Content-Driven Local Response: Supporting Sentence-Level and Message-Level Mobile Email Replies With and Without AI

Research Background and Issues

• Issues and Challenges:
  The authors identified key challenges in current mobile email applications: the lack of flexibility in existing AI integration methods prevents users from balancing efficiency and control. Additionally, the design of current applications leads to a disconnect between AI functionalities and user decision-making. For example, full-text generation may result in inappropriate suggestions, while sentence-by-sentence suggestions might require excessive manual input.

• Significance:
  Mobile email is a crucial form of personal and professional communication in modern society. However, due to limited screen space and low input efficiency, users need more effective solutions in complex or time-sensitive situations while maintaining control over the content of their responses.

• Research Motivation and Related Work:
  Inspired by the principles of microtasks, this study aims to explore how design changes can enable AI to better support users' writing decisions. Related work has compared sentence-level and email-level suggestions but has yet to investigate how to flexibly combine the two.

Solution

• Method and Approach:
  The authors proposed the concept of "Content-Driven Local Response" (CDLR), a design that allows users to directly respond to specific sentences while reading emails. CDLR introduces flexible options to integrate AI functionalities into sentence-level responses, enabling users to choose whether to invoke AI for suggestions and refine responses at the email level.

• Innovations:

  • Using sentence selection as an entry point for expressing user intent.
  • Supporting elastic workflows: spanning from fully manual writing to highly AI-dependent response generation.
  • Integrating contextual content and AI functionalities on mobile devices with limited screen space.

• Implementation Steps and Techniques:

  • Local Response Interface: Users can select specific sentences in an email and input responses or request AI suggestions via a pop-up interaction window.
  • Step-by-Step Workflow: Users progress from local responses to draft-level global replies, with the option to let AI refine the text.
  • Technical Implementation: A prototype was developed in a React web application, utilizing the Llama 3 language model to generate suggestions. It supports both sentence-level and email-level generation while optimizing interface design for enhanced interaction on mobile devices.

Research Outcomes

• Specific Results:

  • Efficiency and Quality Improvements: CDLR significantly reduced user input and error rates while enabling users to compose longer replies.
  • Flexibility Support: Users could dynamically adjust the degree of AI involvement in their workflow, selecting interaction methods that best suited their personal needs.
  • Data-Driven Insights: Log records and user feedback revealed three main workflow types—fully manual, partially AI-assisted sentence generation, and fully AI-refined email-level responses.

• Advantages:
  Compared to full email generation (MSG), CDLR offers a lower "AI footprint," allowing users greater control over response content. Most users found this mode more suitable for tasks requiring detailed and personalized replies.

• Experiments and Evaluation:
  In experiments, users responded to emails using CDLR, MSG, and fully manual modes. CDLR scored higher in user control and quality evaluation compared to MSG, while both outperformed non-AI environments in efficiency ratings.

• Limitations and Future Directions:

  • Limitations:
    • In fast-paced tasks, users may prefer AI-assisted satisfactory choices, potentially leading to information omissions.
    • The simulated scenarios in this study may not fully reflect the dynamic complexity and priority shifts of real-life email tasks.
  • Future Directions:
    • Investigating CDLR's performance in more complex email tasks, such as multi-issue emails or scenarios with extensive background information.
    • Studying the long-term effects of CDLR usage and the impact of social and cultural factors on user response frameworks.
    • Combining MSG and CDLR to create a unified model supporting flexible transitions in multitasking contexts.

In summary, the authors demonstrated how redesigning AI-based mobile email response interfaces with microtask principles at the core can enhance flexibility and user control. This work offers significant insights for the fields of human-computer interaction and AI integration design.