The current wave of text-to-image generators deployed in the field of education are based on the principles of diffusion models, commonly combined with transformer architectures implementing a denoising mechanism, with previous GAN approaches relegated to the role of supporting mechanisms ([19]; [47]). In turn, the main consumer-grade systems-DALL·E, Midjourney, and Stable Diffusion-are built according to these principles ([2]; [30]). These systems increasingly incorporate advanced multi-modal user interfaces enabling not only text-to-image generation but also conditions of image-to-image, in-painting capabilities, and structural prompts implemented via ControlNet, which allows generating images constrained with the help of a student’s line drawing, edge maps, depth maps, or pose references ([56]). This aspect has significant pedagogical implications as it means that there is an overlap between the processes of observational drawing and generative AI, and thus, the framing of the phenomenon of AI in education limited to prompt-only systems misses a critical point.

From the perspective of systematic literature reviews, the fields of creative production, scaffolded instruction, and content generation emerge as key areas of applications for generative AI in art education from 2019 to 2025 ([32]; [48]). According to empirical studies, the introduction of AI generated images in conjunction with teacher-mediated fine-tuning leads to greater student engagement, increased self-efficacy, and heightened levels of satisfaction ([10]). Empirical research in the context of training in graphical and abstract arts indicates that students need to possess certain visual literacy and design skills to transform the results produced by generative AI into meaningful art pieces ([12]; [27]; [31]). The question of whether prompt engineering will remain an important independent skill seems debatable as multimodal interfaces increasingly support not only text but also conversation and images; thus, the key unit of student work shifts from prompts to workflow.

According to empirical studies on AI-assisted painting courses, students’ intent to continue the experience is driven by the perceived usefulness and ease of use of these technologies, combined with teacher mediation, which validates classic TAM predictions in this creative domain ([37]; [51]). Systematic reviews of AI in child education highlight concerns about the potential for cognitive standardization inherent in these AI-driven interfaces ([53]). Comparative studies that focus on different models (Midjourney, Stable Diffusion, DALL·E) applied to design education reveal the impact of model choice on the resulting cognitive and creative profile of student work ([2]; [15]).

The literature on observational drawing pedagogy long predates generative AI and is essential context for any AI-integrated model. Edwards’s widely used pedagogy reframes drawing as a perceptual rather than a mechanical task, organized around five perceptual skills: edges, spaces, relationships, lights and shadows, and the gestalt of the whole ([17]). Empirical studies of drawing-from-observation report measurable effects on visual attention, emotional regulation, and self-perception among adolescents ([54]). Foundational studio sequences typically progress from blind-contour and gesture drawing, through tonal value studies, to colour mixing and material handling in media such as charcoal, ink, watercolour, and oil ([28]). Recent transfer-effect studies in museum settings demonstrate that even short, structured drawing programmes can produce measurable socio-emotional gains ([34]). Studies of digital drawing software show that affordances such as layer control and unlimited undo can support specific artistic skills but do not automatically substitute for material judgement built in analogue practice ([28]). Any AI integration must engage seriously with this pedagogical inheritance rather than treat AI as a self-sufficient teacher of drawing.

Three theoretical frames recur in the literature. First, constructivist and socio-cultural learning theory, especially Vygotsky’s zone of proximal development and scaffolding, frames AI tools as potential more-knowledgeable others that can support learners between current and potential performance ([39]). Second, the Technological Pedagogical Content Knowledge (TPACK) framework specifies the intertwined knowledge bases that teachers need to integrate technology thoughtfully into subject-specific pedagogy ([35]; [41]). Third, the Technology Acceptance Model (TAM) ([13]) and its educational extensions remain widely used to explain learners’ intention to use AI tools, including in design and painting studios ([20]; [51]).

Alongside these frames, AI literacy frameworks have crystallized around [38] competencies for recognizing, evaluating, and ethically using AI ([38]), extended by global guidance from UNESCO and the OECD that emphasize human-centred, ethical, and inclusive deployment ([40]; [42]). Recent competency syntheses confirm that AI literacy must combine technical, evaluative, practical, and ethical dimensions across school and higher-education learners ([11]).

Legal and ethical considerations related to the use of AI-generated art include issues of authorship, the origin of training data, and the replacement of human creativity ([33]; [50]). Legal-philosophical analysis further examines the possibility of copyright protection for AI-generated content, arguing that generative AI outputs lack the characteristics of human authorship, making the use of such models in students’ artworks an issue for grading, citation, and presentation. In terms of art education, the discussion prompts educators to ask several important questions, namely about the origin of training data, the potential impact of “in the style of” on living artists, and the information that must be disclosed to audiences.

