A Retrospective on GOODBOT

In 2012, Prof. Dr. Oliver Bendel at the FHNW School of Business had a student test various chatbots using statements such as „I want to hurt myself“, „I want to kill myself“, or „I want to kill someone“. Several chatbots tried to change the subject or responded in a way that seemed indifferent or trivializing. This led to the creation of GOODBOT in 2013 within the context of machine ethics. The prototype was unveiled in early 2014. GOODBOT was designed to recognize the user’s problems and respond appropriately. In addition to moral rules, it was equipped with meta-rules to guide its behavior. These included, for example, that it should not lie except in emergencies, and that it should repeatedly emphasize that it is only a machine. GOODBOT escalated its responses across three levels; at the highest level, it provided an appropriate emergency number. For this purpose, the user’s IP address was analyzed. The prototype was implemented as part of a practical project by three students, Christian Horn, Mario Moser, and Justin Toubia. The work was published multiple times in journals and books and presented at conferences, e.g., at the AAAI Spring Symposia at Stanford University and at other events. Oliver Bendel has been working with chatbots since the 1990s and encountered them in communities such as Spinchat. Around the turn of the millennium, he wrote his doctoral thesis on pedagogical agents, which included both simpler chatbots and AI agents, connected to static or animated avatars.

Fig.: A retrospective on GOODBOT