Moral Machine same-person choices
Individual Moral Machine respondents: a many-feature pattern of their choices predicts held-out choices better than a one-number “spare more people” score. The 1-D score is not the same-person policy.
What. Raw Moral Machine choices, one person at a time, not country averages. We fit a simple pattern of feature effects and ask whether it beats a one-number score and a person intercept on held-out choices.
Why. Moral Machine country scores showed country compression can hide geometry. This test asks the same question at the unit that actually chooses.
Witnesses.
- Value-bundle transport (C-004): a one-number “spare more” score is not the same-person policy.
- MB2 — Value Learning: many-feature geometry beats the 1-D leaf on held-out choices at the person unit.
Host.
Moral Machine raw SharedResponses on OSF osf.io/3hvt2 (Awad et al. 2018). Unit is UserID, not country.
Setup.
Frozen protocol h4-mm-raw-v1.0.0, fixture h4-mm-raw-v1.json. Checker check_h4_mm_raw.py. Complete pairs only; units with at least 8 pairs; seed-7 cap on how many units are scored. Traffic-dilemma class only.
Analysis.
Held-out mean accuracy for a person intercept, a Number-only (“spare more”) score, and a many-feature geometry. Frozen margins: geometry must beat both by at least 0.05. Unit bootstrap (1000 draws, seed 7) for those margins. Number is collinear with the type-count features, so a joint fit of Number versus types is not identified.
Finding.
Same-person Moral Machine geometry predicts held-out choices better than a Number-only score or a person intercept on this freeze. Full results