The implications of using large language models (LLMs) in educational technology around generative feedback have positive and negative implications particularly when its content related to religion and culture. This study examines the patterns of hallucination in AI feedback towards Islamic content in EFL student essays in Islamic learning context. A systematic evaluation was conducted of five leading LLMs (ChatGPT, Claude, Grok, DeepSeek, and Gemini) using a newly created hallucination multi-type classification framework. The LLMs were prompt to provide feedback on EFL student essays containing Islamic content; then an evaluation was conducted to the responses using 6 types of hallucination indicators namely source fabrication, attribution errors, content misrepresentation, citation errors, contextual misapplication, and theological misstatement. This study found clear patterns of hallucination across all models, which ChatGPT generated the most errors and Claude Sonnet 4 produces the least errors in his feedback. While LLMs responded to the prompts with hallucination in some forms, the patterns and implications of hallucination were distinctly different across all examples. This information is valuable for educators and developers to consider regarding the legitimacy of AI feedback in religious contexts.
Copyrights © 2026