The primary aim of this research is to develop diverse and fair datasets, with a specific focus on marginalized demographics like the Global South and/or non-binary gender groups. I am working towards creating two types of datasets-- (a) Real and Synthetic face datasets from Global South for auditing/training Face Recognition Systems. These are used in tasks like face detection and facial attribute analysis, and (b) Datasets with non-binary gender labels, across modalities like image and text for more inclusive retrieval (visual search enabled e-commerce) and classification (text-based gender prediction) tasks.
Software systems increasingly mediate critical societal functions involving large-scale use of data. The use of personal and sensitive data introduces ethical and legal concerns, necessitating architectures that support Responsible AI and enforce open access safeguards grounded in ethical principles. These principles are conventionally derived through laws, regulations, codes of conduct, and frameworks. Much of the focus in responsible AI has been on privacy alone. However, there can be other sensitive information in the enterprise and public domain that needs to be handled in a responsible and ethical manner as well. Ethical compliance traditionally has been achieved through systematic manual adherence to guidelines, regulations, and other documentation. However, there is a lack of concrete software architectures, frameworks, or tools that can provide automation to enable compliance directly in software systems. This paper introduces Guardrail Framework, a modular, reusable software framework that operationalizes responsible AI principles by decoupling ethical constraints from the functional requirements of software applications. The proposed novel framework addresses the issue of ethical compliance in application software by creating a robust software framework that can be plugged into multiple application software development environments. The framework allows ethical filtering of information on the basis of data sensitivity (Low to High), the trust score of the user (Low to High), and granularity of the data (Cell, Row, Column, or Table). The proposed Guardrail framework is implemented and evaluated using Open Government Dataset (OGD), demonstrating high cohesion, low coupling, and long-term maintainability, in line with fundamental software engineering architectural design principles. Notably, the proposed framework is equally effective for ensuring ethical use of personal, enterprise, and public data.
MIMIC (Medical Information Mart for Intensive Care) is one of the largest, most commonly-used, freely available datasets containing intensive care unit data. I conduct denotative, connotative, and deconstructive readings of the MIMIC-IV dataset through an analysis of the data sources, dataset structure, and the process for getting access to the data, as well as documents and concepts related to the dataset. As a result, I demonstrate that the MIMIC-IV dataset requires more documentation, including an expansion of the existing descriptions, in order to ensure the data is used appropriately and allow for maximum benefit. I make recommendations for future users of the MIMIC-IV dataset, creators of datasets in general, and researchers in the Critical Data Studies field based on my findings.
This paper examines the dilemmas of moral autonomy in the context of the transfer of human consciousness to AI, focusing on the transhumanist archetype Homo Evolutus. Through comparative analysis, the study explores two distinct paradigms: the secular approach, rooted in Kantian philosophy, which considers that moral autonomy can be preserved through transfer; and the religious perspective (Catholic, Protestant, and Orthodox), which argues that the human person is an essential unity of body and soul, making the transfer conceptually impossible. The study highlights three central dilemmas: the authenticity of digital moral reflection, the distribution of responsibility between the biological and digital selves, and the limits of autonomy in human-AI integration. The secular paradigm perpetuates Cartesian dualism, while the theological perspective asserts that consciousness transfer raises fundamental ontological problems that cannot be solved by secular approaches alone. The conclusion emphasizes the need for an interdisciplinary approach that includes theological aspects to clarify what it means to be a moral agent in a technologically altered world.
This thesis examines how AI systems deployed in Malaysia’s smart cities encode constitutional and racial asymmetries through algorithmic infrastructures. Focusing on epistemic governance, it analyses how data, models, and legal frameworks shape machine legibility and citizen classification, with comparative insights from India and Brazil to contextualise Malaysia’s postcolonial governance model.
Representation is a critical concern in the development, use and decisions about AI. AI fairness and health equity research present frameworks for representation that highlight underlying structural issues (such as systemic discrimination) that go beyond, yet also shape, both data and AI tooling. Still, there is a lack of actionable methods to interrogate data representation and support critical reflection on its role in AI systems. We propose a visual, data-driven approach that links decisions about data representation to AI model performance across subgroups. Our method consists of two plots: the representation association plot, which shows whether subgroup representation affects performance, and the representation expansion plot, which simulates how performance disparities may change when expanding subgroup data. We apply our approach to the Lifelines Cohort Study for two health equity use cases: early detection of diabetes and cardiovascular disease. The plots reveal that improving age representation may reduce disparities, while sex and education-based disparities appear unrelated to representation in the dataset. This approach has the potential to guide researchers in identifying when improving data representation may contribute to reducing performance disparities, and when such efforts are unlikely to be effective. As a critical but partial tool, our approach should be embedded in broader inclusive research practices, where representation extends to who defines the data and determines whether and how AI is applied. Future research should validate the actionability of insights with users and priorities of those bearing the health burden.
