Generative AI in Higher Education Instructional Design: A Systematic Review of AI Roles, Educator Competences, and Professional Development Approaches

Synthesis: A PRISMA-guided qualitative systematic review of 78 studies (2022–2026) finds that documented AI-supported instructional design clusters overwhelmingly in the Design phase of ADDIE — 95% of studies — while Implementation (9%) and Evaluation (13%) go largely unexamined, and no study examined AI support across the complete cycle. Most of the corpus (76%) applies no named competence framework, so findings cannot accumulate; human oversight is an explicit commitment in 58% of studies but examined empirically in only 27%. The review's organising claim is that the educator's design expertise, not the technology's generative fluency, remains the centre of the practice — and that professional development should therefore teach evaluation, pedagogical decision-making and accountable oversight rather than tool operation.

Study Design

The review followed PRISMA and synthesised 78 studies published between 2022 and 2026 on educator-facing generative AI use in higher-education instructional design. Rather than pooling effect sizes — which the authors judged inappropriate given a corpus of practitioner reports, small-sample case studies and mixed-methods designs — it conducted a qualitative thematic synthesis organised around four research questions covering the distribution of activity across ADDIE phases and AI roles, the competences required, the operationalisation of human oversight, and the professional-development approaches documented for building them.

The corpus splits into two context strands: 46 practitioner studies of professional design practice and 32 pre-service studies in teacher education. Methodological quality was appraised with MMAT 2018, and that appraisal informed the confidence weighting of findings rather than excluding studies, so weaker designs contribute at correspondingly lower evidential weight. A companion paper earlier mapped 20 of the practitioner studies; the present review reports across the full corpus.

Key Findings

• AI support is structurally concentrated in one phase. Design-phase activity appears in 74 of 78 studies (95%) and Development in 44 (56%), while Analysis reaches only 11 (14%), Evaluation 10 (13%) and Implementation 7 (9%). No study examined AI support across the complete ADDIE cycle, so the cycle is an aspiration that the evidence base does not cover. Implementation, the least documented phase, appears only in practitioner studies.

• Three AI roles, hierarchically distributed. Generative use — AI producing design content at the educator's direction — dominates at 75 of 78 studies (96%). Scaffolding use, where AI structures the educator's own design reasoning, appears in 27 (35%), including a deliberately designed "critical-friend" configuration positioned against the model's tendency to agree. Evaluative use, where AI assesses artefact quality, appears in 16 (21%). The field is overwhelmingly using AI to make things rather than to interrogate its own design decisions.

• Generated learning objectives mostly need repair. Only 38.3% of AI-generated learning objectives met all quality criteria without revision, which makes objective formulation the most rigorously evaluated design decision in the corpus and simultaneously the clearest evidence that generation is not the bottleneck — verification is.

• Most research is theoretically unmoored. 59 of 78 studies (76%) inferred competences from documented practice with no theoretical mapping. Of the 19 that applied a named framework, TPACK and its variants account for 17 — used productively as an analytical lens but stretched beyond its original scope, since TPACK was not designed for AI-specific capacities. AI-specific extensions such as Intelligent-TPACK appear in only five studies. Critically, none of the policy-level frameworks reviewed (DigCompEdu, the UNESCO AI Competency Framework for Teachers) were applied as an analytical framework: research proceeds untheorised while the frameworks designed to guide practice remain empirically untested in this context.

• Seven competence clusters, with ethics the weakest. Thematic synthesis yielded clusters for prompt engineering, critical evaluation of AI outputs, ethical awareness and bias detection, pedagogical–content alignment, conceptual AI literacy, human oversight and validation, and reflective practice and self-regulation. Prompt engineering was the most consistently documented and behaves as a developmental spectrum — from generic instruction-following toward contextually precise, pedagogically grounded prompting — leading the authors to argue it is best understood as pedagogical reasoning conducted through a new medium rather than as a technical skill. Ethical awareness was the weakest documented competence in the pre-service studies, and pre-service teachers systematically overestimate their own evaluation skills, suggesting a perceptual rather than a declarative deficit.

