Uncovering Students' Mental Models of Generative Artificial Intelligence

Synthesis: Ganguly, Garika, and Johri (2026) elicit 64 undergraduate concept maps of generative AI from a required technology-ethics course and identify five distinct mental-model categories — technical-process based, educational-tool based, transitional, consequence-aware, and integrated. The authors find that students do not hold a unified mental model of GenAI, and that declarative ("what") knowledge dominates while procedural ("how") and conditional ("when and why") knowledge remain sparse — only nine maps integrate all three. Because technical literacy and ethical awareness appear to develop separately, the authors argue that AI Literacy curriculum and GenAI guidelines must actively bridge these domains rather than assume a surface familiarity with tools will translate into responsible, calibrated use.

Mental models as a lens on GenAI use

A mental model is an internal cognitive framework for reasoning and decision-making about a system, constructed from an individual's experiences and perceptions. In educational-technology contexts, mental models shape how learners make sense of new tools, decide whether and how to use them, and develop durable behavioral patterns. The authors note that students already use GenAI (ChatGPT, Claude, GitHub Copilot) for academic work — improving drafts, generating ideas, completing tasks — and that short-term performance gains (e.g. higher essay scores) can mask discrepancies between what students believe GenAI can do and how they actually deploy it. Such mismatches produce unintended policy violations, missed opportunities for productive engagement, or resistance to institutional rules that feel disconnected from how students understand the tool.

Prior research on GenAI adoption has largely measured what students know through tests and surveys, or tracked tool adoption and skill development — approaches that capture surface knowledge but rarely expose the underlying cognitive structures students use to reason about AI. This study instead treats mental models as a learnable, researchable object, connecting to the wiki's treatment of AI understanding as a metacognitive skill.

Study design: concept mapping as an elicitation method

The authors leveraged concept mapping as a primary elicitation technique, treating each map as an "expressed model" that externalizes a student's internal conceptual framework. Concept maps are well suited to probing abstract constructs — technical mechanisms, ethical values — that are difficult to verbalize, and can reveal Misconceptions about AI that structured interviews might miss.

• Sample: 86 undergraduate students in a required technology-ethics course within the IT curriculum at George Mason University. As part of an assignment, students created a concept map representing their understanding of GenAI plus a brief written explanation; no further constraints were imposed so as not to bias their thinking.
• Processing: Of 86 students, 15 did not submit, yielding 71 maps. Two researchers independently scored maps holistically on comprehensiveness, organization, and correctness (a 1–3 rubric with plus/minus gradations, per Besterfield-Sacre et al. 2004), mapped to a 9-point scale and grouped into four levels. Eight maps fell into the lowest (Level 1) category; two were retained as exceptional despite weak structure, so six were excluded — leaving 64 concept maps for analysis.
• Analysis: Two complementary approaches — (1) frequency coding of core components via a 16-code inductive codebook, and (2) hierarchical clustering of how concepts co-occur and connect across maps — surfaced recurring structures in students' understanding (RQ1). A qualitative pass coded each map for declarative, procedural, and conditional knowledge (RQ2).

Five mental models of GenAI

Hierarchical clustering of the concept maps revealed five recurring categories of student mental models:

1. Technical-process based — Students connect machine learning, neural networks, and large language models to capabilities like content creation and Multimodal AI output, and to developers/companies. The underlying technology and its generative capacity are seen as inseparable. These students "know how GenAI works" but may lack the human and societal context.
2. Educational-tool based — Students group personalized learning, administrative support, and Accessibility together, perceiving GenAI primarily as something that lowers barriers, adapts to individual needs, and makes education more inclusive. Notably, this cluster sits at a moderate distance from the technical cluster — students who think about GenAI's educational benefits often do not connect those benefits to its technical foundations.
3. Transitional — Students pair productivity and efficiency with policymakers and government, beginning to move beyond individual tool use toward recognizing GenAI as a technology that needs governance. The small size of this cluster suggests this regulatory awareness is still limited.
4. Consequence-aware — Students associate academic integrity (plagiarism, cheating) with long-term impacts, recognizing that decisions about GenAI use today carry implications far into the future.
5. Integrated — Students group ethics, privacy, human oversight, and academic stakeholders together, treating responsible use as a collective, multi-stakeholder concern rather than an individual one. The presence of human oversight alongside ethics and privacy signals solution-oriented thinking. Cluster 5 merges with Cluster 4, indicating that consequence-awareness and ethical governance are two sides of the same coin in the most developed mental models.

