Related Experiment Video
Updated: Oct 14, 2026
Using a Classroom-Based Deese Roediger McDermott Paradigm to Assess the Effects of Imagery on False Memories
Published on: November 14, 2018
Reference Hallucination in AI-Assisted Restorative Dentistry Education: An Educational Risk Assessment Using the
Esra Yildirim Manav1, Sıla Nur Usta2
1Department of Restorative Dentistry, Faculty of Dentistry, Lokman Hekim University, Ankara, Türkiye.
Purpose:
Artificial intelligence (AI) chatbots are increasingly used by dental students and faculty for literature review, coursework preparation, and scholarly writing. In restorative dentistry education, where evidence-based decision-making is fundamental, inaccurate AI-generated citations may pose educational risks. This study evaluated the bibliographic reliability of AI-generated references in restorative dentistry topics and examined their potential implications for dental education.
Methods:
Five large language model (LLM)-based chatbots were tested using ten standardised restorative dentistry prompts representing commonly assigned educational topics. Each chatbot generated 100 references (N = 500). References were manually verified using indexed databases and evaluated using the Reference Hallucination Score (RHS; range 0-11). Chatbot performance was compared using the Friedman test, followed by Dunn-Bonferroni post hoc pairwise comparisons. Binary bibliographic error parameters were analysed using generalised linear mixed models, with chatbot type, prompt complexity, and their interaction as fixed effects and prompt as a random intercept (α = 0.05).
Results:
A significant overall difference in RHS was observed among the evaluated AI chatbots (p < 0.001), with post hoc analyses identifying significant differences between specific model pairs. ChatGPT-5.2 and Perplexity AI demonstrated the lowest hallucination severity, whereas Copilot and DeepSeek-v3.2 demonstrated the highest. Prompt complexity did not significantly influence hallucination severity. Errors most frequently involved author names, journal titles, and DOIs.
Conclusions:
AI-generated references in restorative dentistry demonstrated substantial variability in bibliographic accuracy across the evaluated AI systems. These inaccuracies may pose potential risks to evidence-based learning and academic integrity, highlighting the importance of systematic reference verification and AI literacy in dental education.
