AI Ethics: What 40 Million People Said About Self-Driving Car Crashes

When a self-driving car's brakes fail and a crash is unavoidable, who should the algorithm protect — the passenger inside, or the pedestrian outside? This question sounds abstract until you realise that engineers are encoding an answer right now, invisibly, in every autonomous vehicle on the road.

In 2018, researchers at MIT ran a large-scale experiment to study what people actually believe about these choices. They built a platform called the Moral Machine and collected over 40 million decisions from people in 233 countries — the largest cross-cultural moral survey ever conducted.

What the Moral Machine found

Participants were shown variations of the same scenario: an autonomous vehicle facing an unavoidable crash, with different groups in harm's way. The researchers varied who was inside the car, who was outside, how many people were involved, whether pedestrians were crossing legally, and other factors.

Across those 40 million responses, Awad and colleagues (Nature, 2018) identified several preferences that held broadly across cultures.

• –Save more lives when numbers differ — most people preferred outcomes that minimised total deaths
• –Spare younger people — a preference for saving children over adults, and adults over the elderly, held in most countries
• –Reward lawful behaviour — pedestrians crossing legally were more likely to be spared than those jaywalking
• –Spare humans over animals — among the most consistent cross-cultural findings in the entire dataset

Where the world splits

The headline finding was not consensus — it was variation. The researchers identified three distinct cultural clusters. Western countries (most of Europe and North America) showed a stronger preference for sparing pedestrians over passengers. East Asian respondents weighted the preference for sparing the elderly more than Western participants did. Countries in the Global South formed a third cluster, placing greater weight on saving more lives regardless of other characteristics.

This variation matters enormously for policy. A self-driving car programmed in California will embed the moral preferences of Californian engineers and regulators — then it will be sold in Tokyo, São Paulo, and Berlin. The Moral Machine showed there is no single global answer to who the algorithm should protect.

The deeper problem: who decides?

Even if there were a clear cross-cultural consensus, a second question remains unanswered: whose preferences should the car follow? The passenger who paid for it? The majority of society? A national regulator? An international standards body? None of these answers is obvious.

These are not technical questions. They are political and philosophical ones — about accountability, liability, and the legitimacy of encoding moral choices into code. When a human driver makes a split-second decision, we hold them responsible. When an algorithm does, the accountability chain fractures. Is it the software engineer? The company? The government that approved the car?

Three dilemmas you can vote on now

The self-driving car crash is not the only place AI now makes moral choices. The same logic applies to automated sentencing tools, content moderation systems, and hiring algorithms. Here are three dilemmas that put the questions directly — vote and see how you compare with people worldwide.

If AI is proven 30% more accurate than human judges in criminal trials, should we replace human judges? The accuracy argument is compelling. So is the accountability one.

When AI is projected to eliminate 30% of jobs in a decade, should governments slow it down or let markets and retraining programmes adapt?

Why your vote adds up to something

Surveys about abstract ethics often tell us what people think they believe. Moral dilemmas in a live voting context — where you see results from thousands of others in real time — reveal something closer to revealed preference. When you see that 58% of voters chose to protect the pedestrian, or that the split in your region differs from the global average, it changes how you hold the question.

The Moral Machine's 40 million data points were collected from self-selected participants — not a representative sample. SplitVote's votes come from the same kind of audience. But aggregate moral preferences, even imperfect ones, matter: they are part of how societies form norms, and norms eventually shape law.

SplitVote presents ethical dilemmas for reflection and discussion. Results represent our community's votes, not scientific conclusions. Source: Awad et al., 'The Moral Machine experiment', Nature 558, 59–64 (2018).