How Is Shadow AI Legitimized in Higher Education? A Qualitative Study Among Academic Educators in the Absence of Institutional Guidance

Synthesis: Academic educators use generative AI for teaching preparation whether or not their institution has authorized it, and this study asks how that informal use is made to seem acceptable rather than whether it happens. Drawing on shadow AI research, professional role identity, and moral disengagement, Rameskumar and colleagues conducted 20 semi-structured interviews with academic educators at German higher education institutions and analyzed the transcripts with a Gioia-style coding process. The findings describe a three-step sequence: an institutional guidance vacuum that opens discretionary space, professional role alignment that sorts AI use into compatible, peripheral, and problematic forms, and moral legitimation that frames the practice as efficient, pedagogically valuable, and still under the educator's control. The authors present the result as a conceptual model of shadow AI legitimation in higher education. The contribution is a reframing of ungoverned AI use as bounded professional practice rather than hidden rule-breaking, and the model explains how existing use is justified, not why educators adopt these tools in the first place.

Key Findings

1. The absence of guidance is the enabling condition, not an accident. Participants described rules that were missing, too general, fragmented, or disconnected from concrete teaching situations. As one educator put it, "No one has told me what is or isn't allowed with these tools, there's simply no official guideline" (E01).
2. Discretion came with anticipated risk. Educators decided independently whether, when, and for which tasks to use AI, but several worried about criticism, reputational damage, or sanctions if the institution later judged the practice differently. "I don't want to be seen by colleagues as someone who lets AI do their work" (E10).
3. Professional role alignment produced three categories of use. Educators judged AI as part of the professional role when it supported material preparation, content organization, and examples; as role-external when confined to peripheral tasks such as wording an email (E11); and as problematic when it threatened academic judgment or pedagogical authority.
4. Preserving pedagogical authority was the boundary condition. Participants insisted on reviewing, adapting, and taking responsibility for generated outputs before they reached students. "The rough draft may come from AI, but the substantive refinement is clearly my own work" (E08).
5. Efficiency carried the most explicit justifications. Time savings and reduced workload on routine preparation were offered as sufficient grounds for use: "What used to take two hours of prep I now do in twenty minutes, for me, that justifies using it" (E02).
6. Value-based justification went beyond speed. Educators argued that AI improved the quality, structure, or variety of teaching materials and made past material reusable, so the practice was defended as better teaching rather than merely faster teaching (E05, E09).
7. Responsibility reframing completed the legitimation. Participants located final judgment and accountability with themselves rather than with the tool: "A tool is neither good nor bad, it all depends on how you use it" (E13). The authors read this as moral legitimation in Bandura's sense, but applied to a practice educators consider professionally reasonable rather than deviant.

From shadow IT to shadow AI

The paper separates shadow AI from its parent concept. Shadow IT covers non-sanctioned tools that substitute for approved systems in information processing and communication, and the risks attach mostly to infrastructure. Shadow AI covers systems that generate content, support decisions, and perform knowledge tasks, so control shifts from managing tool access to regulating tool use and the primary risks shift from infrastructure to knowledge. In higher education the phenomenon has a further twist: the tools are often freely accessible and use is typically unrestricted, so the shadow lies in the absence of sanctioning structures rather than in deliberate circumvention. The authors are explicit that openness of use and absence of guidance are not mutually exclusive, and that their definition does not imply concealment. This matters for institutional AI infrastructure: the practical question is not who holds an account but what counts as acceptable practice inside a course.

Three steps of legitimation

The data structure moved from 118 open codes, 13 second-order themes, and five aggregate dimensions to 18 first-order concepts, nine second-order themes, and three aggregate dimensions. Those three dimensions form a sequence.

The interviews spanned universities and universities of applied sciences, professors, lecturers, and research assistants, and averaged 31 minutes and 24 seconds. Preliminary coding suggested that no new themes appeared after approximately 16 conversations, and the remaining interviews confirmed the stability of the emerging categories.

The institutional guidance vacuum creates discretionary space. Missing or underspecified guidance does not produce unrestricted use; perceived organizational risk constrains it. The second step, professional role alignment, is the interpretive mechanism. Educators ask whether a given use supports teaching responsibilities, sits outside the core pedagogical task, or threatens academic judgment, and this sorting establishes the boundaries within which justification is possible. The third step, moral legitimation, supplies the arguments — efficiency, value, and retained responsibility — that convert institutionally uncertain use into a practice presented as acceptable and professionally appropriate. The authors stress that the model does not explain initial adoption. It specifies how existing informal use is evaluated and bounded, and they note a recursive relation in which legitimized shadow AI reveals where formal rules and shared standards remain underspecified.

What this means for practice

• Administrators. Replace general AI statements with role-specific guidance that names typical teaching use cases — material preparation, assignment creation, pedagogical examples, and examination and assessment handling — and that distinguishes permissible, conditionally permissible, and problematic use with the conditions attached.
• Administrators and governance leads. Build regulation on professional judgment rather than control alone. A compliance-only model misreads a practice that educators already conduct selectively and autonomously, and it forfeits the criteria — pedagogical responsibility, academic judgment, didactic quality, transparency — that participants themselves used.
• Faculty developers. Pair technical training with structured reflection on ethical, pedagogical, and professional questions, and teach review, adaptation, and responsible integration of generated content rather than tool mechanics alone.
• Instructors. Make the boundary work visible. The review and adaptation step participants described is the same step that keeps academic integrity and AI literacy defensible when no policy exists, and it is worth documenting at course level.
• Researchers. Treat justification talk as data about professional norms, not as a proxy for usage rates, and design studies that can compare educator accounts with what institutions and students observe.

Limitations

• The study rests on 20 qualitative interviews with academic educators in Germany, so the mechanisms are well described for that setting but not established for other countries, higher education systems, or institutional contexts.
• The sample is concentrated in information systems and computer science departments, whose representatives tend to view generative AI more positively and know it better than colleagues elsewhere, which likely colors both use and justification.
• Only the educator perspective was collected; students, university administrations, IT departments, data protection officers, and centers for higher education teaching were not systematically included, leaving the question of shared expectations about responsible use open.
• The findings derive from self-reported experiences and justifications, which reconstruct subjective interpretations well but reflect actual usage patterns only indirectly.
• No document analysis, diary study, observation, or quantitative survey was combined with the interviews, so the stability of the identified legitimation patterns across usage contexts remains untested.
• The conceptual model describes how existing informal use is evaluated and legitimized, and deliberately does not explain why academic educators begin using generative AI, which limits its reach as an account of adoption.

Citation

Rameskumar, A., Khosrawi-Rad, B., Korn, F., & Schaarschmidt, M. (2027). How Is Shadow AI Legitimized in Higher Education? A Qualitative Study Among Academic Educators in the Absence of Institutional Guidance. Proceedings of the Hawaii International Conference on System Sciences (HICSS 2027).