Abstract
Augmented reality (AR) interfaces increasingly utilize artificial intelligence systems to tailor content and experiences to the user. We explore the effects of one such system — a recommender system — on user trust and perceived accuracy when recommendations are "situated" within the user's physical environment. By comparing traditional 2D web interfaces, interactive 3D models, and immersive AR, we investigate how the modality of presentation influences the psychological relationship between the user and the AI.
Introduction
Recommendation engines are the backbone of modern e-commerce, driving a significant portion of sales for giants like Amazon and views for platforms like Netflix. However, these systems have historically been confined to 2D screens. As Augmented Reality (AR) hardware matures, we are entering an era of situated recommendations, where virtual products are placed directly into a user's physical space.
The central question of this research is how this shift in modality affects user perception. Does seeing a virtual product on your own desk increase trust in the AI's suggestion, or does the novelty of AR simply mask the underlying performance of the algorithm? This study compares traditional 2D browser views, interactive 3D models, and immersive AR to understand their impact on perceived accuracy and system trust.
Methodology
The researchers conducted a $3 \times 2$ within-subjects lab study with 31 participants. The study utilized a custom e-commerce platform designed to deliver product recommendations across three distinct modalities:
- 2D Browser: Standard static images of products.
- 3D Browser: Interactive 3D models that users could rotate via a mouse.
- AR (Microsoft HoloLens): Virtual products anchored to a physical table, allowing users to physically walk around them.
Experimental Design
Participants first completed a preference survey regarding coffee mugs (size, shape, color). A Multi-Attribute Utility Theory (MAUT) algorithm then generated two types of recommendations: High Quality (HQ), which aligned with user preferences, and Low Quality (LQ), which intentionally diverged from them. To mitigate the "novelty effect" of the HoloLens, participants were given a 10-minute acclimation period before the formal study began.
Key Insights
The Value of Context
The study revealed that AR's primary advantage is contextual. By placing a virtual mug on a real table, users could better evaluate scale and utility—a "try before you buy" experience that 2D images cannot provide. This spatial context led to higher ratings for perceived accuracy.
3D vs. AR Parity
Interestingly, the 3D browser performed nearly as well as the AR interface in quantitative trust ratings. This parity is likely due to interaction control. While AR provided better spatial context, the 3D browser allowed users to easily manipulate the object (e.g., looking at the bottom of the mug) using a mouse, whereas the AR objects were fixed to a table, making full inspection more physically demanding.
Trust Dynamics
The research confirmed that trust is asymmetric: it is easily lost and slowly gained. A single low-quality recommendation significantly damaged user trust, while high-quality recommendations built trust at a much slower pace.
Results
- Perceived Accuracy: HQ recommendations were rated significantly higher in AR ($\mu = 3.4$) and 3D ($\mu = 3.2$) compared to 2D ($\mu = 3.12$).
- Trust Increase: For high-quality items, trust in the system was approximately 20% higher in AR and 3D modalities than in the traditional 2D interface.
- User Preference: Qualitative interviews showed a strong preference for AR, with participants highlighting the benefit of visualizing items in-situ.
Limitations and Future Work
While the results are promising, several limitations were identified. The first-generation HoloLens hardware had a limited field of view and produced somewhat "washed-out" colors. Additionally, the fixed position of AR objects limited user interaction. Future research should explore "pick-up" interactions where users can manipulate virtual objects by hand. Furthermore, longitudinal studies are needed to determine how these trust dynamics evolve over extended periods in home environments.
Conclusion
Situated recommendations in AR represent a significant shift in how users interact with AI. By providing physical context, AR enhances both the perceived accuracy of recommendations and the user's trust in the system. As hardware continues to improve, these immersive interfaces are likely to become a standard component of the e-commerce experience.
Summary updated