Research Article
From Tool to Method: Developing AI Literacy in an Upper-Level History Course
Synthesis: Dumančič and Stickle (2026) report a qualitative case study of 15 upper-division history students in a 14-week course on Russian history to 1900 at a regional comprehensive university. The course was built as a three-module arc: an essay written without AI, guided experimentation with AI as a thinking partner, and a final comparative essay where use was optional. Survey and reflection data across the modules show movement from hesitant or uncritical use toward task-specific use — brainstorming, structure and revision — while students kept interpretation for themselves. Trust in the instructor's grading was universal; trust in AI as a grader was not, and fewer than half accepted it as a fair evaluator. About a quarter of the class chose not to use AI for the final essay, which the authors treat as evidence of AI literacy rather than resistance. Their conclusion is that literacy develops through scaffolded, reflective practice rather than exposure to the tool.
Key Findings
- Awareness was near-universal; coursework use was not. Every student had heard of ChatGPT, 11 of 15 had opened it, 9 of 15 had gone beyond curiosity to use it, but only 3 of 15 had used it for coursework before the course — about 20%, a baseline the authors stress because it undercuts the assumption that students arrive as routine users.
- Writing without AI located the actual difficulty. In Module 1, students who struggled named organization and synthesis rather than comprehension: "I knew the content well, but I had a difficult time structuring my points." Twelve of 15 rated the essay valuable or very valuable, the average expected grade was 86.5%, and all 15 strongly agreed that they trusted the instructor to grade fairly.
- Guided use lowered perceived difficulty without inflating confidence. In Module 2, 8 of 15 said the second essay felt slightly easier than the first, and the average expected grade was 85.1%, just over a point below Module 1. Fourteen of 15 rated the assignment valuable, up from 12.
- Students used AI to think, not to generate. The reported uses in Module 2 clustered around brainstorming and ideation (8 or more students), checking and refining ideas (5 or more), organization and structure (4 or more) and finding research directions (3 or more), while 6 or more named its limitations, often through encounters with fabricated content. The authors' summary is that students were using AI to think through essays rather than to produce them.
- Use became more precise over time. By Module 3 comments shifted from "I tried it" to statements of scope: used for transitions but not argument, used to find a second point, ignored when it complicated the work. Self-rated writing improvement rose from 27% of students in Module 1 to 53% in Module 2 and 67% in Module 3.
- Refusal was reasoned, not fearful. Roughly a quarter of the class opted out of AI for the final essay, with explanations such as "I don't find it useful for the kind of thinking this course requires" and "I didn't need it." The authors treat principled nonuse as the outcome the course was designed to produce.
- Grading is where students drew the line. Fewer than half agreed that ChatGPT would be a fair grader, and several disagreed outright; 8 of 15 preferred that the instructor grade alone and 7 were open to shared grading. Their reasoning distinguished what a machine can measure from what a reader can see, including change in a student's work over time.
- The exit survey shows a shift in understanding, with ethical awareness intact. All 15 said their understanding of AI had changed, and 14 of 15 said it changed "a lot" or "extremely"; 14 of 15 called AI valuable in academic contexts; 12 of 15 reported at least moderate ethical concern; 13 of 15 felt moderately to very confident evaluating AI-generated information.
- Students did not expect the humanities to be displaced. Eleven of 15 said AI would probably or definitely not replace human experts, reasoning that it lacks the nuance and subjectivity that historical interpretation requires. Plans for future use were measured: 9 said they might use it, 4 said yes and 2 said no.
- Five design principles summarize the course. Frame AI by modeling your own use and making the thinking visible; scaffold before granting autonomy; build critical reflection into assignments; respect refusal; and keep assessment conversations separate from writing support.
How the study was done
The design is a qualitative case study in the Merriam, Stake and Yin tradition, bounded by a single 14-week semester, a single course and a single regional comprehensive university. The course, Russia to 1900 (HIST 338), enrolls history and social studies majors in their penultimate or final year. Its content — ten centuries of Russian history — was unfamiliar to most students, which the authors describe as producing a level playing field: disciplinary maturity in writing, similar baseline knowledge of the subject.
