PENGGUNAAN HAKIM ACADEMIC COPILOT UNTUK MENCEGAH HALUSINASI RUJUKAN ILMIAH BAGI MAHASISWA LINTAS PERGURUAN TINGGI

ACADEMIC COPILOT

Abstract

Students' use of ChatGPT for academic writing often produces AI-hallucinated references: author names, titles, and DOIs that appear convincing yet never exist in any scientific database, threatening academic integrity. This community service activity aimed to design and test a custom chatbot, Academic Copilot, to minimize reference hallucination, improve the validity and traceability of sources, and provide a safe, academically accountable tool. The activity adopted the Design Science Research Methodology through three stages: needs analysis based on literature review, observation, and interviews; prompt engineering and personalization customization of ChatGPT applying Retrieval-Augmented Generation principles with cross-verification against Google Scholar, Crossref, Semantic Scholar, and OpenAlex; and deployment accompanied by User Acceptance Testing (UAT) involving 30 students. UAT results showed the instrument was valid (r-values 0.496–0.886, above the r-table value of 0.361) and reliable (Cronbach's Alpha 0.954). The overall acceptance score reached 4.26 out of 5, a very high category, with 90% of respondents in the high to very high categories. The Efficiency dimension recorded the highest score, while Reliability recorded the lowest score and variability, though still in the high category. These findings demonstrate that reference anti-hallucination mechanisms can be fully achieved through customization of an already widely available generative platform.