Wide but shallow: the dominance of declarative knowledge

Every concept map reflected declarative knowledge ("what GenAI is"), but far fewer showed procedural knowledge ("how it works", 25 maps) or conditional knowledge ("when and why to use it", 17 maps). The co-occurrence pattern is more revealing than the raw counts:

• Declarative only (31 maps) — the most common profile: rich lists of tools and categories with little process reasoning or evaluative judgment. The authors characterize this as a "wide but shallow" mental model — knowing names and applications without understanding mechanisms or appropriate-use boundaries.
• Declarative + procedural (16 maps) — technically oriented but not critically oriented: students can trace how GenAI works around concepts but show no evaluative or ethical judgment.
• Declarative + conditional (8 maps) — students can name what GenAI is and hold judgments about when/why it should be used, but cannot trace how the technology actually works.
• All three, integrated (9 maps) — the most sophisticated, multilayered mental models.

The authors connect this pattern to the knowledge-structure literature, noting that declarative knowledge is the most readily activated and easily assessed — particularly in novices encountering a domain for the first time. The finding that technical and social-regulatory clusters remain far apart suggests that technical literacy and ethical awareness are developing separately in students, and that education has yet to bridge the two.

What this means for practice

• Instructors. Use the five categories as a diagnostic framework: identify where individual students' maps lie, then design activities that stretch thinking across cluster boundaries toward an integrated model holding technical, educational, ethical, and governance dimensions simultaneously.
• Instructors. Ask for a concept map early in the course and read it against the three knowledge types, because declarative knowledge was universal while procedural and conditional knowledge were not — all 64 maps showed declarative knowledge, 25 showed procedural, 17 showed conditional, and 31 were declarative-only.
• Instructors. Do not let technical literacy stand in for ethical awareness: the technical-process and educational-tool clusters sit at a moderate distance from each other, and only 9 of 64 maps integrated procedural and conditional understanding with declarative knowledge.
• Learners. Audit your own model against the same three questions — what GenAI is, how it works, and when and why to use it — since the most common profile across the 64 maps is the "wide but shallow" one that names tools and applications without mechanism or appropriate-use boundaries.
• Instructors. Widen course and institutional guidelines beyond academic integrity, which matches only one dimension of how students conceptualize GenAI; build guidance in layers, scaffolded from what GenAI is and how it works, through its educational affordances, to its long-term consequences and collective responsibilities.

Limitations

• The study draws on a single required technology-ethics course in the IT curriculum at one university: 86 undergraduates were asked for a concept map as part of an assignment, 15 did not submit, and screening removed 6 more low-rated maps, leaving 64 maps for analysis.
• The five mental-model categories rest on analysis the authors present as preliminary work in progress, and they explicitly leave validation through interviews and think-aloud protocols to future work, so the categories are an interpretation of the maps rather than a confirmed typology.
• Concept maps were self-generated with no constraints so as not to bias student thinking, and were scored holistically by two researchers on comprehensiveness, organization, and correctness using a 1–3 rubric mapped to a 9-point scale — an elicited, rubric-judged artifact rather than a direct measure of reasoning.
• The study cannot connect mental models to behavior: whether students with more integrated mental models make more reflective, responsible use choices is left open, and the small resulting clusters (the transitional category in particular) limit inference from category size.

Citation

Ganguly, A., Garika, S. S., & Johri, A. (2026). Uncovering Students' Mental Models of Generative Artificial Intelligence. arXiv preprint.