AI in the Classroom: Ghana’s Student Teachers Reveal What Makes Generated Teaching Visuals Actually Work

Generative artificial intelligence can conjure a classroom illustration in seconds, but a new study from Ghana suggests that speed is the least interesting thing about it. Researchers followed 201 final-year pre-service teachers at Bagabaga College of Education in Tamale as they used a custom-built AI assistant, the Culturally Responsive Visual Aid Generator (CReVAG), during their mandatory teaching practicum. Their findings, published in Discover Education, paint a nuanced picture of what actually happens when future teachers put generative AI to work on real lessons: the technology’s value depends less on how impressive its outputs look and more on how usable the system feels, how confident teachers are in making culturally informed judgements, and how carefully they evaluate the images it produces.

The study is notable for what it did not do. Rather than asking teachers to imagine how they might use AI someday, the researchers embedded a working tool into the messy reality of school-based teacher training. CReVAG was implemented as a Custom GPT within ChatGPT and configured as a visual-design assistant for culturally responsive teaching resources. Before generating any image, the system prompted users to specify contextual information: country, region or district, learner level, subject and topic. Where relevant, users could add language background, community setting, curriculum focus, cultural sensitivities and intended learning outcomes. The tool’s knowledge base contained curriculum materials from Ghana’s Education Service and National Council for Curriculum and Assessment, and only Canvas and Image Generation capabilities were activated for the study.

Participants received a one-week orientation covering contextual prompting, visual generation, refinement and critical evaluation before applying the tool during their twelve-week practicum. The workflow placed the teacher firmly in charge: context specification, visual generation, teacher review, refinement and instructional use. Teachers remained responsible for checking clarity, curriculum relevance, cultural appropriateness and possible stereotypes before using any generated resource in a practicum lesson. Survey data were collected once, after the sixth week of practicum, once participants had gained genuine hands-on experience with the system in their host schools.

The research team measured six constructs: self-efficacy, culturally responsive teaching self-efficacy, perceived system usability, AI tool acceptance, visual resource quality and lesson delivery effectiveness. Using Partial Least Squares Structural Equation Modelling, complemented by Necessary Condition Analysis, they tested twelve hypothesised relationships. The model explained 39.9 percent of the variance in AI tool acceptance, 39.3 percent in visual resource quality and a striking 61.6 percent in lesson delivery effectiveness, with an SRMR of 0.077 indicating acceptable fit. Eight of the twelve hypothesised associations reached statistical significance at the two-tailed 0.05 level.

Two relationships stood out in magnitude. Perceived system usability showed the largest direct association with AI tool acceptance (β = 0.330, p = .004), suggesting that in the pressured environment of a teaching practicum, where students juggle lesson planning, materials preparation and actual classroom teaching, an AI assistant that is cumbersome to operate becomes another burden rather than a help. Meanwhile, visual resource quality showed the largest direct association with lesson delivery effectiveness (β = 0.309, p < .001). In other words, what mattered most for teaching was not enthusiasm for the tool but the perceived clarity, relevance, cultural appropriateness and pedagogical suitability of the images it produced.

The Necessary Condition Analysis added a complementary perspective. While structural equation modelling describes average associations, NCA asks whether high levels of an outcome are simply unattainable when a condition falls below some empirical threshold. Here, visual resource quality emerged as the strongest statistical necessity condition for high lesson delivery effectiveness (d = 0.190, p = .002), followed by AI tool acceptance (d = 0.158). The bottleneck analysis indicated that at approximately 90 percent of the observed lesson delivery range, visual resource quality needed to reach roughly 3.0 on the five-point scale and system usability about 3.2. The authors stress these are sample-specific thresholds, not universal standards or causal cut-offs.

Perhaps the most theoretically intriguing finding is what failed to reach significance. AI tool acceptance was not significantly associated with visual resource quality (β = 0.167, p = .074), meaning that teachers who liked the tool did not necessarily judge its outputs as pedagogically sound. Similarly, culturally responsive teaching self-efficacy predicted evaluations of the generated resources but not acceptance of the system itself. The authors argue this reveals three conceptually distinct evaluative levels in AI-supported teaching: appraisal of the AI system, evaluation of the instructional artefact it produces, and perceived instructional usefulness during lesson delivery. A teacher can embrace an AI platform wholeheartedly while still, quite rightly, rejecting its images as culturally inaccurate or pedagogically weak.

The subgroup analyses were handled with statistical caution. Because programme groups were markedly unequal in size, with 77.6 percent of participants in Junior High School Education, the researchers used Welch’s ANOVA and Games-Howell comparisons where homogeneity assumptions failed. The most defensible programme differences concerned system usability and lesson delivery effectiveness, with JHS participants reporting higher scores than Primary Education participants. Gender differences were generally small, although males reported significantly higher AI tool acceptance (d = 0.40, p = .006) and somewhat higher visual resource quality ratings. The authors explicitly warn against deficit interpretations, noting these within-sample differences do not establish gender as a cause of AI-related beliefs.

The study’s limitations are candidly acknowledged. Because all constructs were measured once, after exposure to CReVAG, the findings represent cross-sectional explanatory associations rather than causal effects. The sample came from a single institution, the data relied on self-reports, and the saved CReVAG configuration did not lock users to a single underlying ChatGPT model, meaning participants may not have experienced completely standardised technological exposure. Prior AI experience, reported by 77.6 percent of participants, was not part of the prespecified model, though sensitivity analyses showed the focal coefficients were broadly stable when it was added as a control. Notably, prior AI experience was positively associated with visual resource quality, hinting that the depth and character of earlier AI use may shape how teachers evaluate generated materials.

The implications reach well beyond Ghana. The authors argue that teacher-education programmes should move beyond one-off AI demonstrations toward sustained practice in prompting, evaluating outputs, correcting inaccuracies and integrating suitable visuals into actual lessons. Culturally responsive prompting and evaluation deserve explicit attention, with clear institutional criteria for accuracy, clarity, curriculum alignment, cultural appropriateness and accessibility. AI-generated materials, they contend, should be treated as drafts requiring professional evaluation rather than ready-made teaching resources. For developers, CReVAG illustrates the value of systems that request contextual information before generating educational visuals instead of relying on generic prompts. And for everyone watching generative AI reshape education, the message is clear: the technology’s real promise lies not in what it can generate, but in the professional judgement teachers bring to judging, refining and using what it makes.

Subject of Research: Pre-service teachers' use of a culturally responsive generative AI tool for creating visual teaching resources in Ghana

Article Title: Pre-service teachers’ interactions with generative AI for creating culturally grounded visual teaching resources

Article References: Nabang, M., Essel, H. B., Tachie-Menson, A., Setsoafia, P., Kodua, C. A., Kubi, J. A., Acheampong, E. B., Aboagye, G. A., & Nyaaba, M. (2026). Pre-service teachers’ interactions with generative AI for creating culturally grounded visual teaching resources. Discover Education, 5(1), Article 1085. https://doi.org/10.1007/s44217-026-02241-y

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02241-y

Keywords: generative AI, pre-service teachers, culturally responsive teaching, teacher education, Ghana, visual teaching resources, AI tool acceptance, system usability, PLS-SEM, Necessary Condition Analysis, lesson delivery, self-efficacy