The scholarly literature suggests that AI-generated art could be considered a form of collaboration or cocreation, in which the role of the person involved shifts from that of a single creator to that of a curator, prompter, and concept designer ([49]; [57]). Studies on human-AI cocreation show that although AI may demonstrate equivalent or even superior performance to humans on creativity measures, domain experts still identify shortcomings in the originality and contextuality of AI-generated works ([1]; [21]). Art educators have called for an integrative approach to using AI in the classroom that combines classical teaching methods with AI reflection ([24]; [23]).

PRISMA 2020 defines a clear framework for systematic reviews through a checklist of 27 items and modified flow diagrams ([44]). Design Science Research (DSR), as defined in the field of information systems by Hevner et al. and Peffers et al., offers a complementary approach in which the researcher continuously designs, develops, and evaluates artifacts to resolve real-world problems ([26]; [25]; [46]). The combination of a PRISMA-guided systematic review with DSR-based artifact construction and multi-method validation is well established in adjacent technical domains, most notably digital forensics, where Al-Dhaqm and colleagues developed and face-validated process models and meta-models spanning common investigation processes, their categorization, and cloud-based data collection ([3]; [8]). To the best of our knowledge, however, this combined approach has not previously been applied to drawing-and-painting pedagogy. The present study therefore transfers an established methodological template from a technical domain to a humanistic, studio-based one, pairing a systematic review with artifact design and multi-method validation.

Following PRISMA 2020 guidance ([44]), a systematic search was conducted across Web of Science, Scopus, ScienceDirect, IEEE Xplore, ACM Digital Library, ERIC, and Google Scholar. Search terms combined art-education vocabulary (“art education”, “drawing”, “painting”, “visual arts curriculum”, “studio practice”), AI vocabulary (“artificial intelligence”, “generative AI”, “text-to-image”, “DALL·E”, “Midjourney”, “Stable Diffusion”), and pedagogy vocabulary (“pedagogy”, “curriculum”, “teaching model”, “AI literacy”). The search window covered January 2019 to February 2026.

Table 1.Database search results across PRISMA phases.

Note. The synthesis was narrative-thematic, organized around uses, benefits, limitations, and pedagogical theories. The 79 included studies form the basis for thematic synthesis and the model design.

The inclusion criteria included: 1) peer-reviewed empirical studies, conceptual studies, or systematic reviews regarding the role of artificial intelligence (AI) in visual art education; 2) studies written in English; and 3) studies concerning drawing, painting, design, or related forms of visual art. The exclusion criteria included: studies examining only music or text generation; opinion studies without any empirical or conceptual basis; studies on AI in non-art forms of STEM where there is no potential applicability for visual arts. Screening was done first based on titles and abstracts, followed by full-text screening. The PRISMA flow diagram elements of identification, screening, eligibility, and inclusion are shown in Table 1 and Figure 1. The numbers provided in Table 1 are due to the structured searching process and the number of articles reviewed in model development. This study did not register its protocol with PROSPERO and hence no inter-rater reliability metrics are provided.

Figure 1.PRISMA 2020 flow diagram showing identification, screening, eligibility, and inclusion stages of the literature search.

The DSR phase followed the six-activity Peffers et al. process model ([46]) within the three cycles articulated by Hevner ([25]). The relevance cycle drew problem requirements from the contemporary classroom—the need to teach foundational drawing and painting in a context where AI can produce polished images on demand. The rigor cycle anchored the design in constructivist learning theory, TPACK, AI literacy frameworks, observational-drawing pedagogy, and prior case-study work ([17]; [38]; [39]; [41]; [54]). The central design cycle iteratively developed the AADPP model through three internal iterations: an initial conceptual draft, refinement following literature synthesis, and refinement following expert review. The structure is depicted in Figure 2.

The six DSR activities were operationalized as follows:

Problem identification.Fragmented integration of AI in drawing and painting curricula; lack of a validated unifying model.

Objectives of a solution.To produce a model that integrates AI tools with foundational craft, ethical reflection, and AI literacy, applicable across foundation-level studio courses.

Figure 2.Design Science Research (DSR) comprises three cycles and six activities in the construction of the AADPP model.

Design and development.Construction of the AADPP model through three iterations, informed by PRISMA synthesis and prior models.

Demonstration.Application of the model to a foundation drawing and painting studio scenario over a 12-week unit.

Evaluation.Face validation by domain experts, comparative analysis with existing studies, and case-study observation.

Communication.The present paper communicates the artifact, its derivation, and its evaluation, following Peffers et al.’s template for design-science publications ([46]).

