Springe direkt zu Inhalt

Justin Sudbrak:

Designing a Data Donation App for Everyday LLM Conversations

Discipline
Human-AI Interaction / Interaction Design
Degree
Bachelor of Science (B. Sc.)

Contents

This thesis focuses on collecting end users’ interactions with large language models (LLMs), including prompts, model responses, and optional contextual annotations, to support research on human–AI interaction. The aim is to develop a data donation app through which users can voluntarily contribute selected conversations, understand how their contributions will be used, and control what they disclose. The project is situated at the intersection of human-centered AI, usable privacy, and participatory data practices within CHI and CSCW. Recent CHI research highlights the distinctive privacy challenges of conversational LLM platforms: personal information is generated and inferred through ongoing interactions, while mechanisms for accessing, editing, deleting, and sharing data introduce ambiguities around control and ownership (Li et al., 2026). These findings motivate investigating donation interfaces specifically designed for conversational data rather than treating donation as a simple file-upload task. 
Research on data donation emphasizes that meaningful participation requires more than a one-time consent decision. Gómez Ortega et al. (2025) propose supporting donors in exploring their data, setting boundaries, gaining value from participation, and maintaining communication with researchers. Related CHI research shows that users’ perceptions of sensitivity can change when they examine individual interactions alongside patterns and inferences derived from their data (Gómez Ortega et al., 2023). Together, these findings suggest that a donation app should help users inspect and contextualize their conversations before deciding what to share. 
CSCW research provides complementary perspectives on consent and collective responsibility. Leimstädtner et al. (2025) investigate consent interfaces that support decisions aligned with users’ values rather than maximizing disclosure. Walquist et al. (2025) examine the challenges of donating data representing multiple people, raising questions about whose consent matters. Transferred to LLM conversations, these perspectives motivate attention to information about colleagues, family members, or clients that may appear in a donor’s chat history.

Problem Statement

The design challenge is to translate these insights into a usable donation workflow for LLM conversations. Users should be able to understand the research purpose, inspect their data, remove sensitive content, and make informed sharing decisions without excessive effort.
This thesis will investigate the tension between meaningful donor control and the contextual richness needed for research. Its contribution should extend beyond a functional collection tool to provide empirically grounded design knowledge about consent, privacy, and participation in human–AI research.

Research Question

How can a data donation app enable end users to make informed, privacy-conscious decisions about sharing their LLM conversations while preserving useful context for human–AI interaction research?

Sub-questions

  1. What motivations, concerns, and expectations shape users’ willingness to donate LLM conversations, including conversations containing information about other people?
  2. Which interaction techniques best support conversation selection, sensitive-content review, redaction, and understanding of intended data use?
  3. How does a guided donation workflow affect consent comprehension, perceived control, usability, and the alignment between users’ sharing intentions and their actual disclosures?

Approach

  1. Literature and requirements analysis: Review CHI and CSCW research on data donation, usable privacy, consent, and human–AI interaction. Elicit needs of LLM users from this research. Conduct interviews with researchers to identify the minimum contextual information required for research.
  2. Design: Develop and iteratively refine low- and high-fidelity prototypes. 
  3. Implementation: Build a functional web-based prototype supporting one conversation-export format initially. Process raw exports locally before transmitting only user-approved data. 
  4. Evaluation: Compare a basic, ethically adequate selection-and-consent workflow with a guided workflow incorporating contextual explanations and structured review. Use synthetic conversations with seeded sensitive information for controlled tasks, supplemented by a study using participants’ own conversations. 
  5. Discussion: Derive design implications for participant-centered research infrastructure in CHI and CSCW.

References

  • Gómez Ortega, A., Bourgeois, J., & Kortuem, G. (2023). What is Sensitive About (Sensitive) Data? Characterizing Sensitivity and Intimacy with Google Assistant Users. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems.(
  • Gómez Ortega, A., Bourgeois, J., Hutiri, W. T., & Kortuem, G. (2025; first published online 2023). Beyond data transactions: A framework for meaningfully informed data donation. AI & Society, 40, 405–422.
  • Leimstädtner, D., Sörries, P., & Müller-Birn, C. (2025). Designing Value-Centered Consent Interfaces: A Mixed-Methods Approach to Support Patient Values in Data-Sharing Decisions. Proceedings of the ACM on Human-Computer Interaction.
  • Li, Z., Wu, Y., Li, Y., Gui, X., & Luo, Y. (2026). Privacy Control in Conversational LLM Platforms: A Walkthrough Study. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems.
  • Walquist, E., Datey, I., Zheng, W., Zhou, X., Berishaj, K., McDonald, M., Parkhill, M., Zhu, D., & Zytko, D. (2025). Collective Consent: Who Needs to Consent to the Donation of Data Representing Multiple People? Proceedings of the ACM on Human-Computer Interaction.