• Oversight is asserted far more than examined. 75 of 78 studies (96%) describe human review of AI-generated content; 45 (58%) treat it as an explicit commitment backed by a concrete practice. Yet only 21 studies (27%) examine oversight empirically. The review consolidates five oversight mechanisms ordered from proceduralised to attitudinal — ethical decision checkpoints, structured rubric evaluation, iterative prompting with review-and-reprompt, collaborative expert coaching, and a professional cautious stance — and adds two emerging types: configurable human-in-the-loop modes and dialogic critical-friend protocols. Where human-involvement configurations were compared systematically, the fullest co-pilot arrangement outperformed autonomous generation.

• Six professional-development approach families, three of which are well evidenced. The families are compact embedded modules (2–4 weeks), semester-long microlearning, task-based design cycles, collaborative dialogue-based question chains, critical co-discovery, and self-directed practitioner learning. Outcome evidence is strongest for compact embedded modules: a three-week module produced statistically significant gains across all six Intelligent-TPACK domains with medium-to-large effect sizes, and a four-week programme with 436 pre-service teachers and a control condition constitutes the strongest controlled evaluation in the corpus. Three design features distinguish the most rigorously evidenced interventions: ethics embedded as distributed practice rather than a stand-alone session, performance-based competence assessment (lesson plans, decision logs, micro-teaching) motivated directly by the unreliability of self-report, and prompt engineering developed as reflective, pedagogically grounded practice.

• Pre-service education is the laboratory, not the proof. Design-phase activity is universal in the pre-service strand (32/32), and Scaffolding appears in 16 of 32 (50%) there against roughly a quarter of practitioner studies (11/46), because those settings are pedagogically designed interventions. Oversight is also structurally layered, with the teacher educator supervising the pre-service teacher's supervision of the AI. The authors are careful about what this means: responsible AI-supported design has been documented predominantly where practice is institutionally supervised, and the absence of documented examples in the practitioner corpus does not demonstrate their absence from professional practice.

What this means for practice

• Train for evaluation and judgement, not tool operation. Generated learning objectives fail quality criteria 62% of the time unrevised, and pre-service teachers systematically overestimate their own evaluation skills — so professional development should centre output evaluation, pedagogical decision-making and accountable oversight, which is the review's own conclusion.
• Assess competence through performance. Use lesson plans, decision logs and micro-teaching rather than self-report, because the unreliability of self-assessment is demonstrated in the same literature the training addresses.
• Embed ethics as distributed practice with checkpoints at every design task. Alongside performance-based assessment and reflective prompt work, this is one of the three features distinguishing the best-evidenced interventions from the weakest.
• Prefer the fuller co-pilot configuration over autonomous generation for design work — where human-involvement configurations were compared systematically, the fullest arrangement outperformed.
• Do not assume the cycle is covered. 95% of documented activity sits in Design and 9% in Implementation, so treat claims that AI supports full course development as unsupported, and ask for documented oversight mechanisms rather than oversight commitments (58% assert a practice, 27% examine one).

Limitations

The authors identify several limits, and they constrain how far the findings travel.

• Language and retrieval bias. The English-language search under-represents research published in other languages, and 22 reports could not be retrieved despite institutional access. The 2026 literature is necessarily incomplete as indexing continues.
• Single-coder risk on interpretive categories. Screening and coding were performed by the first author with borderline cases resolved by consensus, and coding reliability was not quantified through independent double-coding — so the AI roles and oversight levels in particular carry a risk of single-coder bias, mitigated but not eliminated.
• Self-report dominance. Self-report methodologies dominate the corpus, which matters acutely here because the review itself demonstrates the divergence between self-assessed and externally assessed competence. The caution applies to the included studies and to the review's own competence findings.
• Heterogeneity precludes quantitative synthesis. The corpus mixes practitioner reports, small-sample case studies and mixed-methods designs, so no pooled effect size is available and no study examined the complete ADDIE cycle.
• Tool findings date quickly. The rapid evolution of AI tools means findings about specific tool limitations may not reflect current capabilities. ChatGPT appears in 50 of 78 studies (64%), but the landscape is diversifying toward multi-model comparisons.

Citation

Spyropoulou, N. D., & Kameas, A. D. (2026). Generative AI in Higher Education Instructional Design: A Systematic Review of AI Roles, Educator Competences, and Professional Development Approaches. Computers, 15(11), 745.