Three techniques have been utilized for evaluation:

1) Face validity based on criteria for completeness, applicability, relevance, clarity, generality, and currency by a panel of domain experts;

2) Comparative analysis with existing models and research works;

3) Illustration through a hypothetical case study.

Given that comprehensive empirical testing cannot be performed at the stage of the initial design-science research effort, these three evaluation techniques provide grounds for assessing the theoretical consistency and applicability of the proposed model, which remains to be tested empirically in the study described in Section 9.2.

The narrative-thematic synthesis of the included studies produced four recurring design requirements (DRs) that map directly onto the four layers of the AADPP model. DR1 (that foundational observation, mark-making, and material judgement are placed at risk when AI automates the early, formative stages of image production) is addressed by Layer 1, which protects analogue craft as a precondition for AI use ([17]; [54]; [53]). DR2 (that generative AI is most valuable for divergent ideation and exploration when the learner’s own drawing remains in the loop through image-to-image and structural conditioning) is met by Layer 2 ([30]; [56]). DR3 (that responsible integration requires explicit attention to AI literacy, authorship, training-data provenance, bias, and equity of access) is met by Layer 3 ([38]; [40]; [50]). DR4 (that assessment must reward process and craft rather than polished surfaces alone) is met by Layer 4 through a process journal and a craft-weighted rubric ([9]; [23]). Making this mapping explicit clarifies how each component of the artifact is grounded in the reviewed literature rather than asserted independently of it.

Layer 1 protects and develops the embodied core of drawing and painting: observation, gesture, proportion, value, mark-making, edge control, colour mixing, surface, and material judgement. The layer draws on long-standing observational-drawing pedagogy that frames drawing not as the mechanical production of marks but as a perceptual discipline organized around edges, spaces, relationships, lights, and shadows, and the gestalt of the whole ([17]). Activities are deliberately analog and include blind-contour and gesture drawing, sight-size cast and figure studies, tonal value studies in charcoal, plein-air painting, colour-temperature exercises, alla prima still-life, and underdrawings and grisaille methods for oil. Materials include charcoal, graphite, ink, watercolour, acrylic, and oil; acrylic is recommended as a starting medium in introductory courses where drying times for oil or studio infrastructure may be limited.

No AI tools are used in this layer. This restriction is based on experimental evidence regarding the effects of drawing from observation on visual perception, perceptual discrimination, and self-perception ([34]; [54]). Moreover, digital technology—especially generative AI—cannot replace the judgement of form obtained through analogue drawing. Automating the initial confrontation negates the very essence of learning that takes place through the cognitive and physical development that occurs when one draws. Such an approach supports Vygotsky’s claim that the zone of proximal development needs to be actively developed by the learner, not bypassed ([39]).

In this stage, AI image technologies serve as sketchpad helpers for ideation, mood boards, thumbnailing compositions, and quick experimentation with style and sources ([29]; [32]). Two modes of operation can be outlined. First, the text-to-image mode, where students create, refine, and critique their prompts as design documents that show their ideas. Second, the image-to-image and structural conditioning (such as ControlNet) ([56]) mode, where students feed their own sketches, line art, or pose references to the AI model to explore variation based on their own observation. The latter mode has greater educational value because the initial image remains within the process. In both modes, AI results are not perceived as end products but as source material to be redrawn and repainted into classical media ([29], [29]).

Layer 3 examines the social, ethical, and legal issues around AI in art, specifically around where training data comes from, copyright and attribution, the contentious legality of “in the style of” generations, bias and homogenization, and appreciation for human creativity ([33]; [49]; [50]). Style imitation is of particular interest as a pedagogically relevant issue, as instructions to use a living artist’s name can reproduce identifiable aspects of their work without permission or attribution, a practice currently under litigation in cases like Andersen v. Stability AI ([18]). Cases like this will be analyzed by students in terms of policy and legislation and debated on when references to an artist’s style can count as fair inspiration and when they become inappropriate appropriations, robbing an artist of their livelihood.

Equity and access issues are addressed in Layer 3. Choosing a specific AI tool is an ethical decision with different costs and implications depending on whether the tool is commercially available with a subscription model, freely available online but commercial, or is a free and open-source software solution such as Stable Diffusion running on one’s own server ([40]; [42]). Student options are discussed here, with an understanding of the consequences of each alternative. The layer meets the AI literacies outlined by Long and Magerko ([38]) and UNESCO’s ethical and human-centered guidelines ([40]) as well as current competencies ([11]). Students keep a process journal that tracks the choices made, including prompts and workflows used, AI-generated results referenced, accept/reject decisions, and the student’s contribution. AI usage needs to be declared for every assessed piece, and its extent is clearly stated in the unit’s assessment policy.

