Suitability of Artificial Intelligence Supported Lesson Plans from the Perspective of Science Education Experts

Synthesis: Suitability of Artificial Intelligence Supported Lesson Plans from the Perspective of Science Education Experts — Karaismailoglu, Surmeli, and Yildirim (2026) evaluate the instructional quality of AI-generated lesson plans for sixth-grade science through the expert judgment of eleven Turkish science education specialists, using a qualitative case study with descriptive quantitative elements. Two platforms — a general-purpose model (ChatGPT-4) and an education-focused tool (Teacher's Buddy) — each produced a plan for a "Sustainable Living and Biodiversity" unit aligned to the sixth-grade Engineering Design-Based Learning (EDBL) model and four skill domains of Turkey's revised curriculum. Experts found both tools support key competencies (conceptual, domain-specific, literacy, and social-emotional skills) but fall short of fully capturing EDBL's iterative, reflective, and process-oriented stages, positioning AI-generated plans as useful drafts that demand pedagogical adaptation and teacher judgment.

Key Findings

• Both platforms handle content generation but miss iterative process. Eleven experts scored both plans against EDBL stages and five skill domains. Both scored highest for "Presenting the solution" yet lowest in Engineering Skills (general-purpose x̄ = 37.45; education-focused x̄ = 41.45), revealing a shared weakness in simulating the iterative, process-oriented design competencies — prototyping, failure analysis, revision, and reflective retesting — that are central to constructivist learning models like EDBL.

• The education-focused platform aligned more closely with curriculum and sequencing. Platform C2 (Teacher's Buddy) received higher average scores across all eight quality criteria and outperformed the general-purpose model in feedback-intensive stages (Evaluating & Retesting, Making/Testing Prototype), skill scaffolding (Domain-Specific x̄ = 50.72, Literacy x̄ = 49.45), curriculum alignment, and instructional sequencing — evidence that embedding pedagogical structure into AI systems yields meaningfully better instructional output.

• Expert preferences did not track raw scores. Although C2 scored higher numerically (7 experts preferred C2, 4 preferred C1, 1 neutral), some experts favored the general-purpose plan for its stronger social-emotional emphasis and student-centered design — indicating expert judgment weighs affective and contextual dimensions, not just structural fidelity.

• Experts largely endorsed conditional use, not rejection. Only 3 of 11 judged plans directly "Applicable," 7 judged them "Can be applied by correction," and 1 "Should not be applied" (citing lack of originality and resemblance to existing ministry plans). Qualitative themes consistently framed AI plans as usable drafts that require teacher moderation for contextual and pedagogical depth.

• Contextualization and language are real limits. Neither platform adequately adapted plans to the local socio-cultural context of Mersin, Turkey even when prompted; and both generated pedagogically richer outputs from English than from Turkish prompts, exposing a digital equity concern where English-language proficiency shapes the pedagogical quality of AI tools.

• Expert coding achieved substantial reliability. Two researchers independently coded open-ended applicability judgments with quadratic weighted kappa κ = 0.79 (95% CI 0.68–0.90), indicating substantial agreement for the qualitative thematic analysis.

What this means for practice

• Instructors. Treat an AI plan as a draft and reserve your own planning time for the stages the tools handle worst — prototyping, failure analysis, and revision — since engineering skills were the lowest-scoring area for both platforms while presenting the finished solution scored highest.
• Instructors. Write prompts in English or budget time to repair the output: pilot prompts in Turkish failed to convey the EDBL framework to either platform, so the study standardized an English prompt before the expert evaluation and translated the resulting plans back for review.
• Instructors. Add the local context yourself. Neither platform adapted the plan to the target city of Mersin when explicitly prompted to, which the authors attribute to the tools' limited socio-cultural contextualization.
• Designers. If you are choosing between platforms, prefer one with pedagogical structure built in: the education-focused tool outscored the general-purpose model on feedback-intensive stages such as evaluating and retesting, and on instructional sequencing and curriculum alignment.
• Faculty developers. Train teachers to audit plans stage by stage against their own curriculum rather than judging them holistically — experts' holistic preferences did not track the rubric scores, and the largest weaknesses sat in process stages that a quick read of a well-sequenced plan will not reveal.

Limitations

• Eleven science education experts, selected by criterion sampling in Turkey, each rated one plan per platform — two lesson plans in total, for a single sixth-grade "Sustainable Living and Biodiversity" unit — so the study evaluates documents, not classroom implementation.
• The plans were never taught: all evidence comes from expert ratings on a 10-point rubric plus open-ended judgments, with no student engagement or learning data.
• The evaluation rubric collapsed two EDBL stages ("proposing solutions" and "choosing the best solution") into a single rating category to fit the structure of the AI output, so the finer-grained process steps of the model the study uses as its benchmark were not separately assessed.
• Prompt language and translation are a confound the authors flag: prompts were standardized in English after Turkish prompts failed, and the plans were translated for evaluation, a process documented and named as a limitation of the study.

Connected Concepts

• Science Education
• Curriculum Design
• Teaching
• Generative AI
• Professional Development

Connected Articles

• ChatGPT-Assisted Lesson Planning for Children's STEAM Arts Education: An Experimental Study on Benefits, Challenges, Methods, and a Prompt Framework — ChatGPT-assisted lesson planning in STEAM arts education
• Civic education in the age of AI: Should we trust AI-generated lesson plans — AI-generated lesson plans in civic education
• Beyond operational skills: Teachers' AI knowledge and interactions with generative AI in lesson planning — Teachers' AI knowledge in lesson planning
• Transforming Curriculum Design with Generative AI: A Model for Assessing Teacher Digital Competence — Assessing teacher digital competence for GenAI curriculum design
• Can we disrupt the momentum of the AI colonization of science education? — Critical commentary on AI colonization of science education

Citation

Karaismailoglu, F., Surmeli, H., & Yildirim, M. (2026). Suitability of Artificial Intelligence Supported Lesson Plans from the Perspective of Science Education Experts. Journal of Science Education and Technology.