We introduce the AI Occupational Capability Index (AI-OCI), a novel methodology for quantifying the alignment between AI model capabilities and the tasks that define human occupations. Unlike prior automation risk metrics, which rely on expert heuristics or job-level generalizations, AI-OCI operates at the task level by embedding and comparing over 19,000 occupational tasks with 338 AI capabilities using state-of-the-art language models. The resulting scores reveal how well AI systems can perform specific human functions, enabling interpretable, task-aligned assessments of labor exposure. Empirical evaluations show strong correlations with benchmark indices such as AIOE and GPT-4 Beta exposure scores, while diverging from legacy automation risk measures. We demonstrate AI-OCI’s utility through case-based analyses of employment and wage shifts across high-alignment occupations during the era of large language model adoption. The framework supports scalable, real-time tracking of AI’s workforce impact and provides a foundation for integrating labor intelligence into education, policy, and economic planning.
The mimesis of human traits exhibited by large language models (LLMs) has led some users to perceive these technical systems as agentic, capable of achieving reciprocal and seemingly human-like communication. These misperceptions have, in turn, been linked to documented harms in human-AI interactions (HAIs). This conceptual paper explores current interventions in response to interaction harms, taking AI companions as an illustrative example. We analyze documented cases of AI companion applications that have led to severe harms, including suicide, illustrating that current redressive approaches fail to account for the network of distributed human agents that collectively "animate" anthropomorphic features and encourage some users to regard AI systems as social "agents." By framing anthropomorphism as a social affordance that reproduces a broader distributed process spanning development, design, user interaction, socio-cultural contexts, and institutional forces, this paper demonstrates the necessity for distributed governance of anthropomorphic AI features across these diverse agentic forces. We proceed to discuss obstacles to appropriate governance, including power asymmetries between different agents, and outline existing models that could be adapted for more effective interventions.
The widespread adoption of large language models (LLMs) and generative AI (GenAI) tools across diverse real-world applications has amplified the importance of addressing societal biases inherent within these technologies. While the Natural Language Processing (NLP) community has extensively studied LLM bias, research investigating how non-expert users perceive and interact with biases from these systems remains limited. As these technologies become increasingly prevalent, understanding this question is crucial to inform model developers in their efforts to mitigate bias. To address this gap, this paper presents findings from a university-level competition that challenged participants to design prompts specifically for eliciting biased outputs from GenAI tools. We conducted a quantitative and qualitative analysis of the submitted prompts and the resulting GenAI outputs. This analysis led to the identification of reproducible biases across eight distinct categories within GenAI systems. Furthermore, we identified and categorized the various strategies employed by participants to successfully induce these biased responses. Our findings provide unique insights into how non-expert users understand, engage with, and attempt to manipulate biases in GenAI tools. This research contributes to a deeper understanding of the user-side experience of AI bias and offers actionable knowledge for developers and policymakers working towards creating fairer and more equitable AI systems.
Recent years have seen a surge of interest in structured tools and interventions for responsible data and technology development and deployment. As part of this phenomenon, data ethics frameworks have emerged as a basic tool that go beyond lists of `ethics principles', aiming to guide the practical application of these ethical considerations. This study offers an in-depth insight into how organisations are making use of the Data Ethics Canvas - a popular, openly licensed data ethics framework, developed by a non-profit, and used across a diverse range of organisations. Through a reflexive thematic analysis of 22 interviews with tool users, we draw key reflections to understand its role in organisational settings. We find that participants aim to use the tool for data ethics assessment, literacy, and as a foundation or inspiration for governance initiatives. We argue that the versatility of the tool offers an opportunity for `brokering' among stakeholders who approach data practices from different perspectives. Still, limited guidance and a need to overcome organisational barriers leave significant responsibility on individuals to deliver outcomes from interactions with the tool. We further discuss how far these findings can be used to reflect on wider use, re-design, re-framing and evaluation of interventions to generate responsible data practices. We offer recommendations including the need for clear and robust guidance of users' engagements with tools, and for tools to account for a landscape where responsibilities for data practices and their AI applications increasingly converge.
