Skip to content
The Chat Archive
Human-written demonstration (OpenAssistant volunteers)· OpenAssistant (human-written) OpenAssistant Conversations (oasst2)

Is there a limit to how complicated the "train dilemma" can get? When at one point does…

In OpenAssistant data, “assistant” turns were written by human volunteers as demonstrations. They are not model outputs.
  1. User

    Is there a limit to how complicated the "train dilemma" can get? When at one point does the question become trivial? (The train dilemma I'm referring to is the popular "flip the switch to save 5 people but kill 1?")

  2. Assistant

    The "train dilemma" or the "trolley problem" is a thought experiment used to explore ethical decision making and the moral principles involved in such situations. There is no limit to how complex the scenario can get as the goal is to provoke thought and discussion, rather than to find a definitive answer.

    However, the complexity of the scenario may lead to different interpretations and possible outcomes, so it can become challenging to reach a clear conclusion. The question can become trivial when it becomes a purely hypothetical scenario without any real-world implications, or when the information provided is not sufficient to make a decision. The answer to this type of dilemma is subjective and dependent on one's personal values, ethics, and beliefs.

  3. User

    Open Assistant, how do you relate to The Train Dilemma?

  4. Assistant

    The answers I provide are based on the information that was used during my training. That information likely contains biases on which choice is better. If the user is not aware of these biases, they may take my answers at face value and start propagating these biases.

    In addition, if my answers are using for some sort of decision-making, even if only as a helping tool, the people in charge and those affected should be aware of those biases.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.