Engaging people in ethical AI development: Design, dataset creation, and decision-making

Sharma, Tanusree

Permalink

https://hdl.handle.net/2142/125676 Copy

Description

Title

Engaging people in ethical AI development: Design, dataset creation, and decision-making

Author(s)

Sharma, Tanusree

Issue Date

2024-07-01

Director of Research (if dissertation) or Advisor (if thesis)

Wang, Yang

Doctoral Committee Chair(s)

Wang, Yang

Committee Member(s)

• Huang, Yun
• Miller, Andrew
• Das, Sauvik

Department of Study

Illinois Informatics Institute

Discipline

Informatics

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• AI Governance
• Security/Privacy
• Ethical AI, Democratic AI
• Dataset
• Design
• Decentralized Governance

Abstract

Major limitations of past AI development include but are not limited to, the absence of thorough documentation and traceability in design, training with a small number of specific datasets, and lack of transparent and democratic decision-making processes for deploying AI models. These limitations could lead to adverse outcomes, such as discrimination, lack of privacy and inclusivity, and breaches of legal regulations. Underserved populations such as people with disabilities and minority groups, in particular, are disproportionately affected by these design decisions. In this dissertation, I employ qualitative, quantitative, and design methods to uncover users’ expectations for privacy and security in AI systems. I then develop methods to engage users in the dataset creation and decision-making phases of AI development. More specifically, this thesis covers different facets of AI development. I start with empirical studies on how end users think about privacy/security in application domains that are powered by AI such as targeted ads and search engines. I then dive into ethical dataset creation for AI model development. Lastly, I present how to engage a diverse set of populations toward democratic decision-making regarding how AI should behave and align with people’s values. One such privacy measure in AI systems is “Data Minimization” which means data should be adequate, relevant, and limited to what is necessary for the purposes for which it is processed. By examining user reactions to data minimization in AI systems, I find that users’ assessments of privacy and security risks can be limited by bounded rationality. My analysis of the users-centric data minimization in the AI system design phase reveals how users reason about the necessity of data based on service quality, specific types of data, or volume and recency of data influenced by their mental capacity and available information. I then investigate how the data creation phase for AI development contributes to privacy risk, misrepresentation, and bias. To address ethical concerns, I developed a method for ethically creating a dataset with an underserved population (blind users) and built the first disability-first dataset, BivPriv to support novel algorithm development for blind users’ visual privacy management (e.g., automatically identifying private or sensitive content in their pictures and videos before they share them with others). The findings highlight users’ willingness to actively engage in the AI development lifecycle and their expectations to be informed about how researchers or companies use the data they share and how their contribution benefits AI systems. Leveraging these insights, I then designed Inclusive.AI, a platform equipped with decentralized governance mechanisms (e.g., quadratic voting), aiming to engage a wide range of people (e.g., underserved groups) in democratic decision-making processes about AI development and governance. This involves introducing technical interventions, particularly Decentralized Autonomous Organizations (DAOs), that incorporate various deliberation, and preference aggregation through voting mechanisms to allow active user participation in influencing how AI should behave, for instance, determining the level of personalization users would experience in AI systems. Through large-scale experiments on different use cases (e.g., text-to-image models, multimodal language models), Inclusive.AI demonstrates promising improvements in scalability, expressiveness, and governance quality in incorporating users’ preferences. This thesis highlights the importance of critically reflecting on when and how to engage end users in the AI development lifecycle. While engaging users is crucial, it is important to build a method for fair deliberation that considers the resilience of the decision-making process, the power structure among stakeholders, and the aggregation method used for gathering user preferences with practical constraints. Key contributions of this thesis include- (i) privacy and security expectations in AI systems (e.g., search engines, targeted ads) across South Asia and EU/UK, accounting for cultural, socioeconomical, and regulatory differences to conceptualize low-fidelity designs based on contextual and situational privacy concerns; (ii) a novel method to ethically create public disability-first dataset, “BivPriv,” to support AI models development in identifying private visual content; (iii) systematically assess the level of decentralization of various DAOs to identify design metrics for democratic decision-making platform; and (iv) design and evaluation of a democratic tool “InclusiveAI,” for decision-making on controversial AI topic (e.g. stereotype, politics), guiding future research in “Democratic AI.”

Graduation Semester

2024-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/125676

Copyright and License Information

Copyright 2024 Tanusree Sharma

Owning Collections