns are shaped by perceived gains in human agency

Abstract: As conversational AI systems become more deeply integrated into daily life, the implications for human agency are increasingly urgent to understand. AI's potential to amplify capability sits alongside risks of individual and collective disempowerment, yet empirical, ecologically-valid evidence about cumulative usage is scarce. We analyze deep ethnographic data from a study of daily AI chatbot users (n = 51) in the United States, Germany, and Singapore to illuminate conversational AI usage in situated context as a sociotechnical practice. We show that people consistently link sustained AI usage to perceived gains in individual agency. Crucially, these perceived gains often outweigh concerns about accuracy, reliability, and consistency to shape usage patterns. Our findings challenge prevailing assumptions about how and why humans use AI systems over time, suggesting that traditional trust-based models are not sufficient for explaining human behavior with conversational AI. Finally, we expose a critical tension: immediate psychological boosts to perceived agency may not necessarily translate into material effects, structural empowerment, or long-term capacity. Our results help establish a new foundation for novel behavioral frameworks, measurement tools, and AI benchmarks to ensure conversational AI strengthens human agency in substantial, sustained ways.

• The paper reveals that cumulative, multidimensional gains in human agency drive sustained AI usage, outweighing occasional technical errors.
• The methodology employs ethnographic techniques such as diary studies, in situ observations, and interviews across diverse cultural contexts.
• The study challenges traditional trust frameworks by demonstrating that users prioritize psychological empowerment over strict technical reliability.

Ethnographic Insights into Human Agency and AI Chatbot Usage Patterns

Study Overview and Methodological Framework

This paper undertakes an ethnographic exploration of daily AI chatbot usage among 51 participants across the United States, Germany, and Singapore. The study adopts a rigorous, contextually anchored qualitative methodology incorporating diary studies, semi-structured interviews, in situ observations, and follow-up interviews, oriented by principles of constructivist grounded theory. The participant cohort is composed of active daily users of systems such as ChatGPT, Gemini, Grok, Perplexity, Microsoft Copilot, and Meta AI, with research protocols emphasizing ecological validity and participant pseudonymization.

The investigatory lens reframes agency anthropologically—as the capacity for individuals to shape their environment—while acknowledging mediation by sociotechnical and sociocultural determinants. Rather than focusing on isolated human-AI interactions, the analysis foregrounds cumulative, situated patterns within broader personal and systemic contexts.

Key Empirical Findings

Perceived Agency Gains as the Primary Driver of Sustained AI Usage

Across diverse domains, participants overwhelmingly reported that cumulative engagement with conversational AI generated pronounced, multidimensional enhancements in their perceived agency. These gains manifested in five principal dimensions:

• Instrumental agency: Augmented efficacy in discrete, task-oriented domains (e.g., productivity, information retrieval)
• Cognitive agency: Enhanced clarity and organization of thought, aided by AI as a “sounding board”
• Affective agency: Improved ability to understand and regulate emotion, sometimes extending to self-concept transformation
• Relational agency: Strengthened skills in communication and navigation of interpersonal and social contexts
• Structural agency: Increased perceived competence in interacting with complex systems (e.g., healthcare, bureaucracy)

Importantly, these effects were often holistic and enduring, with individuals explicitly linking AI-mediated empowerment to material outcomes ranging from career advancement and interpersonal decisions to overcoming institutional challenges.

Error Tolerance and the Limits of Trust Models

A critical empirical observation is that perceived agency gains consistently outweighed the influence of AI failures—including those with acute personal or professional consequences—on usage patterns. Participants’ cost–benefit calculus privileged sustained or cumulative feelings of empowered capacity above traditional technical benchmarks such as accuracy or reliability. There was explicit resistance to the notion that one-time errors, or even unreliability at large, should diminish continued AI engagement.

This empirical pattern directly challenges dominant trust-based frameworks in HCI and human-automation interaction, which posit reliability as foundational for adoption and continued use. Participants instead displayed risk-balancing heuristics: if perceived agency improvements exceeded the anticipated costs of occasional errors, usage continued largely unabated.

