Artificial Intelligence in Teacher Education: A Systematic Literature Review of Pedagogical Frameworks

Synthesis: Shukla (2026) reports a PRISMA-informed scoping review of how Professional Development programs prepare future teachers for generative AI, searching open databases for work published between January 2020 and October 2026 and arriving at a corpus of seven sources: three framework or guidance documents and five empirical studies, with Prilop et al. (2025) counted in both groups because it reports a survey alongside a conceptual contribution. Frameworks extend Technological Pedagogical Content Knowledge (TPACK) and place ethics inside pedagogy; the empirical work, small-scale and cross-sectional, finds uneven and narrow adoption, heavy concentration on ideation, and overconfidence that corrects when pre-service teachers are made to evaluate AI output. In a Finnish pilot, 27 percent of participants had never used generative AI at all, and no study in the corpus measured pupil learning or the quality of AI-supported teaching products. The review's value is an honest map of a gap that widens where it matters: ethics is named in every source but never operationalized as something a program can teach or assess, so anyone building teacher AI competency inherits vocabulary without evidence.

Key Findings

1. The corpus holds seven sources, not seven studies. Three framework or guidance documents and five empirical studies make the corpus, with Prilop et al. (2025) counted twice because it also reports survey data.
2. Frameworks converge on ethics and diverge on instruction. TPACK extensions, the UNESCO framework's 15 competencies and the triadic account all treat AI knowledge as more than technical operation, but none specifies instruction.
3. Adoption is uneven and narrow. In a Finnish pilot (N = 77), 27 percent of pre-service teachers had never used generative AI, and users concentrated on ideation rather than adaptive or self-regulated learning uses.
4. Guided critique recalibrates confidence. Among 59 pre-service teachers, confidence that ChatGPT could write complete, sufficiently detailed lesson plans fell significantly after guided analysis, and participants then treated it as a starting point needing revision.
5. Personal use does not produce pedagogical thinking. In the Irish study, personal and academic use of generative AI was more frequent than lesson-planning use and showed no significant correlation with perceived educational opportunities.
6. Teacher educators' own use shapes their views. A Danish survey (n = 91) found generative AI use mediated the link between motivation and confidence and perceived potential, and educators requested formal and informal development.
7. No study measured learning. Dependent variables across the five empirical studies were perceptions, intentions, readiness and self-reported use; none assessed lesson-plan quality, pupil learning, or change in practice over time.

What the reviewer searched and what it produced

Shukla ran the search in October 2026 across web-based scholarly search covering publisher sites, institutional repositories and open-access journals, combining artificial intelligence, generative AI, ChatGPT, AI literacy and AI competency terms with teacher education, pre-service teacher, initial teacher education and teacher educator terms. A second query added TPACK, competency framework and ethical guidelines. The window was January 2020 to October 2026. Criteria were applied at title and abstract, then to full text or abstract, and the European Commission guidelines were carried as supporting guidance rather than counted, because they address educators in general rather than teacher preparation. Screening produced seven distinct sources. Synthesis was narrative and thematic, with no meta-analysis attempted and no formal quality-scoring instrument applied, and the mapping of each study onto the five UNESCO dimensions is presented as the reviewer's own interpretation of the abstracts.

Why the frameworks answer different questions

Three families appear: TPACK extensions, the UNESCO competency framework, and ethical guidance. Mishra and colleagues reconsider the knowledge domains in light of properties they attribute to generative AI (protean, opaque, unstable, generative, social) and treat ethical integration as the purpose of the update rather than a separate domain. Prilop and colleagues split AI Literacy into AI as teaching tool, teaching content and learning tool, which makes visible that a program can be strong on one facet and silent on another. UNESCO offers the most extensive structure and the only one organized to inform national policy: 15 competencies in 5 dimensions at three progression levels, Acquire, Deepen and Create. Asked what should be taught, how it should be taught, and how learning would be known, no source answers all three, and none answers the third with an instrument. That absence explains the measurement problem downstream.

Why the evidence stops at perceptions

The five empirical studies are uneven in both senses of the word. The Finnish pilot (N = 77) found more than a quarter of participants had never used generative AI and most of the rest used it only sporadically, mostly for ideation and content creation; the authors read this as limited awareness, though a reading in which students apply tools to the tasks in front of them is equally open to a cross-sectional design. The Irish study found personal and academic use more frequent than lesson-planning use, with no significant correlation between personal use and perceived educational opportunity. Kalenda and colleagues (n = 59) supplied the one intervention: after guided analysis of ChatGPT-generated lesson plans, confidence in the tool's completeness and detail declined significantly, and participants named missing content, insufficient Scaffolding and weak differentiation as the weaknesses needing human revision. Estocapio and Bilog applied the Technology Adoption Models and found positive attitudes alongside awareness of ethical risk. Across all five, the dependent variables are perceptions, intentions, readiness and self-reported use.

Where ethics stops

Every source names Ethics. Frameworks place it as purpose or as a dimension, teacher educators raise it as a concern, pre-service teachers recognize it as a risk, and one study treats structured training as the response to it. The consensus is thinner than it looks: the corpus says little about what ethical competence consists in for a pre-service teacher, how it is taught, or how it is assessed, and no empirical study uses the UNESCO ethics dimension to design instruction or measure outcomes. Three issues recur — accuracy and bias in generated content, preservation of basic skills and assessment integrity, and whose interests AI serves in education — and the third is absent from every empirical study. Geography compounds the problem: the identifiable settings are Finland, Ireland and Denmark, so claims about equity for low-connectivity, large-class or non-European systems rest on a framework's principles rather than on evidence.

What this means for practice

• Instructors. Make evaluation an assessed task, not an assumed skill: confidence in AI lesson plans declined only after students were required to analyze them, and participants in the Finnish pilot modified outputs and verified externally without always making that critique explicit.
• Faculty developers. Build development around hands-on use rather than information sessions, since educators' own generative AI use mediated the link between their motivation and confidence and their sense of the technology's potential.
• Curriculum designers. Audit the program against the three facets of AI literacy — AI as teaching tool, as teaching content, as learning tool — because the triad exists precisely to expose facets a program neglects.
• Administrators. Reconsider what counts as evidence of competence when lesson plans and case analyses can be generated in minutes; observed teaching, verbal defense of a plan, and critique tasks give programs artifacts they can actually warrant.
• Researchers. Measure products and pupils rather than attitudes, and report tool versions and prompts, since a confidence shift observed after analyzing output from one model version may not replicate with the next.

Limitations

• The search ran on open web-based scholarly search without Scopus, Web of Science or ERIC, screening and extraction were done by a single reviewer with no independent verification, no quality-scoring instrument was used, and several sources were extracted from abstracts and publisher metadata alone — hence a corpus of seven.
• Identifiable settings are Finland, Ireland and Denmark; there is no evidence from lower-resource or non-European teacher education systems, which matters because benefits and harms are unlikely to distribute evenly.
• Samples ran from 59 to 91 where size was stated; the Finnish study describes itself as a pilot that also tested a new instrument, and the intervention study measured confidence before and after a task without an apparent comparison group, so the shift cannot be shown to last.
• Studies from 2024 used tools that have since been updated, so a decline in confidence after analyzing one model's lesson plans may be smaller or absent with a later version.

Citation

Shukla, V. K. (2026). Artificial Intelligence in Teacher Education: A Systematic Literature Review of Pedagogical Frameworks, Empirical Evidence, and Ethical Implications (2020-2026). EdArXiv preprint.