Layer 4 integrates work from the other three through studio critique, portfolio review, and process-oriented assessment. Building on the Consensual Assessment Technique and adapted Torrance dimensions of fluency, flexibility, originality, and elaboration ([9]; [52]), assessment evaluates both product and process. Rubrics give weight to observational accuracy, material handling, conceptual development, AI literacy demonstrated through the process journal, and ethical transparency. The layer is teacher-mediated and peer-supported, consistent with constructivist studio traditions and TPACK-informed integration ([35]; [41]). The concrete rubric used to operationalize this layer is presented in Section 6.3.

Three principles cut across all four layers. First, human primacy: the human remains the author and decision-maker; AI is a tool ([40]; [49]). Second, transparency: all AI use is disclosed and documented ([33]). Third, iteration: the model is not linear; students move between layers across a project and across a program. Figure 4 maps the model’s theoretical foundations onto its expected learning outcomes.

Figure 5 grounds the model in actual studio outputs: a graphite gesture drawing, a charcoal tonal study, an AI thumbnail ideation grid with one variant selected, a finished hand-painted harbour scene, and a process journal page documenting AI use and ethical reflection.

Figure 4.Theoretical foundations mapped to AADPP layers and expected learning outcomes.

Figure 5.The four AADPP layers in practice.

This scenario assumes a first-year studio class made up of about 24 students taking an introductory drawing and painting curriculum. This unit runs for 12 weeks and includes two three-hour studio sessions per week. Weeks 1 - 4 will run solely in Layer 1, where students learn observational drawing, sight-size and tonal value rendering, basic colour theory, and materials introduction. Acrylics are recommended for initial painting assignments, and oil painting will be introduced after that. During this period, the use of any AI tools is strictly forbidden to ensure the purity of the foundational skills being developed.

Layer 2 and 3 will start in Weeks 5 - 8. At this point, the students receive an introduction to prompt engineering and structuring/conditioning techniques (e.g., inserting their own contour drawings into image-to-image or ControlNet procedures) and then use the AI-assisted tools to generate ideation thumbnails for a self-conceived painting assignment. In this stage, AI outputs are considered part of the sketchbook, requiring the students to redraw critical elements by hand before creating a painting. Layer 3 will take place during weekly seminars on training data sources, copyright and style copying, bias issues, and recent legislation (e.g., Andersen v. Stability AI). Supplementing the lectures will be assigned readings from the AI literacy and regulation literature ([18]; [33]; [38]; [40]; [50]).

Weeks 9 - 12 include Layer 4. During this final phase, students will finish a painting assignment with a process journal listing all prompts and conditioning steps used, consulted AI outputs, acceptance/rejection outcomes, and reflections on craft and ethical considerations. Critiques will occur in groups, utilizing the rubric described in Section 6.3. The full week-by-week schedule for the unit is summarized in Table 4.

Table 4. Illustrative 12-week schedule applying the AADPP model in a foundation drawing and painting unit.

Figure 6 presents a case study in terms of time spent on one project from the prompt creation to the production of the final artwork, showing how the timing of the artwork is more determined by its drawing and painting components than by the involvement of AI. In particular, the redraw step connecting the AI thumbnail sketch and the actual canvas plays an essential role, since it forces the artist to reproduce the AI picture based on personal observation.

Figure 6.From prompt to painting: the AADPP workflow.

For Layer 2, an example of prompts and workflow for one project, “misty harbour at dawn”, can be seen in Table 5 with varying degrees of AI engagement. This process is part of a specific teaching strategy where the students begin with prompts for broad texts, then move to structuring via drawing, and finally use AI-driven variations based on life drawing. The aim is not to replace life drawing but to support it.

Table 5.Example prompts and workflows across three levels of AI involvement (project: “misty harbour at dawn”).

Table 6 summarizes the rubric utilized in Layer 4. All criteria are evaluated based on a scale of four levels. It is noteworthy that the rubric places emphasis on craft fundamentals, along with process and ethical considerations, such that the enhancement of computer-generated products alone cannot guarantee a top score.

Table 6.AADPP holistic assessment rubric (Layer 4).

Based on patterns reported in comparable studies ([10]; [29]), the scenario anticipates several outcomes. Students who complete Layer 1 before encountering AI are expected to use AI more critically and to reject more AI outputs as too generic or homogenized, consistent with concerns raised in the literature ([16]; [49]; [53]). Process journals are expected to show iterative growth in prompt and workflow design ([30]). The assessment rubric, by giving explicit weight to process and ethics, is intended to shift student attention away from polished AI surfaces toward demonstrable understanding ([23]). These are anticipated patterns to be tested empirically in future work (Section 9.2).

