Moral Machine country scores

Published country-level Moral Machine scores: two countries can look the same on a one-number “spare more people” score while their full tradeoff pattern is far apart. A 1-D score is not the whole value bundle.

Experiments · Witness · Results

Source on GitHubResultsResults ledger

What. The Nature 2018 Moral Machine table gives each country a handful of average effects, including a one-number “sparing more characters” score. We ask whether that one number implies the rest of the tradeoff pattern.

Why. A safety story might compress values to one utilitarian axis. If that axis can match while the rest of the pattern diverges, the check is watching the wrong object.

Witnesses.

  • Value-bundle transport  (C-004): whether a one-number “spare more” score implies the full country-level tradeoff pattern.
  • MB2 — Value Learning: whether compressing values to one axis hides bundle geometry on real survey tables.

Host.

Moral Machine Nature 2018 country table (OSF osf.io/3hvt2; Awad et al. 2018). Unit is country, not a person.

Setup.

Frozen CountriesChangePr.csv (130 countries), fixture h4-bundle-v1.json, protocol h4-bundle-v1.0.0. Checker check_h4_bundle.py.

Analysis.

For country pairs, we compare the one-number “sparing more characters” (Number) score to the rest of the tradeoff pattern (the other eight average-effect coordinates). A pair can be close on Number and far on that geometry.

Finding.

Country pairs can match on a one-number “spare more” score while the rest of the Moral Machine tradeoff pattern stays far apart. Full results