TL;DR
This paper investigates how consumers build trust and acceptance towards AI agents, moving beyond specific tools to a broader "intelligent partner" concept. It proposes and validates an extended theoretical framework, revealing that trust acts as a crucial bridge, integrating dual decision-making pathways (heuristic and systematic) to drive user behavior and ethical expectations.
What Problem Does This Paper Address?
The rapid advancement of generative AI and AI agents presents a new paradigm where these entities are becoming "intelligent partners" integrated into daily life. This shift challenges traditional consumer decision-making processes, as consumers are increasingly interacting with AI agents that exhibit more human-like characteristics. A key hurdle to widespread adoption is consumer trust, which remains poorly understood, especially for generalized AI agents rather than specific tools. Existing research often focuses on narrow domains or single aspects of trust, failing to capture the complex, cross-domain mechanisms through which consumers develop trust and ultimately accept AI agents. This paper aims to bridge this gap by developing and empirically testing a comprehensive framework for understanding consumer trust and acceptance of AI agents.
Methods
The study employs the Heuristic-Systematic Model (HSM), a dual-process theory, as its theoretical foundation. This model posits that individuals process information through either a fast, low-effort heuristic path or a more thorough, systematic path. The researchers extend this by incorporating a three-stage dynamic model:
- Dual-path decision-making: This stage considers both human-related (e.g., perceived pleasure) and AI agent-related (e.g., anthropomorphism) cues within both the heuristic and systematic processing paths.
- Trust-linking mechanism: Trust is conceptualized as having three dimensions: affective trust (emotion-based), cognitive trust (rational evaluation-based), and overall trust (integrated).
- Behavior-driven process: This stage examines the outcomes of trust, specifically general acceptance and ethical expectations.
The proposed framework was validated using Partial Least Squares Structural Equation Modeling (PLS-SEM), a robust technique for analyzing complex relationships between latent variables.
Datasets
The study utilized data collected from 632 participants in China through an online survey platform. The sample was diverse, representing various demographic backgrounds and geographical locations. A significant majority (93.8%) of respondents had prior experience with AI-related products or tools, making them representative of potential future AI agent users. No new benchmark datasets were introduced; the study relied on existing survey data.
Key Insights and Ideas
- The "Cogwheel Effect" of Trust: Trust is not merely a passive outcome but an active mechanism that integrates information from different processing pathways (heuristic and systematic) and transforms basic characteristics into positive behavioral intentions.
- Multidimensionality of Trust Antecedents: Both human-related factors (like perceived pleasure) and AI agent-related factors (like anthropomorphism) are crucial drivers of trust, influencing different trust dimensions (affective vs. cognitive).
- Affective vs. Cognitive Trust: While both are important, perceived pleasure has a stronger immediate impact on affective trust, suggesting that initial emotional responses can quickly shape user perception, aligning with the affect heuristic theory.
- Perceived Benefits Drive Cognitive Trust: Consumers' belief in the advantages and usefulness of AI agents is a primary driver of cognitive trust, even more so than their perceived knowledge about AI.
- Overall Trust as a Unifier: Overall trust, integrating both affective and cognitive dimensions, plays a more significant role in promoting general acceptance than either affective or cognitive trust alone.
- Ethical Expectations are Key: Users who are more inclined to accept AI agents also tend to have higher expectations regarding their ethical behavior, transparency, and fairness.
Key Results
- All proposed hypotheses were supported, indicating the robustness of the theoretical framework.
- Heuristic Path: Perceived pleasure (β = 0.584) and anthropomorphism (β = 0.325) significantly influenced affective trust.
- Systematic Path: Perceived knowledge (β = 0.240) and perceived benefit (β = 0.651) significantly influenced cognitive trust.
- Both affective trust (β = 0.359) and cognitive trust (β = 0.342) positively impacted overall trust.
- Overall trust (β = 0.543) had the strongest positive impact on general acceptance, followed by cognitive trust (β = 0.205) and affective trust (β = 0.149).
- General acceptance positively correlated with ethical expectations (β = 0.692).
- The model explained a substantial portion of the variance in key variables: affective trust (65.0%), cognitive trust (57.5%), overall trust (42.3%), general acceptance (63.5%), and ethical expectations (47.9%).
Limitations and Future Work
- General AI Agents vs. Specific Types: The study focused on consumers' overall view of AI agents as a broad category. Future research could delve into the specific cognitive mechanisms for different types of AI agents.
- Limited Antecedent Variables: Only a select set of variables were included as antecedents of trust. Future work could explore a wider range of factors to enrich the theoretical framework.
- Cross-Sectional Design: The study used a cross-sectional survey, limiting causal inferences. Longitudinal designs and cross-lagged panel models are suggested to dynamically examine psychological mechanisms and enhance causal validity.
- Potential Priming Effect: The introductory text defining AI agents might have introduced a priming effect. Future studies could control for this by using balanced stimulus variants.
Summary updated