This practice-based PhD develops a socio-technical data literacy framework for generative AI, emphasizing participatory engagement with the socio-cultural dimensions of data. Through methods such as participatory workshops and critical making, the research demonstrates how non-technical stakeholders can decode and intervene in algorithmic bias through engagement with classification data practices. The resulting toolkit and evaluative framework ofer practical strategies for inclusive, culturally-aware participation in educational and civic contexts. By conceptualizing data literacy as a critical, situated and collective critical practice, the research contributes to HCI by advancing more equitable and relational human-AI interaction.
Generative AI technologies are increasingly central to the infrastructures that mediate perception,
representation, and decision-making; however, these systems also replicate and intensify existing social
hierarchies through historically embedded biases in training data and classification architectures [
The term socio-technical underscores that these systems are not purely computational but are shaped by cultural assumptions, institutional structures, and power relations. Data refers not only to training corpora but also to the classification schemas that organize and give meaning to information. These classifications—often hidden behind polished outputs—play a crucial role in determining what is made visible, normative, or excluded. Literacy, in this context, is not merely a technical skill but a relational, critical ability to interpret, question, and reshape algorithmic representations in a situated context. It enables users to surface bias, negotiate meaning, and collectively reimagine the epistemic structures embedded in generative systems.
Rather than treating users as passive recipients of generative technologies, this work emphasizes their role as epistemic agents with ethical responsibility in shaping model behavior through interactions such as prompt design, content selection, and sense-making. Within this perspective, the research investigates how non-technical stakeholders can actively challenge dominant representations and co-construct alternative classificatory logics within generative AI systems.
The work is guided by two central questions: How can socio-technical data literacy function as a creative-critical method for decoding and intervening in bias with generative AI systems, while supporting user agency in the interpretation and manipulation of classification processes? How can participatory approaches to data and AI move beyond technical performativity to foster more relational, care-oriented engagements with generative technologies and data?
Informed by feminist epistemologies [
Situated within HCI, this research advances a set of tools designed to evaluate and inform nontechnical users’ engagement with generative AI systems, with a focus on fostering agency and critical awareness. As these systems become increasingly embedded in everyday contexts, understanding how users interpret, challenge, and influence generative outputs becomes essential.
Designed for educational and civic settings, these tools support reflective engagement and promote individual and collective responsibility. By foregrounding user agency in shaping model behavior, the framework invites critical attention to accountability, transparency, and the classificatory systems that underpin algorithmic outputs. It thus expands the scope of data literacy toward more equitable, situated, and ethically responsive human–AI relations.
This research is currently at the beginning of its second year, with the methodology defined and under implementation. Ethics approval has been obtained from the RMIT University Human Research Ethics Committee in Melbourne, and expert interviews and participatory workshops are currently underway. The following sections outline the theoretical foundations, participatory methodology, and practical contributions of this research.
The ontological foundation of this practice-based PhD engages with interdisciplinary discourses from
philosophy, history, and science and technology studies to examine how technological systems, bias,
and discrimination are co-produced. Understanding AI as a material and discursive infrastructure—
shaped by practices of classification, representation, and decision-making—requires a historically and
conceptually grounded perspective [
Recent researches demonstrate that both large language models (LLMs) and text-to-image (TTI) systems systematically reproduce and amplify stereotypes related to gender, sexuality, and ethnicity [12, 13, 14], marginalize non-Western epistemologies [15], and constrain identity representations [16], while ultimately excluding disability and neurodivergent identities [17]. These patterns of exclusion are deeply embedded in the construction of training datasets, where selection, annotation, and classification practices are shaped by assumptions about what and who should be made visible. Such processes confer epistemic legitimacy on particular worldviews, embedding them within algorithmic systems under the guise of technical objectivity.
Data annotation has been shown to be a situated and power-laden process that mediates subjectivity
and institutional authority [18, 19]. Visual taxonomies, such as those found in ImageNet, have been
shown to rely on normative assumptions about what identities, roles, and expressions should look like,
reducing individuals to predefined labels that claim universality but reflect narrow, culturally specific
worldviews [
As argued by Bowker and Star, classification systems embody institutional and cultural logics that, although often obscured, carry significant epistemic and political consequences [ 20]. In this sense, datasets do not merely reflect reality, but actively construct it through the worldviews embedded in their structures and labeling practices [21]. The act of selecting, labeling, and categorizing images is not neutral or technical—it is a political intervention with lasting impact on how people are seen, sorted, and acted upon by AI systems. These classificatory regimes not only reproduce harm but have become increasingly opaque as commercial AI systems scale, limiting public scrutiny of how representations are produced and deployed.
This research contributes to ongoing debates by framing data literacy as a critical, collective practice specific to the context of generative AI. It addresses the epistemic and political dimensions of classification systems, proposing tools and methods that make these structures accessible and open to contestation by non-technical audiences.
A central ambition of this research is to involve non-technical experts, community members, educators, activists, and other stakeholders who have been historically marginalized or excluded from the design and classification processes underpinning AI systems.
These individuals are not merely end-users; they are co-constructors of knowledge [22, 23] whose lived experiences, values, and perspectives are essential for surfacing biases, contesting dominant narratives, and envisioning alternative futures for AI [24].
