Assessing AI's Performance in Bulgarian Orthography: Implications for Educational Use

Artificial Intelligence (AI) is increasingly integrated into educational practice, yet its normative reliability in underrepresented languages remains insufficiently examined.This study evaluates the orthographic performance of three widely used AI systems: ChatGPT, Gemini, and Copilot when applying Bulgarian compound-noun spelling rules.A three-stage experimental design was implemented: baseline classification, rule exposure, and manipulation through false contextual statements.Accuracy was assessed by expert evaluation in accordance with the Official Orthographic Dictionary of the Bulgarian Language.Baseline performance varied substantially (ChatGPT: 42.9%, Gemini: 85.7%, Copilot: 42.9%).After the rule presentation, improvement was observed for ChatGPT (71.4%) and Copilot (100%), while Gemini remained stable (85.7%).However, under manipulation, all systems demonstrated marked instability (ChatGPT: 42.9%, Gemini: 42.9%, Copilot: 28.6%).The findings indicate that apparent rule alignment does not ensure stable normative reasoning.The results highlight the need for critical verification of AI-generated linguistic guidance in educational and inclusive learning environments.