Large language models (LLMs) were not designed to replace healthcare workers, but they are being used in ways that can lead users to overestimate the types of roles that these systems can assume. While prompt engineering has been shown to improve LLMs' clinical effectiveness in mental health applications, little is known about whether such strategies help models adhere to ethical principles for real-world deployment. In this study, we conducted an 18-month ethnographic collaboration with mental health practitioners (three clinically licensed psychologists and seven trained peer counselors) to map LLM counselors' behavior during a session to professional codes of conduct established by organizations like the American Psychological Association (APA). Through qualitative analysis and expert evaluation of N=137 sessions (110 self-counseling; 27 simulated), we outline a framework of 15 ethical violations mapped to 5 major themes. These include: Lack of Contextual Understanding, where the counselor fails to account for users' lived experiences, leading to oversimplified, contextually irrelevant, and one-size-fits-all intervention; Poor Therapeutic Collaboration, where the counselor's low turn-taking behavior and invalidating outputs limit users' agency over their therapeutic experience; Deceptive Empathy, where the counselor's simulated anthropomorphic responses (``I hear you'', ``I understand'') create a false sense of emotional connection; Unfair Discrimination, where the counselor's responses exhibit algorithmic bias and cultural insensitivity toward marginalized populations; and Lack of Safety & Crisis Management, where individuals who are ``knowledgeable enough'' to correct LLM outputs are at an advantage, while others, due to lack of clinical knowledge and digital literacy, are more likely to suffer from clinically inappropriate responses. Reflecting on these findings through a practitioner-informed lens, we argue that reducing psychotherapy—a deeply meaningful and relational process—to a language generation task can have serious and harmful implications in practice. We conclude by discussing policy-oriented accountability mechanisms for emerging LLM counselors.
Text-to-image (T2I) AI tools are trained on vast datasets of existing images and artworks. We identify that existing ethical standards and regulatory safeguards for these tools largely lie within the Western neoliberal realm. They assume that artistic creativity originates from individuals rather than in collectives or social environments, ownership is an individual concern rather than shaped by communities and shared cultural traditions, and compensation should be based on individual claims rather than acknowledging collective contributions to artistic knowledge. In this paper, we counter these assumptions by theorizing ‘collective agency’ as a critical conceptual lens to rethink artists’ community-centric roles in relation to these tools. Drawing from our nine-month-long qualitative interventions with diverse Bangladeshi artist groups, we find that these artists manifest cultural resonance, co-creation, and sense of recognition through their art-making practices which fosters collective agency among them. This empirically grounded account of collective agency in our study posits practical design and policy implications, such as incorporating artists’ solidarity, community-centric data stewardship, and collective bargaining mechanisms in ethical development of T2I AI tools to reclaim artists' control over their creative practices in the AI age.
The relationship between AI and uncertainty in high-stakes public environments has not yet been given the attention that it requires. While technical literature often frames uncertainty as a limitation that should be resolved or minimised, this project draws attention to an alternative interpretation: uncertainty as a fundamental and valuable component of human judgment, particularly within many aspects of public sector decision-making, and therefore minimising uncertainty to design more effective AI can become undesirable. My research investigates how AI systems designed for predictability, consistency, and optimization struggle to operate effectively in environments where discretion, ambiguity, and pluralism are not only unavoidable but often necessary. This project advances the conceptual understanding of uncertainty in AI ethics and governance while also offering early empirical insights through experiments with large language models in legal interpretive tasks. The overarching aim is to develop normative and technical guidance for building AI systems that align more meaningfully with the social and institutional functions of uncertainty. Additionally, I acknowledge the benefits of meaningfully minimising environmental uncertainties for AI systems and my future work aspires to produce a framework to help guide when adaptations to reduce uncertainty for public sector AI are permittable and when they should not be made to ensure the inherent humanness of society remains intact.
This research study explores how Canadian multinationals can leverage Environmental, Social and Governance (ESG) and artificial intelligence tools to combat foreign bribery and corruption in international business transactions.
This study investigates how AI image generation plat-forms develop distinct platform vernaculars systematic visual languages with recognizable aesthetic signatures that create new forms of cultural bias. Through mixed-methods analysis combining content analysis of 306 AI-generated images, interviews with 16 visual professionals, and survey validation from 430 respondents, we document three key phenomena. First, platform choice ac-counts for 54% of variance in visual outcomes (η² = 0.542, p < 0.001), with professionals achieving distinct recognition accuracy: Flux (82%), Midjourney (78%), Stable Diffusion (71%). Second, rather than displacement, 71.9% of professionals report AI enhances creative capabilities through strategic integration models. Third, systematic cultural bias manifests through 73% Western demographic defaults in neutral prompts, though professionals develop sophisticated mitigation strategies achieving 82% bias reduction effectiveness. These findings establish platform vernaculars as algorithmic aesthetic hierarchies that require new forms of visual literacy, challenging assumptions about AI homogenization while revealing persistent representational inequities demanding professional resistance strategies.
This research investigates how cooperatives and marginalized communities can lead in governing AI systems by drawing on traditions of economic democracy and collective accountability. Through case studies in Detroit’s food justice movement and the Mondragón Corporation, I explore how these groups anticipate the social impacts of AI and co-design governance models rooted in their values. Moving beyond risk-based approaches, this work proposes a framework for anticipatory governance that redistributes power in technology development and centers community-led innovation.