Conventional participatory practices in AI often confine stakeholder involvement to consultative or tokenistic roles, where input is solicited only at discrete stages—typically after key design decisions have already been made—or restricted to superficial aspects such as user interface feedback [ 25]. Moreover, current practice frequently relies on proxies—such as UX professionals or algorithmic models—to represent stakeholder voices, rather than enabling direct and sustained engagement [26].
This results in constrained agency and limited influence over foundational classificatory structures. Such forms of engagement are not merely desirable but constitute necessary conditions for ensuring accountability, transparency, and contextual relevance in AI systems. Within this framework, the redistribution of epistemic authority and the collective shaping of classification processes are understood as central to advancing more just, reflexive, and socially responsive technological futures [ 27, 28, 26, 29].
This requires politically informed understandings of how technology and citizenship are entangled, making visible the power relations embedded in digital systems and supporting emancipatory practices aimed at social justice[30]. As D’Ignazio and Klein emphasize in their framework of data feminism, Expanding who participates in the design and interpretation of data is not simply a matter of broadening access, but a deliberate epistemic choice—one that challenges dominant knowledge systems and afirms the value of situated, plural forms of understanding that are often excluded from mainstream data practices [31].
This PhD research adopts a practice-based methodology that weaves together critical theoretical inquiry and participatory experimentation. The approach is structured around four interconnected methodological pillars, each designed to foreground the epistemic and ethical complexities of generative AI while centering the agency of non-technical stakeholders.
The first pillar consists of a literature review conceived as a form of critical infrastructure mapping. Rather than summarizing existing work, it delineates the conceptual landscape of data classification, representational bias, and user agency within generative AI. These systems are approached as sociotechnical infrastructures shaped by historical, cultural, and political conditions.
As part of this phase, a cross-disciplinary bias cartography is assembled—drawing from media studies, data science, informatics, and psychology—not to isolate technological failures, but to examine how algorithmic and human biases intersect. This mapping supports a relational understanding of classification processes and provides a shared foundation for reflection and discussion with experts and participants in following methods.
The second pillar involves semi-structured interviews with designers, educators, AI practitioners, and activists engaged in critical work on algorithmic systems. Conceived as dialogical rather than extractive, these interviews support knowledge co-production through reflective prompts and in-depth discussion.
Participants are invited to critically engage with and contribute to the cross-disciplinary bias cartography developed during the literature review, bringing insights from their respective domains of practice. This process surfaces tensions between ethical commitments and technical constraints, and informs the iterative co-design of workshop formats and the development of evaluation criteria.
The third methodological pillar centers on participatory workshops, which are conceived as epistemic interventions and critical making spaces. These workshops invite participants to engage with generative AI through a combination of reflective inquiry and speculative experimentation.
Activities include prompt hacking—where participants iteratively test and annotate generative models to reveal hidden biases and zine-making, which draws on feminist traditions of storytelling to document personal encounters with algorithmic classification. Further, dataset remixing enables the co-creation of alternative taxonomies through collaborative data curation, while image annotation and reclassification exercises encourage participants to challenge normative visual grammars and disrupt established hierarchies.
These workshops are guided by principles of constructionism [32], care ethics [33], and pedagogical
co-production [
The final phase of the research involves the development and assessment of a modular toolkit designed to support the practice of socio-technical data literacy in educational and civic contexts. The toolkit includes adaptable facilitation formats, design probes, and guiding materials, and is examined through an evaluative lens that attends to multiple dimensions of participant engagement.
These include epistemic understanding—the ability to articulate and identify bias in generative AI systems; social engagement—reflected in the willingness to discuss and intervene in classification processes; and critical awareness—the recognition of AI systems as situated, value-laden infrastructures.
Evaluation is not treated as a conclusive step, but rather as an iterative and generative moment within the research, one that reflects on how theoretical commitments are translated into situated practice, and how collective inquiry can inform more inclusive and contextually grounded ways of engaging with AI.
This practice-based doctoral research contributes to HCI by reframing AI literacy through a sociotechnical lens that centers interpretive agency, ethical responsibility, and participatory engagement. Generative AI is approached not as a neutral infrastructure but as a site where bias is embedded, enacted, and contested through interaction.
The development of a socio-technical data literacy framework enables non-technical stakeholders— often excluded from processes of design, governance, and interpretation—to critically engage with the classificatory logics shaping generative outputs. The project investigates how engagement with generative AI can support a shift from bias awareness to the cultivation of ethical agency. It focuses on how individuals recognize and respond to algorithmic representations, and how moments of interpretation can become sites of negotiation, reflexivity, and shared accountability. Bias is understood not only as a property of data but as something reproduced through interaction, sense-making, and institutional context.
Through participatory workshops, dialogical interviews, and critical making practices, the research explores how collective, situated interventions can foster more inclusive and pluralistic approaches to knowledge production. The resulting toolkit and evaluative framework ofer practical resources for educational and civic contexts, while proposing new ways of assessing agency, awareness, and engagement in AI-mediated environments. Ultimately, this work expands the scope of data literacy and participatory design toward more equitable, reflexive, and socially responsive human–AI interactions.
The author would like to thank their supervisors, Brad Haylock and Laurene Vaughan, from the RMIT Design School for their unwavering support and for guiding the development of this research.
During the preparation of this work, the author used Grammarly in order to: Grammar and spelling check. Further, the author used Deeply in order to: Syntax review.