Antisemitic Content in Chatbot Responses
LLM-based classification for a study at the University of Bern
A three-week research assistantship (August 2026) at the Institute of Communication and Media Studies, University of Bern, with Dr. Mykola Makhortykh.
I built an LLM-based pipeline to classify antisemitic content, following the IHRA definition, in about 30,000 chatbot responses. I compared six LLMs as classifiers (F1 = 0.86). The study also tested how well LLM safety filters hold up against different prompting strategies, such as role-play framing. The findings were presented at the Weizenbaum Institute.
Methods: Python, LLM-based classification, LLM evaluation.