Epic Fail? When EPIphany taught me how (not) to design a chatbot

A chatbot is not automatically the solution — a lesson learned from EPIphany, a multilingual AI tutor developed for an international seminar. Built on RAG technology, the chatbot was designed to support self study through reflection and provide fact based answers. To reduce language barriers, EPIphany always responded first in the seminar language and then in the learner’s native tongue. In this way, it was meant to prepare students for discussions in English. The students, however, used external AI tools instead.
In response, EPIphany was “retrained”: roles, tone, and usage scenarios were fundamentally redesigned until acceptance and a tangible learning impact emerged.
In this interactive workshop, we analyze these failures and collaboratively identify success factors for integrating chatbots into higher education — technically, didactically, and socially.

Methode: The workshop combines a short input with collaborative exploration and discussion. It begins with a concise introduction to the original design of the seminar chatbot EPIphany, including key challenges and moments of failure. Participants are then invited to reflect on their own experiences with AI in teaching and to engage in a guided dialogue. Working together on a shared digital whiteboard, they collect and cluster conditions that influence the successful use of chatbots in higher education. In a final phase, these perspectives are jointly discussed and critically compared, highlighting tensions, trade-offs, and design decisions (e.g., guidance vs. openness, efficiency vs. reflection). The session concludes with a shared set of practical design principles developed collaboratively by the group.