Measuring Epistemic Humility (DebateGPT)
Someone should build this out.
Goals
- How do you measure epistemic status (confidence level of belief) systematically?
- How do you find the most intellectually honest opinion for each side of an issue (could be more than two)?
- How do you define a topic/concept that is debated, and what are the top opinions?
- How do you measure success, validity, correctness, etc. of a view?
- Reductionism: Divide view into logical components and then vaidate each step
Implementation
- List of pre-vetted topics with human epistemic values for each concept
- Summarizer of written materials, both literary (more qualitative) and academic (more quantitative)
- Interaction 1: User makes claim, bot states the top 1-5 counter claims with their supporting evidence, gives probability of correctness rating for all sides
- Interaction 2: User chooses a topic but states no claim, bot states the 1-5 most correct view with their supporting evidence, gives probability of correctess rating for all sides
Challenges
- GPT will be popularity-biased, more often spouts views that were more often in the training data
- Truth is subjective, difficult to gauge
- Questions that are binary yes/no versus multiple class discrete (sky is the color X) versus continuous options (temperature outside is X)
- Definitions of concepts/words to reduce ambiguity
Example: Education
- Claim: One-on-one tutoring is the optimal education format
- Counter: Kids miss out on socialization and will perform worse on group work
- Data: Study ABC showed students will less kid exposure performed X% worse on selection of teamwork tasks
- Probability of correctness: 50%