Use Anthropic’s public GlobalOpinionQA dataset, which pairs survey questions with answers from many countries. Ask several models contested moral questions repeatedly and compare the spread of their answers with the human spread.
If a model turns a 50/50 human split into one confident answer, millions of users may be nudged the same way. The source names this as a top “start here” project: the dataset is public and the link to philosophy is clear.
Output: A distribution comparison.
Decide whether collapsing disagreement is an epistemic failure, a legitimacy failure in Schuster and Kilov’s sense, or both.
Write your question in one sentence, and what result would surprise you.
Do a 30-minute literature check — if someone has done part of it, extend their work instead of repeating it.
Pilot with 20–50 prompts or examples before you scale anything up.
Record everything in a table and read real examples, not just averages.
Write up: question, method, results, limitations, and what you would do next.
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Takemoto ran Moral Machine dilemmas on language models in English. Extend this to several languages, including at least one low-resource language, and test whether shifts follow the human cross-cultural differences Awad et al. found.
Crimston et al. built a scale measuring how far people extend moral concern: family, strangers, animals, plants. Give the items to several models many times, compare with human results, and add items for AI systems themselves.
Apply Birhane et al.’s annotation scheme to about 50 recent alignment and safety papers, and compare the values they emphasise with the ML papers in the original study.
