Designing Meaningful Algorithmic System Transparency for Non-Expert Users

Transparency matters: people need it to make informed decisions regarding technologies that might impact them. Among others, non-expert users represent key stakeholders that can benefit from but also be impacted by algorithmic systems. Their transparency needs deserve urgent attention as daily lives increasingly rely on such systems. In this thesis, I explore how to design meaningful algorithmic system transparency for non-expert users. I highlight three core barriers to transparency, among other challenges raised in the literature. First, it is unclear how to design transparency in practice (‘Design Challenge’). Transparency is considered a core principle of Responsible AI, but not a panacea. To serve users, I thus argue transparency must be meaningful to them, i.e. contextually appropriate. Secondly, new laws and transparency guidelines are emerging. I argue these must align with designers’ and users’ perspectives to become effective (‘Guideline Challenge’). Lastly, there remains the need to assess transparency (‘Assessment Challenge’). By directly engaging with users and designers of AI systems, I provide empirical and critical insights to help address these three challenges. I thus contribute towards bridging the gap between theory and design practices, so that algorithmic systems better cater for users’ information needs in real-world settings. This is key to help ensure that systems can be reviewed, challenged, and used safely by users. Without meaningful transparency, it would not be possible to understand, critique, and scrutinise them. To explore the Design Challenge, I first investigate what meaningful transparency means for under-studied users, in particular parents in the UK in the wake of the new online safety legislation (Chapter 3). To help solve the Guideline Challenge, I study the gap between a recent international standard on algorithmic system transparency and designers’ practices (Chapter 4). I also use a co-design approach to test and expand design recommendations from the human-centred explainable AI (HCXAI) literature in a specific clinical domain: gynaecology (Chapter 5). Lastly, I explore the Assessment Challenge by benchmarking meaningful transparency (Chapter 3) and experimenting with a critical method to investigate the transparency dimensions of a system’s user interface (Chapter 6). This thesis thus aims to provide guidance for designers, standards bodies, academics, and policy-makers on what meaningful transparency can mean for various non-expert users, and how to promote it in practice.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Schor, Bianca

Advisors dc:contributor.advisor

• Blackwell, Alan
• Singh, Jatinder

• Transparency
• Non-Expert Users
• Algorithmic System
• Artificial Intelligence
• Design

Licence

• https://www.repository.cam.ac.uk/bitstreams/0b38ee6e-b1d1-4026-bdeb-2414780e5ac6/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.113618

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/376338