# Towards Real Smart Apps: Investigating Human-AI Interactions in Smartphone On-Device AI Apps

**Authors:** Jason Ching Yuen Siu, Jieshan Chen, Yujin Huang, Zhenchang Xing, Chunyang Chen  
**arXiv:** [2307.00756](https://arxiv.org/abs/2307.00756)  
**Date:** 2023-07-03  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2307.00756v1/

## Key points

- Deep learning techniques have led to on-device AI features in smartphone apps
- AI features enable advanced tasks like speech translation, attracting users and increasing market competitiveness
- Designing good interaction for AI features is crucial for usability and understandability
- Existing guidelines and tools do not adequately address the challenges of AI features
- The authors conducted an empirical study to explore user-AI interactions in mobile apps
- They identified 255 distinct AI features and summarized 759 implementations into three primary interaction pattern types
- A multi-faceted search-enabled gallery was developed based on their findings
- A user study demonstrated the usefulness and effectiveness of their findings
- Effective interactions for on-device AI features enhance user experience and make apps more competitive in the market.

## Summary

The emergence of deep learning techniques has led to the integration of on-device AI features in smartphone apps. These AI features enable advanced tasks such as speech translation, which not only attract users but also increase market competitiveness. However, designing a good interaction for these AI features is crucial to ensure their usability and understandability. AI features pose unique challenges, including sensitiveness to input, dynamic behaviors, and output uncertainty. Unfortunately, existing guidelines and tools do not adequately address these challenges. This was confirmed through informal interviews with professional designers who pointed out that current guidelines either do not cover AI features or are focused on mobile apps in general. To address these issues, the authors conducted the first empirical study to explore user-AI interactions specifically in mobile apps. Their objective was to understand the current status of on-device AI usage by investigating 176 AI apps out of a total of 62,822 apps. Through their analysis, they identified 255 distinct AI features and summarized 759 implementations into three primary interaction pattern types. Building upon their findings, the authors developed a multi-faceted search-enabled gallery that incorporates their insights. The results of a user study conducted on this gallery demonstrated the usefulness and effectiveness of their findings. Overall, this research sheds light on the importance of designing effective interactions for on-device AI features in mobile apps. By addressing the unique challenges associated with these features – such as sensitiveness to input, dynamic behaviors and output uncertainty – developers can enhance user experience and make their apps more competitive in the market.