Co-designing value sensitive conversational agents: framework, toolkit, and design guidelines
File(s)
Author(s)
Sadek, Malak
Type
Thesis
Abstract
As Conversational Agents (CAs) are increasingly deployed in sensitive domains such as healthcare and education, aligning them with human values remains a critical challenge. Co-design practices using translational resources have engaged stakeholders in CA design, but rarely address the The Value Alignment Problem (VAP). This PhD thesis leverages these practices, and Value-Sensitive Design (VSD), to investigate VAP mitigation in CAs through research that included: systematically reviewing 52 CA co-design studies, revealing inconsistent stakeholder and value inclusion, leading to recommendations for incorporating both during CA co-design; systematically analysing 7 AI design processes and frameworks, identifying gaps and proposing guidelines for improved VSD integration; conducting design workshops with 12 AI practitioners, establishing best practices for toolkit design decisions to support VSD; and interviewing 30 AI and CA practitioners, identifying barriers to stakeholder collaboration and value integration.
Findings showed that effectively embedding intended values into CAs requires eliciting users’ values and translating them into technically useful design artifacts for CA creators. The Value-Sensitive Conversational Agent Framework and Toolkit were developed, enabling CA creators and users to co-design three translational resources, facilitating value embodiment within CA prototypes. A mixed-method evaluation including 3 design workshops with 21 CA creators and users and 13 follow-up interviews, as well as a survey of 406 CA users, demonstrated that framework and toolkit usage enabled effective value elicitation and CA user empowerment, practical integration into CA creators’ workflows, and significantly better CA prototype alignment with intended values.
Practically, this research offers a structured approach for CA creators to elicit and embed user values in CA prototypes. Theoretically, it extends VSD by operationalising it using co-design, demonstrating how translational resources can bridge the gap between abstract values and concrete system design. It positions co-design as a practical means to embed stakeholder values in AI, contributing to the value alignment discourse.
Findings showed that effectively embedding intended values into CAs requires eliciting users’ values and translating them into technically useful design artifacts for CA creators. The Value-Sensitive Conversational Agent Framework and Toolkit were developed, enabling CA creators and users to co-design three translational resources, facilitating value embodiment within CA prototypes. A mixed-method evaluation including 3 design workshops with 21 CA creators and users and 13 follow-up interviews, as well as a survey of 406 CA users, demonstrated that framework and toolkit usage enabled effective value elicitation and CA user empowerment, practical integration into CA creators’ workflows, and significantly better CA prototype alignment with intended values.
Practically, this research offers a structured approach for CA creators to elicit and embed user values in CA prototypes. Theoretically, it extends VSD by operationalising it using co-design, demonstrating how translational resources can bridge the gap between abstract values and concrete system design. It positions co-design as a practical means to embed stakeholder values in AI, contributing to the value alignment discourse.
Version
Open Access
Date Issued
2025-03-02
Date Awarded
2025-11-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Mougenot, Céline
Calvo, Rafael
Sponsor
Leverhulme Centre for the Future of Intelligence
Grant Number
RC-2015-067
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)