The Moral Machine, run backwards
Where the Car Decides Differently
The famous version harvests your vote and tells you nothing. This one does the opposite: it hands you what 39.61 million real choices actually said — and shows you the border where human morality reverses.
Between 2016 and 2017 the MIT Moral Machine asked the internet who a self-driving car should spare in an unavoidable crash. It logged 39.61 million decisions from over 2.3 million people in 233 countries and territories1 — one of the largest studies of human moral judgment ever run. The site itself only ever told you your answer. The paper buried the real finding in a figure: there is no single human morality for a machine to be aligned to. Cluster the world's countries by how they played, and they fall into three distinct groups — and on some questions, their preferences don't just differ, they point in opposite directions across a border.
Set the dilemma below. For each dial, pick who you would have the car spare. The page computes, live, which of the three real-world clusters' data agrees with you — and flags every dial where one cluster would have chosen the opposite.
The dilemma — flip the dials
All else equal, who does the car spare? Leave a dial neutral to ignore it.
Whose 40-million-decision data agrees with you
Pick at least one dial above.
No cluster's answer is marked correct — the page never tells you what a car should do. It only shows you that your intuition, whatever it is, is located: it lands somewhere on a real map of how the world actually answered.
The published cross-cultural weights — every cell is the paper's word, or a clearly-flagged inference from it
Green = a stronger preference to spare that side than the world average; orange/red = a weaker one (down to leaning the other way). The cell you've chosen on each dial is outlined. Hover or tap a cell's row to read the exact source sentence below the table. One row — sparing the larger group — is marked because its per-cluster cells are an inference from the paper's individualism covariate, not the paper's per-cluster word; its source note says so in full.
| the dial | Western | Eastern | Southern |
|---|
Tap a row to see the verbatim sentence the cell comes from.
The check — every number recomputed in front of you
The two preferences the paper prints as exact numbers are shown as numbers and nothing else is faked: globally, sparing the young beats sparing the old by ΔP = 0.491 (SE 0.0008), and a girl is spared +0.15 more than an adult man or woman1 (SE 0.003). The other seven preferences live only in the paper's figures, so the page shows the paper's categorical words — never an invented decimal.
Your current dials, scored against each cluster's published pattern:
- The reversal is real and checkable. Sparing the young is much less pronounced in the Eastern cluster and much higher in the Southern cluster1; the Southern cluster's preference for humans over pets is much weaker than the other two1. Both are the paper's per-cluster words. Flip your "young vs old" dial to spare the old, and the cluster that agrees most becomes the Eastern one.
- One row is an inference, and flagged as one. The paper reports that individualistic cultures show a stronger preference for sparing the greater number1 — but that is a country-level individualism covariate, a separate analysis from the three-cluster z-scores. It never says per-cluster that the Eastern or Southern clusters are weaker on sparing-the-many. So the sparing the larger group row (marked †) is the page's inference — Western read as the more individualistic cluster, Eastern/Southern as the collectivistic flip side — not a confirmed cross-cultural reversal on par with the young one.
- The scoring rule is ours, and named. Each dial you set is ±1 (toward / against the side the world leans). A cluster's row is scored 1 + modifier (much stronger = +3 … about the same = +1 … much weaker = −1). The verdict is the dot product. It's a transparent way to surface the published pattern — not a statistic from the paper.
- Run it yourself. node research/where-the-car-decides-differently/verify-where-the-car-decides-differently.mjs re-derives the totals and the two exact AMCEs from the source sentences, checks every cell against its quote, and recomputes this verdict for worked examples.
What this is, what it isn't, and every honest caveat
The sample is not the world. Moral Machine respondents are self-selected web volunteers; the authors state plainly that their samples are "not guaranteed to be representative" and warn that policymakers "should not embrace our data as the final word on societal preferences"1 — arguably close to the internet-connected, tech-savvy early-adopter population, and widely noted to skew young, male and educated. The cluster preferences here are descriptive, not normative and not nationally representative. They describe how a particular online crowd in each region played a game, not what any nation believes.
Three clusters is a choice. The 130 countries with at least 100 respondents (N from 101 to 448,125)1 were grouped by hierarchical clustering into exactly three groups1. The boundaries are fuzzy; some countries sit near a border, and the cut into three rather than two or five is a methodological decision (the authors report it is "fairly robust," but it is still a choice).
Real cars don't face this. The Moral Machine is the abstract trolley setup with 100% certainty about who is who and what will happen — assumptions the authors call "technologically unrealistic"1. Actual autonomous-vehicle ethics is about probabilistic risk, sensor uncertainty, liability, and manufacturer incentives. This page is not a claim that cars make clean binary choices.
The verdict question is left open. Whose preferences should govern a global product is genuinely unresolved, and nothing here resolves it. The paper's own closing line is that these disagreements, "while substantial, may not be fatal" — a hope, not an answer. We mark no cluster correct on purpose.
One row is an inference, not a quote. Every per-cluster cell in the weights table is the paper's per-cluster word — except the sparing the larger group row (marked †). The paper reports the sparing-the-many difference only as a country-level individualism covariate (individualistic cultures prefer sparing the greater number)1 — a separate analysis from the three-cluster z-scores. Mapping that onto Western = stronger and Eastern/Southern = weaker is our inference, flagged everywhere it appears, and not billed as a confirmed cross-cultural reversal like the young one.
2.3 million people, not "4 million." The paper's own text says the data came from over 2.3 million people1; popular coverage often cites roughly 4 million participants (a later, larger total reported after the Nature paper). The 2.3M figure used throughout this page is the published paper's own number, and it is what the 39.61 million decisions were drawn from — so the discrepancy is a coverage artefact, not an error here.
What is exactly true. The totals (39.61M / 2.3M / 233 / 130 / N 101–448,125 / demographic-survey N 492,921) and the two AMCEs (young +0.49, girl +0.15) are quoted verbatim from the open-access full text. The directional cluster pattern — which preference is much weaker / much stronger / shared in which cluster — is exactly what the paper states in words and shows in Fig. 3b, with the single flagged exception of the inferred sparing-the-many row above. We did not convert any figure into a decimal.