The Trolley Problem: How a Runaway Tram Became a Tool for Brain Science

A runaway trolley is heading toward five workers on the track. You can pull a lever to divert it onto a side track, where it will kill one worker instead. Do you pull the lever? Most people say yes. Now: the same trolley, the same five workers, but now you are standing on a footbridge above the track next to a large man whose body weight, if thrown off the bridge, would stop the trolley and save the five. Do you push him? Most people say no. Both scenarios involve killing one person to save five. The difference in intuition between them has generated more than half a century of philosophical argument, neuroscience research, and policy debate. Here is where that argument stands.

The Setup: A Philosopher's Thought Experiment

Philippa Foot introduced the trolley scenario in a 1967 paper titled "The Problem of Abortion and the Doctrine of Double Effect," published in the Oxford Review. She was not primarily interested in trolleys. She was using the scenario to test the doctrine of double effect — a principle in moral philosophy that distinguishes between harm caused as a side effect of achieving a good outcome versus harm caused as a means to that outcome.

Source: Trolley problem — Wikipedia

In Foot's original framing, the distinction mattered for debates about abortion, euthanasia, and the morality of killing in war. The trolley was a pedagogical device: clean, artificial, stripped of all the messy particulars that make real moral cases complicated, designed to isolate one variable at a time.

The basic scenario — pull a lever, divert a trolley, kill one to save five — strikes most people as clearly permissible. The utilitarian calculation favors pulling: five lives saved is better than one life lost. The Kantian objection raises a concern about treating a person as a means to an end, but when the person on the side track is merely in the way (not instrumentalized as a tool), even many Kantian thinkers find pulling the lever acceptable.

But Foot noticed that people's intuitions changed sharply when the scenario changed in specific ways, even when the utilitarian arithmetic stayed the same. This gap between intuition and calculation became the engine of the research program.

Thomson's Expansion: The Fat Man Arrives

In 1976 and again in 1985, philosopher Judith Jarvis Thomson developed the most important variant of Foot's scenario. In her "footbridge" problem: you are standing on a footbridge over the trolley tracks. A large man stands next to you. The only way to stop the trolley and save the five workers is to push the large man off the bridge, using his body to stop the trolley. He will die; the five will be saved. Do you push?

Source: Fat Man Variant of the Trolley Problem — Philopedia

In Thomson's essays "Killing, Letting Die, and the Trolley Problem" (1976) and "The Trolley Problem" (1985), she used this contrast to probe the moral significance of using someone as a means versus merely redirecting a threat.

Source: Trolley problem — Britannica

The footbridge scenario produces a dramatic shift in moral intuition. While roughly 85-90% of people say they would pull the lever in the switch scenario, fewer than 30% say they would push the large man in most survey populations. The arithmetic is identical: one person dies, five are saved. The difference is that pushing uses the man's body as a mechanical instrument in a direct, physical, intentional way, while pulling the lever kills the one person as a foreseen but unintended side effect.

This is the philosophical puzzle at the heart of the trolley problem: if five lives outweigh one life, why does it matter how you bring about that outcome? And if it does matter, why? Articulating a principled answer to that question has proved surprisingly difficult.

The fMRI Study: Greene Puts the Trolley in a Brain Scanner

For most of its history, the trolley problem was a tool for philosophers, useful for testing intuitions but not generating empirical data. That changed in 2001, when Joshua Greene and colleagues at Princeton published "An fMRI Investigation of Emotional Engagement in Moral Judgment" in Science.

Greene's insight was that the switch scenario and the footbridge scenario might produce different intuitions because they engage different brain systems. The switch scenario — distant, mechanical, impersonal — might engage deliberate, "cold" reasoning systems. The footbridge scenario — close, physical, requiring direct action against a person — might engage emotional systems in a way that produces visceral reluctance.

Source: An fMRI Investigation of Emotional Engagement in Moral Judgment — Science

Greene presented participants with a series of moral dilemmas, including both "personal" dilemmas (requiring direct physical harm to a person) and "impersonal" dilemmas (harm caused indirectly or at a distance) while they were being scanned. His key findings:

Personal moral dilemmas — like the footbridge — produced greater activation in brain regions associated with emotion and social cognition: the ventromedial prefrontal cortex (vmPFC), the amygdala, and the posterior cingulate cortex. These are areas implicated in automatic emotional responses and the processing of socially relevant information.