Coexistence of Agency Gains and Critical Reservations

Despite widely reported increases in agency, participants frequently harbored substantial concerns regarding overreliance, skill atrophy, dependency, cognitive outsourcing, and potential stigmatization. Many articulated a tension between immediate psychological empowerment and the specter of long-term disempowerment, both individual (e.g., cognitive atrophy) and collective (e.g., systemic automation and existential risks).

Notably, AI-enabled agency in personal or professional spheres did not generally extend to a sense of structural empowerment. Structural agency gains were rare, with broader social, institutional, or political empowerment perceived as unaffected or even eroded.

Contextual Moderators: Situational Stability and Internal Conviction

The magnitude and nature of perceived agency gains were significantly modulated by two contextual variables:

• Situational Stability: Those experiencing high day-to-day unpredictability or lower social/structural support reported more expansive and emotionally salient agency gains from AI usage, often turning to chatbots for high-stakes relational or affective needs. In contrast, individuals with stable environments deployed AI narrowly and reported more bounded, instrumental agency effects.
• Internal Conviction: Individuals with higher self-confidence and clearer values tended to attribute AI performance failures to technical limitations, maintaining stable self-worth. Those with low internal conviction were more prone to defer to AI “expertise,” with self-evaluation oscillating alongside perceived AI performance, sometimes undermining overall self-efficacy.

Some individuals with both high situational stability and internal conviction reported no change in agency, suggesting that personal histories and sociocultural context interact nonlinearly with AI-mediated empowerment.

Theoretical and Practical Implications

Toward Post-Trust Behavioral Frameworks in Human-AI Interaction

These findings undermine the explanatory sufficiency of traditional trust-centric models for AI adoption and usage, particularly those rooted in reliability, predictability, and staged accretion of trust. Instead, the dominant mediating variable is perceived enhancement of self-capacity, suggesting a shift of evaluative focus from attributes of the machine to the outcome for the user. This reversal complicates prevailing contract- or reliability-based trust models, necessitating the development of new theoretical frameworks that foreground experiential agency and dynamic risk negotiation.

Decoupling Psychological and Structural Agency

The study raises the possibility that AI-induced psychological agency gains may often lack corresponding material or structural empowerment, generating “thin” or possibly illusory senses of enhanced capacity. This dissociation aligns with critical perspectives in social theory regarding the positional and relational nature of agency, and introduces a cautionary note regarding technological augmentation: short-term capability boosts may not translate to durable, system-level empowerment. The risk of cumulative cognitive deskilling, or “cognitive surrender,” remains salient and underexplored in longitudinal, ecologically valid settings.

Methodological Recommendations and Limitations

This ethnographic approach demonstrates unique advantages in uncovering nuanced, context-bound human-AI dynamics, relative to experimental or log-based studies. However, qualitative sampling and self-report introduce generalizability constraints, necessitating complementary large-scale quantitative investigations and cross-cultural comparative work. The observed findings are temporally and sociotechnically situated, and thus require cautious extrapolation to other populations or future AI modalities.

Implications for AI Development and Future Research

Practically, the results highlight the necessity for AI system development, evaluation, and benchmarking to integrate measures of long-term, multidimensional human agency as primary objectives. Algorithmic alignment—moving beyond narrow preference satisfaction toward substantive empowerment—appears crucial for sustained positive impact. Benchmarks such as HumanAgencyBench and efforts in sociotechnical alignment are likely to become increasingly central in both academic and practical AI work.

These findings invite several lines of future inquiry:

• Operationalizing and quantifying the relationship between subjective agency gains and objective, longitudinal capability changes.
• Exploring the mechanisms and thresholds by which perceived agency counterbalances tolerance for technical error or unreliability.
• Investigating how sociocultural, positional, and psychological factors mediate human–AI co-adaptation, with attention to risk of cognitive or affective overreliance.

Conclusion

This study provides a deeply contextualized account of how daily users integrate AI chatbots into the fabric of their lives, revealing that perceived multidimensional agency gains fundamentally shape usage patterns—frequently eclipsing traditional trust or reliability concerns. The ethnographic evidence challenges extant behavioral models in human–AI interaction, advocating for a paradigm shift that centers perceived agency and its sociotechnical determinants. At the same time, the research underscores a critical analytic distinction between psychological and structural empowerment, raising important open questions for the measurement, alignment, and optimization of AI systems to robustly support human flourishing.