Research Article
A critical review and actionable framework for integrating generative AI into teacher professional development
No primary data: a critical review synthesizing 20 published studies, with no sample, instrument or fieldwork of its own.
Synthesis: Shenoy and Saarela (2026) synthesize 20 peer-reviewed studies — 16 empirical and 4 conceptual, published 2023–2025 — on generative AI in teacher professional development, proposing the GenAI-TPD framework: five competency domains (Technical Fluency, Pedagogical Integration, Critical Evaluation, Ethical Reasoning, Collaborative Practice) across four stages (Exploration, Integration, Critical Reflection, Peer Exchange), operationalized through six testable propositions and observable indicators. Functional, cognitive, social and reflective affordances recur, but deeper critical reflection is least observed and needs explicit scaffolding; constructivist principles appear in approximately two-thirds of studies, fewer than one-third engage TPACK, and probabilistic, confidently wrong outputs make technological knowledge evaluative rather than accumulative.
Key Findings
- Four affordance categories organize what GenAI offers teachers: functional (content co-creation, instructional design, assessment development), cognitive (higher-order thinking, knowledge construction), social (human–AI partnership, collaboration) and reflective (belief revision, self-efficacy).
- Cognitive gains stop short of critical reflection, the least observed engagement level, so programs must scaffold it explicitly.
- Self-efficacy findings are near-universal and should be discounted. Latent class analysis finds optimistic, critical, critically reflected and neutral profiles, and only the optimistic one is uniformly positive.
- TPACK extensions retain an assumption GenAI breaks — stable, deterministic tools under teacher control — whereas outputs are probabilistic, variable and capable of confident error.
- Only a minority taught prompt engineering explicitly, about one-third guided critical evaluation, and most relied on self-report rather than performance-based assessment.
- Over-reliance is designable: require independent judgment before consulting AI, teach critical evaluation explicitly, and add structured peer observation and mentorship.
- The framework turns description into testable theory through six falsifiable propositions and observable indicators such as prompt complexity scores, seeded-error detection rates and lesson-plan rubrics.
The corpus and how it was synthesized
The search ran from January to March 2025 across Scopus, Web of Science, IEEE Xplore, ACM and ERIC, covering 2020–2025; all 20 included studies appeared between 2023 and 2025. PRISMA flow runs from 1,098 records to 523 after deduplication, 237 assessed at full text and 20 included. The corpus concentrates in East Asia (11 studies) and is dominated by ChatGPT. Two researchers coded all 20 studies across 37 binary categories: mean Cohen's κ = 0.69 (SD = 0.30), mean agreement 86%, with 17 observational categories at κ = 1.00 and interpretive categories lagging (constructivist κ = 0.10, adult learning κ = 0.15). Adapted CASP appraisal scored 7.5–10.0 (mean = 9.3), with 17 of 20 at 9.0 or above; lower scores tracked limited reflexivity reporting.
Affordances and the reflection gap
Teachers used GenAI first for content co-creation, lesson planning, objective formulation, rubric development and formative feedback. They then reported critical evaluation and synthesis of AI-generated content and described the tool as a thought partner, with human–AI partnership framed as complementary rather than replacement; teachers paired with mentors outperformed those working with AI alone. Reflective gains included belief revision and confidence: a brief prompt-engineering intervention produced large effects on AI literacy (Cohen's d = 0.91) and medium anxiety reductions (d = 0.56), while an extended program left perceptions of output originality unchanged. Every study measured self-reported outcomes, which is atypical for technology integration research and likely reflects social desirability as much as competence.
Why TPACK does not fit generative AI
Only a minority engaged TPACK explicitly; most studies that omitted it organized teacher knowledge around constructivist process, discrete competency or affordance function, and none reproduces TPACK's intersection of content, pedagogy and technology. Three misalignments follow: TPACK assumes static, deterministic tools while GenAI behaves dynamically; it treats technology as an instrument under human control while GenAI acts as a collaborative agent; and it treats technological knowledge as bounded while generative systems produce emergent affordances. Because outputs vary across identical prompts, knowledge of the tool cannot accumulate; the needed competence is interrogating output.
The GenAI-TPD framework
Technical Fluency covers prompt design, navigation and troubleshooting; Pedagogical Integration aligns GenAI with curricular goals and subject matter; Critical Evaluation covers fact-checking, bias detection, reflective critique and architectural literacy; Ethical Reasoning covers integrity, fairness, transparency and data handling; Collaborative Practice covers co-design, peer feedback and student-centered use. Stages run Exploration (about 6–10 hours), Integration (6–12 weeks), Critical Reflection and continuous Peer Exchange, with iteration built in. Six propositions specify validation designs: proficiency predicts sophisticated prompting; critical-evaluation training improves seeded-error detection; ethical reasoning predicts classroom AI-use policies; peer observation supports six-month retention; peer exchange amplifies other domains; combined GenAI-plus-mentorship programs outperform either alone. The indicators remain operationalized constructs, not validated instruments.
What this means for practice
- Instructors. Require an independent attempt before GenAI access, and teach prompt engineering and output verification explicitly.
- Instructional designers. Sequence Exploration, Integration, Critical Reflection and Peer Exchange, using microlearning with a community of practice and mentoring.
- Educational technology developers. Build verification in — seeded-error tasks, bias-detection prompts, provenance cues — and expose performance measures rather than usage dashboards.
- Administrators. Align rollout with capacity building, clear policies and dedicated development time that preserves autonomy; fund equitable AI-literacy access.
- Researchers. Test P1–P6 with performance outcomes, prioritize P2 and P6, validate the Table 2 instruments, and run cross-context studies.
Limitations
- English-language peer-reviewed restriction and East Asian concentration (11 of 20) limit transfer; subject-specific professional development is underrepresented.
- Self-report dominance invites social desirability bias; few studies used performance-based assessment, and absent negative outcomes are atypical.
- Pre-service teachers are overrepresented relative to in-service teachers.
- 2023–2025 findings may not capture later developments, and proposed indicators await instrument validation.
- The framework requires implementation research, and no reviewed study linked teacher competency gains to student outcomes.
Citation
Shenoy, P., & Saarela, M. (2026). A critical review and actionable framework for integrating generative AI into teacher professional development. Discover Education, 5, 570.