Constructing Selves in AI-Mediated Apprenticeship Learning: Students of Colour's Narratives of Agency, Recognition and Negotiation

Synthesis: Gyekye (2026) conducted a 9-month qualitative narrative inquiry with five secondary-school apprentices of color in an online apprenticeship program, examining how they constructed narratives of agency, recognition, and negotiation while engaging with generative AI (ChatGPT). Students authored themselves through AI in three moves: refusal (non-use as alignment with institutional rules or epistemic values), negotiation (repairing institutional mismatches; treating privacy as a "distant abstraction"), and recognition (equating fairness with the absence of visible bias — a "cultural flatness" in which identity-specific experiences go unrepresented in AI responses). The study introduces two interpretive concepts — Institutional AI Literacy (the capacity to make AI use institutionally legible) and Recognition Bandwidth (the room learners have to see themselves recognized in AI outputs) — arguing that effective AI-mediated learning extends beyond technical proficiency to include agency, legitimate authorship, and meaningful recognition. A five-student single-program qualitative study: its contribution is showing that AI use, limitation, and refusal are bound up with how learners understand themselves as authors and emerging professionals.

Summary

Generative AI is rapidly transforming computer-assisted learning, yet research has focused primarily on academic integrity, efficiency, and adoption — with comparatively less attention to students' own experiences, particularly those of students of color in authentic educational settings. This study asked how five secondary-school apprentices of color in an online apprenticeship program constructed narratives of agency, recognition, and negotiation while engaging with ChatGPT, and what those narratives reveal about AI-mediated learning in apprenticeship education.

The study used qualitative narrative inquiry over 9 months, with data from semi-structured interviews, reflective journals, and observations. Analysis followed a systematic process — familiarization, initial coding, pattern search, theme review, theme definition — sensitized by the study's theoretical framework (Rest's Four-Component Model of moral development, recognition theory, and narrative identity theory). Three overarching constructs emerged, organized around the research questions: narratives of agency (RQ1), narratives of recognition (RQ2), and narratives of negotiation (RQ3), with five subsections (non-use, institutional mismatch and privacy, equity and design feedback) identified through narrative analysis.

Key Findings

• Refusal as identity construction, not a failure of adoption. Students constructed their identities around ChatGPT partly by refusing it: non-use was an intentional practice asserting control, aligning with institutional expectations ("I don't use ChatGPT for school reasons since it is not allowed at my school"), or signaling epistemic values ("I don't think ChatGPT is something I would ever use unless I'm told to"). Refusal performed compliance, caution, or depth-oriented learning — agency as the capacity to position oneself within institutional and technological constraints.
• Institutional mismatch as a negotiation site. Students repaired mismatches between what AI produced and what their institutions would accept — reframing AI outputs to make them institutionally legible. This negotiation work is invisible to policy framings that treat AI use as a binary choice.
• Privacy as a distant abstraction. Contrary to the assumption that youth are highly privacy-concerned, most students described privacy as a concept acknowledged in general terms but rarely guiding everyday practice ("I don't think ChatGPT invades your data privacy…unless you are providing it with your personal information") — privacy surfacing occasionally but remaining removed from decision-making.
• Cultural flatness: neutrality read as equity. Nearly all students equated fairness with the absence of visible bias — invisibility interpreted as equitable treatment. But this neutrality often produced "cultural flatness": identity-specific experiences, perspectives, and cultural markers not meaningfully represented in AI responses. When asked about bias, students saw none where the authors saw erasure.
• Students as co-designers of equitable AI. Student redesign proposals positioned them as co-designers: they proposed design features that would better support transparency, learning, and equitable participation — evidence that learners can articulate what recognition would require of AI systems.
• Two interpretive concepts proposed. Institutional AI Literacy — the capacity to navigate AI within institutional expectations and make AI use institutionally legible; Recognition Bandwidth — the room learners have to see themselves recognized in AI outputs. Both are proposed as lenses requiring further empirical examination across settings.

What this means for practice

• Instructors. Treat refusal and non-use as legitimate identity work, not resistance to be overcome: students' reasons for not using AI (institutional alignment, epistemic values, caution) are pedagogically meaningful data about how they position themselves as learners.
• Instructors. Expect students to negotiate AI outputs into institutionally acceptable forms — the mismatch between fluent AI prose and what a program accepts is real work students are already doing, and it is invisible to policies that ask only whether AI was used.
• Curriculum designers. Design for Recognition Bandwidth: ask whether learners can see their identities, cultural contexts, and ways of knowing represented in AI responses — neutrality-as-fairness can itself be the exclusion mechanism, because students will read the absence of visible bias as equity even when their experiences are absent.
• Program leaders. The privacy finding complicates instruction: students hold privacy as an abstraction, not a decision rule, so one-time privacy lectures are unlikely to change practice — privacy reasoning needs to be exercised in situ, tied to concrete AI-use decisions.
• Researchers. The Institutional AI Literacy concept reframes AI literacy as institutionally situated: navigating AI within apprenticeship or workplace expectations is a distinct competence from using AI tools well, and both are distinct from the ethical-awareness framings that dominate AI-literacy instruments.

Limitations

• Five students of color in a single online apprenticeship program: as a qualitative narrative inquiry, the study was designed for in-depth understanding of participants' experiences, not generalizability — the author states the concepts' applicability across settings and populations requires further empirical examination.
• The Institutional AI Literacy and Recognition Bandwidth concepts are proposed as interpretive lenses derived from this dataset, not validated constructs with measurement instruments.
• ChatGPT specifically: the study examines one tool, and findings about cultural flatness and refusal may not transfer to other AI systems with different design and content defaults.
• The author is a practitioner-researcher in the studied program (NYC Public Schools), and the paper does not detail how that position was managed in data collection and analysis beyond general trustworthiness procedures.
• The recruitment frame (apprentices who volunteered to discuss AI) may over-represent students already engaged with the topic.

Citation

Gyekye, E. K. (2026). Constructing selves in AI-mediated apprenticeship learning: Students of colour's narratives of agency, recognition and negotiation. Journal of Computer Assisted Learning, 42(6), e70351. https://doi.org/10.1002/jcal.70351