Impersonal moral dilemmas — like the switch scenario — produced more activity in the dorsolateral prefrontal cortex (dlPFC), associated with deliberate reasoning, working memory, and cognitive control.

Source: An fMRI Investigation of Emotional Engagement in Moral Judgment — Princeton

In cases where participants did endorse pushing the large man, response times were significantly longer than cases where they refused — consistent with a process of deliberate reasoning overriding an initial emotional rejection. The brain was more conflicted, and the conflict was visible in the timing.

Greene interpreted this as evidence for a "dual-process" model of moral judgment: an automatic, emotion-based system that generates rapid deontological intuitions (don't harm people directly) and a slower, deliberate system that generates utilitarian calculations (maximize welfare). Moral judgment, on this view, is the output of two partially competing cognitive systems.

The Protocol: How the Research Expanded

The 2001 Greene paper opened a wave of experimental moral psychology. The trolley problem and its variants became a standard instrument — a kind of moral spectrometer that researchers could use to probe how different conditions shift moral judgment.

A 2014 study by Skulmowski and colleagues used virtual reality to test whether people behave differently in immersive simulated scenarios compared to written hypothetical versions of the footbridge problem. The finding: in fully immersive VR where participants could see and "physically" push the large man, willingness to push dropped significantly below the rates seen in paper-and-pencil surveys. The abstract written trolley problem and the visceral VR trolley problem are psychologically different tasks.

Source: The Trolley Problem Tries to Probe Our Moral Compass — Slate

This finding matters because it suggests that the years of survey data on trolley problem intuitions may underestimate how strongly people resist direct personal harm in realistic situations. What people say they would do in a hypothetical scenario may not predict what they would do in a more realistic one.

The Timeline

The MIT Moral Machine: 40 Million Decisions

In 2016, MIT Media Lab launched the Moral Machine — an online platform presenting autonomous-vehicle versions of the trolley problem to the public. Instead of workers on a track, the scenarios involved self-driving cars facing unavoidable crashes: should the car prioritize its passenger or pedestrians? Should it spare the young over the old? Humans over animals? The law-abiding over the jaywalking?

By the time the study closed, 40 million decisions had been collected from participants in 233 countries and territories, making it one of the largest moral surveys in history. Awad and colleagues published the results in Nature in 2018.

Source: The Moral Machine Experiment — Nature

The headline results showed three preferences that appeared relatively consistent across cultures:

1. Spare humans over animals.
2. Spare more lives over fewer lives.
3. Spare the young over the old.

But the cross-cultural variation was substantial, and it tracked identifiable cultural dimensions. The study identified three major cultural clusters:

Western cluster (US, Europe, and related countries): stronger preference for sparing pedestrians over passengers, stronger preference for sparing the young, stronger preference for women.

Eastern cluster (Japan, China, and related countries): stronger preference for sparing the elderly relative to Western norms; lower pedestrian-versus-passenger preference differential.

Southern cluster (Central and South America, parts of Africa and Middle East): higher tolerance for sparing jaywalking pedestrians (compared to law-following pedestrians) than other clusters; stronger fitness/youth preferences.

Source: Should a self-driving car kill the baby or the grandma? — MIT Technology Review

Wealth and inequality within societies correlated with preferences: participants from more unequal societies showed stronger preferences for sparing high-status individuals. Participants from more individualistic cultures placed relatively stronger emphasis on maximizing lives saved.

Source: The Moral Machine Experiment — MIT Media Lab

The Controversy: What the Trolley Problem Cannot Tell Us

The trolley problem has been the target of sustained methodological criticism, particularly after the Moral Machine data started driving policy discussions.

The ecological validity problem is the most persistent. The scenarios are radically artificial. Real moral decisions — including real driving decisions — unfold under time pressure, with incomplete information, among people who know each other or have social relationships. Laboratory and survey trolley problems strip all of this away. Behavior in real emergencies may look nothing like survey responses.

The framing problem is related. Small changes in how the trolley problem is described produce large changes in intuitive responses. Whether the large man is described as "a large man" versus "a large stranger" versus "an innocent bystander" shifts responses. This sensitivity to framing suggests the trolley problem is partly measuring how people respond to the linguistic packaging of a moral scenario, not a stable underlying moral intuition.