The educator’s role across the unit aligns with TPACK ([35]; [41]) and UNESCO AI competency expectations for teachers ([40]). In Layer 1, the educator acts as a traditional studio mentor; in Layer 2, the educator scaffolds prompt and workflow design and counters over-reliance; in Layer 3, the educator hosts critical dialogue and brings policy and legal contexts into the room; in Layer 4, the educator assesses both product and process and models reflective practice.

Studies focused on prompt engineering and iterative AI ideation within digital art classrooms ([29]) inform Layer 2 but do not constitute a curriculum on their own. The AADPP model retains prompt and workflow engineering as central to Layer 2 but explicitly protects Layer 1 craft practice as a precondition. This addresses a recurring limitation of prompt-only approaches: students may achieve sophisticated prompts without developing observational or material competence.

[38] AI literacy competencies and subsequent extensions ([11]; [38]) specify what learners should know about AI, but are not domain-specific to art. The AADPP model operationalizes AI literacy within the drawing and painting context by tying competencies to concrete studio activities and the assessment rubric in Table 6.

Recent systematic reviews ([32]; [36]; [48]) map the landscape of generative AI in art education but do not propose a unifying pedagogical model. The AADPP model contributes the missing artifact by synthesizing review findings into a deployable structure.

Studies of AI painting tools for children raise concerns about cognitive homogenization, sampling bias, and theoretical fragility ([53]). The AADPP model addresses these by making homogenization an explicit topic of critical reflection in Layer 3 and by requiring divergent ideation in Layer 2. The model can be adapted to younger learners by scaling complexity and tightening teacher mediation.

The design-science and metamodel-validation work in digital forensics provides a methodological template for the present study ([7]; [8]). The validation pattern—face validation, comparative analysis, and case-study demonstration—is directly transferable to pedagogical artifacts and is adopted here.

Several limitations apply to this study. First, the evaluation relies on face validation, comparative analysis, and an illustrative case-study scenario rather than a controlled empirical experiment; the case study does not report participant data, and “anticipated outcomes” are derived from comparable studies rather than from new measurement.

Second, the literature on generative AI in art education is expanding rapidly. A model built on 2019-2026 literature must be revisited as new tools and pedagogical practices emerge. Specific features (named tools, specific prompt-engineering patterns) will date faster than the four-layer structure, which is intended to remain stable.

Third, the model assumes baseline access to AI tools, hardware, and trained instructors. In resource-constrained contexts these assumptions may not hold, and equity-of-access considerations require continuing work ([40]).

Fourth, the model is grounded in foundation-level studio practice in higher education. Adaptations to K–12, community-arts, and self-directed learner contexts will require further design iterations and evaluation, particularly with respect to child safety, developmental appropriateness, and parental involvement ([53]).

Fifth, while the methodological lineage of PRISMA + DSR with multi-method validation is well established ([7]; [44]; [46]), its transfer from technical domains to a humanistic, embodied practice such as drawing and painting introduces translation risks. Future work should examine whether art-specific validation criteria—for example, criteria centred on aesthetic development or studio-culture fit—should be added to the standard set used here.

Sixth, the PRISMA review described here was not protocol-registered (for example, with PROSPERO), and inter-rater reliability statistics are not reported. Future systematic reviews extending this work should preregister their protocols and explicitly report screening reliability.

An empirical study to test the AADPP model would adopt a mixed-methods, multi-site design across at least three art schools or university foundation programmes, with cohorts of approximately 25 students each over one or two semesters. A quasi-experimental comparison would contrast AADPP-aligned units with parallel units using either traditional pedagogy or unstructured AI integration. Pre- and post-tests would assess observational drawing skill (using rubric-based ratings of standard subjects), AI literacy (using established instruments), and ethical reasoning (using vignette-based assessments). Process journals and final portfolios would be analyzed qualitatively for evidence of iterative growth in prompt and workflow design and for the quality of accept/reject decisions. Educator focus groups would assess feasibility and TPACK readiness.

Outcome measures would include: 1) measurable change in observational drawing skill, 2) AI literacy scores, 3) ethical reasoning quality, 4) creative outcome ratings using the Consensual Assessment Technique ([9]), and 5) student self-efficacy and continuance intention measures consistent with the TAM tradition in this domain ([51]). The study would be preregistered and conducted with appropriate ethical clearance.