Students' multimodal prompting practices as epistemic work in AI literacy development

Synthesis: Students' multimodal prompting practices as epistemic work in AI literacy development

Key Findings

• Prompting strategies vary along a continuum from basic input-output use to strategic, iterative, and dialogic practices. Across the eight student groups, the authors mapped four prompting methods — Input-Output, Role-Play, Chain-of-Thought, and Generated Knowledge — with one group (G5) interrogating the system about the components of a good story before generating and another (G3) instructing it step-by-step through the assignment.
• Prompting emerges as a central epistemic practice through which students critically interpret, refine, and negotiate AI-generated outputs — for example, one group's 16-step prompting history moved from exploring genre conventions to challenging the system's character design choices and requesting explanations of its revisions.
• Multimodal engagement exposes challenges in translating abstract meaning into machine-readable prompts, fostering awareness of system limitations and bias: students concluded that "prompt literacy is different between prompting for text than it is for pictures," and that image outputs became increasingly divergent from their intended vision without a precise visual vocabulary ("We realized we don't have as much experience prompting for images").
• Students develop the need to actively construct coherence across modalities when producing text and image outputs, and learn tool-specific constraints — Copilot's image generation was judged less effective than ChatGPT or DALL·E, and groups switched tools after copyright-related refusals, evaluating systems on performance and output quality.
• While students demonstrate developing competence in evaluation and creation, the ethical dimensions of AI Literacy remain underdeveloped: reflections focused on functionality and alignment rather than fairness, transparency, privacy, or responsible use.

Study Design & Method

The study was conducted as an exploratory workshop with 28 postgraduate students engaged in collaborative multimodal prompting tasks, including the creation of short stories or poems and corresponding images using a university-provided GenAI tool. The participants (aged 23–44) came from diverse cultural, linguistic, and academic backgrounds spanning Europe, Asia, Africa, and the Americas, with prior GenAI experience ranging from information retrieval and summarization to image generation and coding tasks. The workshop lasted 1h 45min across two sessions of four groups each; all four groups chose short stories in the first session and the majority chose poems in the second, yielding a balanced exploration of narrative and poetic formats. The corpus comprised eight self-documented group reports (6–22 pages each, 109 pages in total) containing prompting histories, motivations, and reflections, analyzed qualitatively using reflexive thematic analysis, guided by frameworks for prompting methods and AI literacy. The design aimed to provide empirical insight into two questions: which prompting strategies students develop when interacting with open-ended GenAI tools, and how engagement in prompt engineering activities shapes their understanding of GenAI and AI Literacy more broadly.

What this means for practice

• Instructors. Design prompting tasks as iterative, dialogic work rather than one-shot formula practice: students' approaches ranged across Input-Output, Role-Play, Chain-of-Thought, and Generated Knowledge, and refinement came through repeated attempts.
• Build image generation into the task alongside text so students confront the translation problem directly: text prompting yielded satisfactory results with fewer attempts than image prompting, and groups needed a precise visual vocabulary to keep outputs aligned with their intended vision.
• Require students to document and interrogate their prompting histories — the eight groups produced 109 pages of self-documented reports — so the interpretive choices behind each prompt become visible and discussable rather than invisible.
• Teach the ethical dimensions of AI Literacy as explicit curriculum content: students' reflections focused on functionality and alignment rather than fairness, transparency, privacy, or responsible use.
• Learners. Evaluate tools on performance and output quality and be ready to switch: groups judged Copilot's image generation less effective than ChatGPT or DALL·E and changed tools after copyright-related refusals.

Limitations

• The sample of 28 postgraduate students was self-selected and enrolled in a single course module that may have attracted students with a pre-existing interest in technology or digital innovation.
• The intervention was confined to a single workshop lasting 1h and 45 min within one course context, which restricts the scope of the intervention and the depth of longitudinal insight into students' evolving AI literacy.
• Evidence comes from eight self-documented group reports (109 pages in total) analyzed with reflexive thematic analysis, so the ethical-reasoning gap is read from students' own accounts rather than observed practice.
• Participants ranged in age from 23 to 44 and came from diverse cultural, linguistic, and academic backgrounds, so the findings should be treated as exploratory and context-specific.

Citation

Sofkova Hashemi, S. (2026). Students' multimodal prompting practices as epistemic work in AI literacy development.