Research Article
The integration of generative artificial intelligence in secondary education: A systematic review
Synthesis: Pérez-Montesdeoca, Rodríguez-Rodríguez, and Fernández-Sogorb (2026) searched Scopus and Web of Science on 11 November 2024 and synthesized 33 studies of generative AI in secondary education under PRISMA (κ = .89). Only 21 collected original classroom data; MMAT appraisal put that empirical work at moderate to high quality, with small samples the recurring weakness. Twenty-six studies describe how the technology is used — teacher-initiated writing support, STEM problem solving, assessment workflows — and 7 examine what changes for students. The review's argument is conditional: writing, motivation, and critical thinking gains appear where teachers mediate the tool and students judge its output, and fade where the tool produces the work.
Key Findings
- Adoption is teacher-led and scattered. Of the 26 use-focused studies, 7 examined problem solving, 6 assessment, 3 question generation, 3 teacher training, and 2 each for narrative creation, image generation, and language learning — nearly all teacher-initiated and run in isolation.
- Accuracy collapses as tasks get harder. BingChat and Bard reached 92.11% on basic biology questions, yet every tool fell below 20% on higher-order items, and ChatGPT-3.5 dropped from 83.4% to 63.8% as physics content advanced. GPT-4 and GPT-4o solved 82%-87% of mathematics problems; Gemini solved 32%.
- Automated grading is close but not substitutable. LLM agreement with human evaluators reached only 50.8%, and a fine-tuned ChatGPT-3.5 beat BERT by 9.1% in scoring accuracy — useful inside a supervised workflow, not grounds for removing the teacher.
- Feedback is the strongest application. Meyer et al. found LLM feedback improved text revisions and motivation, and Khan et al. reported that immediate, tailored vocabulary feedback nearly doubled scores.
- Students who evaluate gain; students who delegate do not. Lee et al. found full delegation narrowed learning while comparing and rejecting output deepened engagement; Chen and Chang found ChatGPT with game-based learning raised performance (M = 49.06 vs. M = 44.46) and lowered cognitive load (M = 3.24/3.32 vs. 3.74) against the game alone.
- Self-reported gains are thinner than the enthusiasm. In Tamimi et al., 37.4% of students reported better performance and 44.3% reported no significant change in productivity.
What the corpus contains
The search covered Scopus and Web of Science, limited to peer-reviewed primary research from 2019 to 2024 in English or Spanish. From 3,381 records, deduplication and criteria left 676; screening dropped 322 on title and abstract, 3 for access and 2 as retracted, then 257 with no GenAI focus, 46 without a learning-improvement focus, and 13 outside secondary education — 33 studies. The corpus is young: 2 appeared in 2021, 2 in 2022, 6 in 2023, and 23 in 2024; 8 came from the United States. Twelve were technical evaluations or conceptual discussions with no human participants; the MMAT covered only the 21 involving students, teachers, or classrooms.
Where schools are using it
Writing and material production dominate. Levine et al. found students using ChatGPT to plan, revise, and translate, treating output as a starting point rather than an answer — though they often accepted or rejected whole suggestions without inspecting them. Bulut and Yildirim-Erbasli and Carrasco generated stories, quizzes, and history materials.
STEM problem solving is the largest category, with 7 studies, and its pattern is consistent: strong on routine items, confident failure on complex ones. Parra et al. found chatbots could not turn geometry descriptions into visual representations; Polverini and Gregorcic reported ChatGPT at 41.7% correct on kinematics graphs against 47% for students, with correct visual interpretation in only 30.9% of responses.
Assessment follows with 6 studies. Ciampa et al. described five levels of teacher-led integration, only the highest of which asks students to critique the model's output.
What changes for students
The seven impact studies report gains that are real but conditional. In Chen and Chang's comparison the ChatGPT group reported higher perceived competence (M = 2.63/2.54 vs. 2.21) while intrinsic motivation stayed similar, suggesting the tool acted as scaffolding rather than as a motivator. Kangasharju et al. found poetry-writing support benefited students equally regardless of prior grades — one of the few results pointing toward equity.
Critical evaluation does not transfer on its own. Ali et al. found no change in deepfake detection after workshops (M = 54.68 pre-test, M = 53.87 post-test), though students afterward named source verification (53.33%) and sender reliability (46.66%) as credibility criteria. Aprin et al. did find a 57% case-level improvement in judging altered images with AI support, but the effect weakened where little online information existed to check against.
Through the ICAP lens the authors favor, the corpus sorts by engagement rather than by tool: passive copying and delegation produce little, while arguing with and correcting the model is where gains sit. cognitive offloading is the failure mode the review keeps circling, and teacher mediation separates the outcomes.
What this means for practice
- Instructors: Assign GenAI to stages you can inspect — planning notes, revision rationales, comparisons of model answers — not the finished product, since gains appear where students judge output.
- Instructors: Budget editing time; every study producing texts, quizzes, or lab procedures reported errors or missing context.
- Administrators: Treat AI literacy as taught curriculum, not assumption. Deepfake workshops left detection flat, so skepticism must be practiced against real examples.
- Designers: Build outputs that invite checking — flag uncertainty, show sources, expose reasoning — because the documented failures are confident errors, not refusals.
- Researchers: Measure process alongside post-test scores; the corpus cannot say whether gains persist.
Limitations
- No meta-analysis. Narrative synthesis across 33 heterogeneous studies, with no pooled effect or reporting-bias assessment.
- Twelve of the 33 studies tested tools, not learners. They support claims about capability, not evidence about learning.
- The empirical base is small and short. Most studies used small samples, exploratory designs, or single-session interventions, with partial MMAT criteria on sampling and validation.
- The search stops in November 2024, across two databases, in English and Spanish only; terminology was also inconsistent, forcing case-by-case relevance judgments.
Citation
Pérez-Montesdeoca, H., Rodríguez-Rodríguez, D., & Fernández-Sogorb, A. (2026). The integration of generative artificial intelligence in secondary education: A systematic review. Contemporary Educational Technology, 18(3), Article ep667.