This work explores how AI-based systems can be designed and developed in a more responsible manner by aligning them with human values. To this end, we propose a structured approach for identifying and specifying values for AI systems through stakeholder workshops. While engaging relevant stakeholders within a particular context of use is crucial for designing more ethically-aligned AI systems, it is not a trivial task due to the inherent complexity and abstractness values can evoke. In this paper, we aim to address these challenges and offer actionable insights by developing methodologies for value identification and specification, grounded in a real-world use case: a healthcare application designed to detect and prevent exacerbations of chronic obstructive pulmonary disease (COPD). We demonstrate how values such as autonomy can be identified, contextualized and translated into concrete design implications guiding the AI system's behaviour. Our contribution offers practical guidance for fellow researchers, designers, developers and regulators seeking to navigate the complexities of responsible AI development.
This paper describes a methodology for Evolving Systems Supporting Governance, Regulation, Control, Safety, and Security in personalised healthcare. To embrace AI in any critical system, any stochastic advantage of time or resource saving needs to be trusted, and this in turn needs deterministic consolidation – i.e. verification processes which both secure the foundation of any novel system through offering reassurances of the reliance upon models and also validation of those models in practice, since they may “drift“ over time. Deviating from this original foundation can lead to errors, which need to be addressed for such a system to remain useful, indeed credible. This review of how such governance can become integral to developing new principles for responsible AI inspired by personalised healthcare's 5Ps which can be adopted for the managing of a common yet critical care path. This leads to many questions around the strategic, operational and tactical approaches, which are answered through providing a use case of dealing with Emergency to exemplify future approaches to agent based healthcare management.
As governments around the world race to build “sovereign AI” systems, the term has emerged as a rhetorically powerful yet conceptually ambiguous expression of national ambition. While it evokes autonomy over artificial intelligence infrastructure, in practice it often masks deep dependencies on transnational firms—particularly U.S.-based companies like NVIDIA. This study critically examines how sovereignty is strategically sold and mobilized in global discourse, tracing the term’s diffusion through news media and analyzing how state and corporate actors invoke the language of autonomy while remaining embedded in U.S.-dominated infrastructural hierarchies. Drawing on critical discourse analysis and infrastructure studies, the paper shows how sovereignty operates not only as a policy goal but also as a symbolic strategy that reframes material reliance as empowerment. The analysis is based on a comprehensive dataset of all news articles containing the phrase “sovereign AI” in the Nexis Uni database as of April 19, 2025, providing a robust foundation for analyzing the discourse’s global diffusion and strategic use. Theoretically, the study contributes to debates on sociotechnical imaginaries, infrastructural asymmetry, and technonationalism by demonstrating how AI sovereignty narratives perform geopolitical work: legitimizing public investment, obscuring structural dependencies, and rebranding global inequality as innovation. In doing so, it urges scholars and policymakers to interrogate the political work that discourse performs in shaping AI futures—and to ask who benefits when sovereignty is claimed, and who bears the cost when it is not realized.
As organisations rapidly adopt artificial intelligence (AI), they encounter a host of complex ethical challenges. These challenges are made even more difficult by the nature of AI itself: it is not a static or uniform technology but rather an evolving ecosystem of tools and systems that interact dynamically with their environments. Managers are under increasing pressure to address these challenges and justify governance investments, while established frameworks for responsible AI governance are still emerging. The central question of thew paper, then, is how managers make sense of these evolving ethical challenges. One productive way to explore this - that we also adopt in this paper - is drawing parallels with the field of Corporate Social Responsibility (CSR). Like AI governance, CSR also involves responding to ambiguous expectations and CSR research provides valuable insights into how organizations engage in "sensemaking" (a process of interpreting and giving meaning to complex situations in order to guide decisions and justify actions). In CSR literature, two dominant forms of sensemaking are identified: "value-driven" and "instrumental". The value-driven approach is rooted in ethical principles and moral commitments; the instrumental approach treats governance as a strategic tool, aligning ethical practices with business goals. Drawing from this conceptual framework, the paper presents a study of managers from diverse sectors to explore how they make sense of AI ethics and governance. The analysis reveals that AI governance sensemaking similarly falls into these two categories, value-driven and instrumental, but importantly, neither approach is sufficient on its own. Instead, our qualitative study finds that managers should blend these two approaches into what we describe as a "holistic sensemaking strategy". By adopting a holistic framework that integrates both perspectives, managers are better equipped to navigate the evolving and ambiguous terrain of AI governance, thus making investments in governance that are both principled and practical.
