"I Can't Reply with That": Characterizing Problematic Email Reply Suggestions

Title of the Paper

“I Can’t Reply with That”: Characterizing Problematic Email Reply Suggestions

Paper Information

• Field of Study: User interface design and AI-assisted communication in HCI
• Keywords: email, CMC, AI-MC, smart reply, algorithm auditing, AI-assisted writing, contextual politeness, socio-linguistics

Research Background and Issues

• Identified Problems or Challenges:

  • Email reply suggestion systems may provide inappropriate suggestions, affecting users' social interactions.
  • Current research often focuses on technical implementation while neglecting social contextual issues in user experience.
  • There are direct impacts (e.g., users' perception of suggestions) and indirect impacts (e.g., misunderstandings caused by suggestions).

• Significance:

  • Email is a widely used communication tool, with smart reply suggestions accounting for approximately 10% of total emails in major email services.
  • Inappropriate suggestions can undermine users' trust in the platform and affect communication efficiency.

• Research Motivation and Related Work:

  • Research on smart reply technology primarily focuses on technical performance metrics, such as click-through rates.
  • There is a lack of systematic studies on how suggestions are perceived in social contexts.
  • The study involves the impact of language, culture, and social relationships on communication.

Solution

• Proposed Methods or Solutions:

  • Develop a mixed-method framework, including qualitative interviews and controlled experiments, to identify issues in email suggestions.
  • Design experiments to systematically examine the impact of structural features of emails (e.g., opening and closing) and social relationships on the perception of suggestions.

• Innovations:

  • Highlight the social complexity of the issue, complementing previous research focused on technical implementation.
  • Provide a combined approach of quantitative measurement and qualitative insights, generating a manually annotated dataset of email scenarios.

• Implementation Steps and Key Techniques:

  • Phase 1: Conduct semi-structured interviews to identify issues and relevant themes (e.g., excessive positivity, cultural assumptions, inappropriateness).
  • Phase 2: Design and run controlled experiments to evaluate how suggestions are perceived in different social and content contexts. Collect user feedback and refine suggestions accordingly.

Research Findings

• Specific Findings:

  • Identified four main issues: excessive positivity, semantic mismatch, gender assumptions, and socio-cultural misalignment.
  • Social relationships and context significantly influence the perception of email suggestions.
  • Provided a dataset containing problematic scenario descriptions and user-corrected suggestions, offering a basis for design optimization.

• Advantages:

  • Offers a detailed user experience perspective, expanding the understanding of email reply suggestions.
  • Research findings are applicable to other AI-mediated communication systems.

• Experimental and Evaluation Results:

  • Significant scoring biases observed across different types of scenarios.
  • Adjustments to suggestions primarily involved adding vocabulary, especially polite terms like “thank you” and “sorry.”
  • Certain social relationships (e.g., superiors) were deemed to require more cautious adjustments to suggestion content.

• Limitations and Future Directions:

  • Current research only covers English email scenarios, lacking validation in multilingual and cross-cultural contexts.
  • Data may be limited by differences between experimental settings and real-world usage scenarios.
  • Future studies could explore long-term user behavior changes with AI suggestion systems and examine usage differences across device platforms.

Conclusion

This study demonstrates that current email reply suggestion systems still fall short in addressing social relationships and nuanced communication needs. Future development should more carefully consider user personalization and social network contexts while exploring ways to reduce cultural bias and prejudice. The proposed framework can be applied to evaluate other AI-assisted communication technologies.