How Does ChatGPT Decide on the Trolley Problem?

On March 25, 2024, Takashi Kido opened the symposium „Impact of GenAI on Social and Individual Well-being“. It is part of the AAAI 2024 Spring Symposia at Stanford University. Among other things, he reported on a study on the trolley problem. In it, not only humans but also ChatGPT answered the question of whether to set the switch (version by Philippa Foot) or push the fat man off the bridge (version by Judith Jarvis Thomson). In the discussion that followed, Oliver Bendel reported on an experiment that his Data Science student Sahan Hatemo is conducting as part of a mini-challenge. He developed dozens of personas that should lead to corresponding GPTs. They have different ages, ethnicities, genders, life plans, and so on. Once completed, they are also supposed to answer the aforementioned questions of the trolley problem. However, the student is considering reducing the complexity first and using an open source language model. There was a lively discussion about this project. Further information on this symposium is available here.

Fig.: A trolley in San Francisco