THEORETICAL AND APPLIED ASPECTS OF GENERATIVE AI: FROM LANGUAGE MODELS TO PRACTICAL APPLICATIONS IN EDUCATIONAL CONTENT CREATION

Generative AI (GenAI) presents transformative potential for educational content creation, yet a significant gap exists between its theoretical development and practical application. This leads to a high risk of misapplication and the propagation of factually unreliable “fluent hallucinations.” This research bridges this theory-practice gap by constructing and validating a novel techno-pedagogical framework, aiming to quantitatively link GenAI’s theoretical properties (e.g., training data) to its applied performance. A sequential explanatory mixed-methods design was used. We codified the theoretical aspects of five GenAI models and conducted a quasi-experiment, generating 2,500 content pieces from 500 prompts. This corpus was evaluated by 15 domain experts using a validated Pedagogical Content Quality Rubric (PCQR). A weak correlation (r = .19) was found between output fluency and factual accuracy, confirming the “fluent hallucination” phenomenon. Multiple regression (R^2 = .68) identified training data composition (\beta = .55) and instruction-tuning (\beta = .24) as the strongest predictors of pedagogical quality; model parameter size was non-significant. The study concludes that GenAI’s pedagogical utility is predictable based on its theoretical architecture, moving the evaluation from a “black box” to a “gray box” model. We recommend a shift toward verifiable, domain-specific tools and repositioning educators as critical validators.