The three modules ran in sequence. Weeks 1 to 4 excluded AI entirely so that students would establish a personal writing baseline and so the instructor could see each student's writing profile. Weeks 5 to 9 introduced the tool through the instructor's own practice: he shared an AI philosophy document he had co-written with ChatGPT, walked through the prompts and revisions behind it, and then ran a 40-minute exploratory session in week 6 in which students questioned ChatGPT as if it were Catherine the Great. A week later he supplied a bank of prompts for a Reacting to the Past simulation, asking how a Russian noble or a Ukrainian Cossack might argue for gradual serfdom reform. Week 8 added a mini-AI assignment in which each student chose a writing task, submitted the prompts and output, and reflected on whether the exchange produced new ideas. Weeks 10 to 14 removed new AI content and handed decisions to students, ending with a comparative essay on Russian and American empire-building.
Data were surveys with Likert items and open responses, reflections embedded in assignments, the essays themselves, and a final exit survey. Analysis followed Braun and Clarke's thematic analysis: familiarization, inductive coding, theme development, iterative comparison across modules, and deliberate attention to outlier and resistant voices. The instructor was also the researcher, and the paper is candid about the trade-off: insider access to classroom dynamics in exchange for findings filtered through one person's pedagogical values. Participation was voluntary and ungraded, surveys were collected through an external platform, and the instructor kept reflexive memos through the semester. There was no second coder; peer discussion in the final weeks served as informal member checking. Likert data are reported as directional percentages, not as statistical claims.
What this means for practice
- Instructors introducing AI into a writing course. Start with a task written without it. The baseline essay gave students something to compare against and gave the instructor a record of each writer's unaided work; without it, the later claims about AI's effect on structure and confidence have no reference point. Then model the tool rather than describing good practice: show your own prompts, drafts and revisions. Students in this course named the instructor's transparency as the reason they took the tool seriously, and the AI philosophy document did more work than a policy statement would have.
- Course designers. Make the conversation the assignment. Submitting prompts, outputs and a short reflection turned prompting into evidence of metacognitive work and gave the instructor a record of what students actually asked for.
- Assessment designers. Keep grading out of the tool. Students' objections were about judgment rather than accuracy: a model can score features, but it cannot see the difference between a student's usual work and this piece. That is an argument for keeping automated scoring away from interpretive writing, and for saying so explicitly in the syllabus.
- Instructors worried about nonuse. Treat refusal as a legitimate outcome and design the final task so that AI is optional. About a quarter of this class declined and still met the course goals, which is the evidence that the choice was real rather than rhetorical.
- Department chairs and faculty developers. The arc costs time: three sequenced modules, a low-stakes exploration session and a reflective assignment. A one-off workshop cannot reproduce it; the precision students developed depends on repetition across modules.
Limitations
- Fifteen students in one course at one institution. The authors claim analytical rather than statistical generalization, and readers are asked to judge transferability by whether the case resembles their own teaching. The refusal finding is the thinnest cell of all — "about a quarter" of 15 is three or four students — so it is illustrative rather than a rate that should be generalized.
- All outcomes are self-reported. Essays were not externally rated, and the expected grades students gave are estimates, not records; the paper treats their consistency as reassurance rather than evidence.
- The instructor was the researcher, there was no second coder, and member checking was informal. Reflexive memos and transparent coding mitigate but do not remove the risk that the account reflects the instructor's expectations.
- The course conditions may have inflated receptivity: students knew little Russian history, so AI scaffolding addressed a genuine gap, and the institution serves many resource-constrained students. Results may not transfer to introductory surveys, research universities or STEM courses.
- The study does not record which model or version students used, so the account is tied to an unrecorded tool vintage.
- Likert items are reported descriptively, with no validated pre/post instrument and no comparison course, so the developmental arc is an interpretation of self-report rather than a measured change.
Citation
Dumančič, M., & Stickle, T. (2026). From Tool to Method: Developing AI Literacy in an Upper-Level History Course. Journal of Teaching and Learning with Technology, 14(Special Issue), 20–34.