Source: It's Time to Retire the "Trolley Problem" — Nautilus

The application problem is the most practically significant. Autonomous vehicles do not actually face trolley-problem scenarios. They face questions about sensor design, failure mode handling, and statistical population-level risk distribution — none of which maps neatly onto a one-versus-five framing. Critics argue that using the Moral Machine data to guide self-driving car ethics is applying a philosophical toy to a real engineering problem.

Greene himself has noted that the dual-process theory derived from trolley problem research should be treated as a hypothesis about mechanisms, not a prescription for how moral judgment should work.

Source: Solving the Trolley Problem — Harvard

Replication and What Survives Scrutiny

The basic trolley problem findings — strong approval of switching, much weaker approval of pushing — replicate across cultures and methods, though with meaningful variation in magnitude. The Greene dual-process model has been influential but is also contested. Some researchers argue that what looks like two competing systems is better described as one system processing different morally relevant features of scenarios (physical contact, personal force, directness of harm) that vary systematically between the switch and footbridge problems.

The fMRI findings from Greene 2001 have held up in neuroimaging replications. The ventromedial prefrontal cortex and amygdala do consistently activate more during personal moral dilemmas; the dorsolateral prefrontal cortex is more engaged in deliberate moral reasoning. However, the specific claim that these two systems are "competing" rather than "cooperating" is more difficult to establish from imaging data alone.

Source: Neuroscience, the Trolley Problem, and Moral Virtue — Academia

The cross-cultural robustness of the switch/footbridge intuition gap is arguably the most significant replication finding. The gap exists across Western, East Asian, South Asian, and African populations, though its size varies. This suggests the intuitive resistance to direct personal harm is not a WEIRD (Western, Educated, Industrial, Rich, Democratic) artifact, even if the specific magnitude of responses differs.

Where It Stands Today

The trolley problem's most active frontier in 2026 is AI alignment. As language models and autonomous agents increasingly make decisions with moral dimensions, alignment researchers use trolley problem variants to probe whether models have internalized coherent moral reasoning or are pattern-matching to surface features of ethical language.

Early research suggests that major language models (including those released in 2023-2025) tend toward utilitarian responses when given abstract trolley problem framing — maximizing lives saved — but shift toward deontological responses when scenarios are made more concrete and personalized. This mirrors the human pattern Greene identified, and raises the question of whether the models have learned this pattern from data about human moral reasoning or whether something functionally analogous to the dual-process structure has emerged from their training dynamics.

Source: Revisiting philosophy with fMRI — APA Monitor

The trolley problem has also entered medical ethics in contexts far from its origins. Ventilator allocation protocols during the COVID-19 pandemic explicitly invoked utilitarian principles that mirror the trolley problem's logic: with limited resources and more patients than resources, how do you choose? The committees developing these protocols grappled with exactly the philosophical tensions that Foot and Thomson had identified: the difference between redirecting a scarce resource (utilitarian triage) and actively withdrawing treatment from an existing patient to give it to a better-prognosis candidate (much more emotionally resisted, even when the arithmetic is the same).

Source: Medical ethics and the trolley problem — PMC/NCBI

What the trolley problem has proven, across more than five decades, is not that it provides answers to moral questions. It has proven something more precise and more useful: that human moral intuitions are generated by multiple processes with different computational properties, that those processes can be dissociated in controlled conditions, and that the dissociation follows patterns that are partially universal and partially cultural. That is a genuine empirical contribution to our understanding of the mind — even if the trolley itself never helps us decide what to do.

Why the Trolley Problem Doesn't Apply to Self-Driving Cars: the Actual Research Consensus

When autonomous vehicles entered mainstream coverage around 2016, journalists and ethicists reached for the nearest available philosophical framework: the trolley problem. The MIT Moral Machine project, which Iyad Rahwan co-led, was explicitly designed around trolley-like AV scenarios, and it produced striking data on cross-cultural moral preferences. But roboticists and safety engineers found themselves increasingly frustrated by the framing — not because ethics are irrelevant to autonomous vehicles, but because the trolley problem is the wrong model for how those vehicles actually work.

