Literacy demands placed on students in student–GenAI entanglements in higher education

Synthesis: Sun, Dohn and Rehm ask not what GenAI can do in higher education but what each form of student–GenAI engagement demands of the student, and answer with a structured rapid review of 102 empirical studies classified against Strobel et al.'s taxonomy of GenAI applications. The distribution is lopsided: Assistant and Enabler entanglements account for 42.16% of the corpus each, while Generator, Reimaginator and Synthesizer together make up less than one fifth. Because the same large language models can support all five, the authors argue the skew is not a technological ceiling but an institutional one — Assistant and Enabler work keeps competence attributable to the individual student, while the marginal forms distribute epistemic work across the student–GenAI relation in ways the institution has not learned to recognize. Reading each type through Dohn's three context levels (domain-internal, activity-internal, activity-framing) shows demands moving from coherence to open conflict, and the paper's response is three design principles for teachers: make the demand transparent, sequence activities across the entanglement spectrum, and build explicit reflection on how epistemic responsibility was distributed.

Key Findings

1. From 5,762 Scopus records screened against pre-specified criteria, 102 empirical studies of GenAI in higher education were retained; a requirement that the abstract identify GenAI performing a specific function in a teaching or learning activity excluded adoption, attitude and general-impact studies.
2. Two authors independently classified the corpus against Strobel et al.'s five types, with inter-rater reliability of 91.09% and disagreements resolved by consensus; each study received one primary category.
3. The distribution is dominated by Assistant (43 studies, 42.16%) and Enabler (43, 42.16%), with Generator (8, 7.84%), Synthesizer (5, 4.90%) and Reimaginator (3, 2.94%) together under one fifth of the corpus.
4. The authors rule out technological limitation as the explanation, since the same models that support Assistant and Enabler applications can also generate, transform and synthesize; they read the skew as a question of which distributions of epistemic work higher education has made legitimate and adoptable.
5. Analyzing each type through Dohn's three context levels shows a gradient from coherence to tension: Assistant is coherent across all three levels, Enabler is coherent but rests on an unexamined blurring of individually possessed competence and "person-plus" competence, Synthesizer and Generator introduce tension where the activity's demands diverge from what the institution recognizes, and Reimaginator produces the most acute conflict because students take epistemic responsibility for outputs they did not produce while still being assessed as individual authors.
6. Three design principles follow: transparency (make the literacy demands of each entanglement type explicit, including where legitimate use ends and overreliance begins), progression (sequence activities from Assistant/Enabler toward Synthesizer, Generator and Reimaginator rather than exposing students only to the most institutionally comfortable forms), and reflection and negotiation (build reflection into GenAI activities so students can name what an activity demanded, where they felt tension, and how they navigated it).
7. The paper's conclusion reframes GenAI literacy itself: not competence in using AI, but competence in navigating what it means to know, act and take responsibility within entangled human–AI practices.

Five forms of student–GenAI entanglement

The taxonomy distinguishes types by how epistemic work is distributed between the student and the system. Assistant describes bounded support on work the student authored — feedback, corrections, suggestions on their own essay. Enabler is reinterpreted from Strobel et al.'s infrastructure category into a competence-Scaffolding role, where the pedagogical purpose is developing higher-order competences (Critical Thinking, Self-Regulated Learning, Creativity, literacy) rather than completing a task. Synthesizer covers artificial but disciplinarily grounded materials — synthetic questions, cases or scenarios for practice. Generator describes the system producing the educational artifact from minimal student input, leaving the student to prompt, curate and evaluate. Reimaginator covers cross-modal or stylistic transformation of existing material, such as converting a health-information script into a personalized video in another language.

The interesting move is not the list but the demand analysis layered on it. Dohn's framework separates domain-internal demands (what the discipline requires), activity-internal demands (how this activity is performed), and activity-framing demands (what the setting recognizes as legitimate). Demands at the three levels can reinforce each other, interfere, or contradict, and the paper's claim is that the marginal entanglement types are marginal because their demands fracture across the levels rather than because they are hard to implement: in a Generator activity the design may be sound while the institution still expects the primary artifact to be student-authored, a contradiction students are left to navigate without being told it exists.

Why the distribution is an institutional pattern, not a capability limit

The paper's explanation is that higher education's practice logic retains an individualistic model of competence: what a student knows and can do should be attributable to that student alone, with tools used to express that competence rather than to share in producing it. Assistant and Enabler entanglements fit that logic — Assistant keeps GenAI at the margins of student-authored work, and Enabler looks coherent because the activity-framing level still credits the developed competence to the student, even though the competences it builds are exercised through continued entanglement, blurring the boundary between individually possessed and "person-plus" competence.

For the marginal types the contradiction becomes visible. Generator engagements become institutionally acceptable mainly where an additional professional framing justifies the distribution — clinical assessment items, nursing care planning, visual arts practice — and stay less visible where no such framing exists. Reimaginator engagements are the sharpest case: the student must judge whether a transformed output remains epistemically faithful to a source they may not fully command, and the institution must decide who authored the result. What is missing from the corpus, on this reading, is not tool capability but legitimacy — a pedagogical and governance question rather than a technical one.

What this means for practice

• Name the demand instead of assuming students infer it. For Assistant and Enabler work, make explicit what individual competence means when GenAI is involved and where the line between legitimate use and overreliance sits. For Synthesizer, Generator and Reimaginator work, say the tension out loud: explain how epistemic work is distributed, what counts as legitimate participation, and how the activity relates to the institution's expectations.
• Sequence across entanglement types rather than teaching one comfortable pattern. Start with Assistant or Enabler activities where demands cohere, then introduce Synthesizer, Generator and Reimaginator work so students develop the confidence to manage more distributed epistemic work — many graduates otherwise meet only the institutionally safest forms before entering workplaces where the others are routine.
• Build reflection on the distribution itself. Ask students to articulate what an activity demanded of them, where they felt tension or uncertainty, and how they navigated it. The target is meta-level: recognizing implicit expectations in a novel entanglement rather than mastering one type.
• Treat literacy as responsibility, not fluency. On this account an AI-literate graduate is one who can judge how epistemic work and responsibility are distributed in a given human–AI arrangement and act appropriately within it — which makes AI Literacy an institutional design problem as much as a student skill.

Limitations

• Each study was assigned a single primary entanglement type, which simplifies genuinely hybrid uses; the authors note that multi-label coding could examine those combinations in future work.
• The literacy demands were inferred theoretically through Dohn's framework rather than validated empirically with students, so the account rests on how GenAI use is described in published studies rather than on students' own reports of what felt demanding or contradictory.
• The corpus reflects what has been studied and published rather than what is enacted in classrooms, so the observed distribution may partly track publication conventions, research tractability and funding priorities rather than institutional accommodation alone.

Citation

Sun, Z., Dohn, N. B., & Rehm, M. (2026). Literacy demands placed on students in student–GenAI entanglements in higher education. Postdigital Science and Education.