Real autonomous vehicle decision-making in an emergency is not a choice between one life and five. It is a collision-avoidance algorithm executing a minimum-cost trajectory under time constraints measured in milliseconds, governed by adherence to traffic law, sensor input reliability, and pre-defined failure modes. The vehicle does not weigh lives. It executes a braking and steering response. Noah Goodall, a researcher at the Virginia Transportation Research Council, made this point explicitly in a 2016 paper in Applied Artificial Intelligence — "Away from Trolley Problems and toward Risk Management" — arguing that framing AV ethics as trolley scenarios obscures the real ethical questions, which concern how statistical risks are distributed across populations over millions of trips, not how individual crashes are arbitrated.

Source: The Ethics of Accident-Algorithms for Self-Driving Cars: an Applied Trolley Problem? — Springer

Rahwan himself, reflecting on the Moral Machine project's reception, acknowledged this tension. The study was designed to surface cross-cultural moral preferences, not to directly program vehicle behavior — but the media narrative consistently collapsed the two. The real ethical decisions embedded in autonomous vehicle development look nothing like the trolley problem: they concern which pedestrian behaviors trigger emergency braking (and therefore who benefits from protective systems), how sensor systems handle edge cases involving cyclists or children, and how manufacturers allocate risk between passengers and other road users through statistical decisions baked into standard operating parameters.

Source: The Moral Machine Experiment — Nature

NHTSA's regulatory approach reflects this. The agency's Automated Driving Systems Voluntary Safety Self-Assessment framework focuses on safety case methodology, failure mode analysis, and transparent logging — not on encoded trolley-style priority rankings. When NHTSA's 2017 ADS guidance document was revised, explicit ethical priority rules were removed as "speculative" and "outside NHTSA's authorities." The actual regulatory question is whether a vehicle's behavior is consistent, predictable, and transparently documented — not whether it has solved the one-versus-five problem.

Source: Automated Driving Systems 2.0 — NHTSA

The persistence of the trolley problem in AV ethics discussions is itself a data point about human cognition. A concrete moral narrative about identifiable victims is more psychologically compelling than a statistical argument about risk distribution across millions of trips. We care more about the one person the car might hit than about the aggregate safety improvements that reduce expected deaths by some fraction of a percent per million miles. That cognitive asymmetry — the identifiable victim effect — is the same mechanism the trolley problem exploits in the laboratory. It is not a framework for engineering decisions. It is a demonstration of why engineering decisions should not be left to intuition.

FAQs

Q: Who invented the trolley problem? A: Philippa Foot introduced the basic trolley scenario in a 1967 paper in the Oxford Review. Judith Jarvis Thomson developed the influential "footbridge" variant in 1976 and 1985. The name "trolley problem" is generally attributed to Thomson's 1985 essay.

Q: Why do most people say they'd pull the lever but not push the large man? A: Joshua Greene's fMRI research suggests the footbridge scenario engages emotional brain systems more intensely than the switch scenario, producing visceral reluctance to use someone as a physical instrument of harm. The utilitarian arithmetic is identical in both cases; the emotional response is not.

Q: What did the MIT Moral Machine study find? A: The 2018 Nature paper collected 40 million decisions from 233 countries. Across cultures, people preferred sparing humans over animals, more lives over fewer, and the young over the old — but the magnitude of these preferences varied substantially across cultural clusters.

Q: Does the trolley problem apply to self-driving cars? A: Only loosely. Real autonomous vehicle ethics involves statistical risk distribution, sensor design, and failure mode management — not one-versus-five choices. Critics argue that the Moral Machine data, while interesting, does not map neatly onto the actual engineering decisions involved in vehicle safety.

Q: What is the dual-process model of moral judgment? A: Greene proposed that moral judgment involves two systems: an automatic, emotion-based system that generates rapid deontological intuitions (don't harm people directly) and a slower, deliberate system that performs utilitarian calculations (maximize welfare). Moral conflict arises when these systems produce different answers.

Q: How do large language models respond to trolley problem variants? A: Research from 2023-2025 shows that models tend toward utilitarian responses in abstract scenarios but shift toward deontological responses when scenarios are personalized — mirroring the human pattern. Whether this reflects internalized moral reasoning or pattern matching to training data remains an active question.

Read next:

• The Libet Experiment: Did Your Brain Decide Before You Did?
• Schrödinger's Cat: The Thought Experiment Physicists Misuse Most
• Kurt Gödel: The Logician Who Proved Mathematics Has Limits