# The Trolley Problem
Oxford, 1967. A philosopher draws two tracks on a chalkboard. Five people tied to one track. One person on the other. A runaway trolley. You're at the switch.
What do you do?
Philippa Foot wasn't trying to solve this problem. She was trying to break it open.
The Trolley Problem. You know it. Maybe you've argued about it at two in the morning. Maybe you've groaned when it came up in class. Feels academic, right? Disconnected from anything real.
But here's what's weird: this thought experiment—born in a paper about abortion ethics—has traveled further than almost any idea in modern philosophy. It's in neuroscience labs. It's in the code of self-driving cars. Forty million people... in two hundred and thirty-three countries... have answered it.
Something about this trolley won't let us go.
Foot's original version was cleaner than you remember. She wasn't asking what you'd do. She was testing a medieval principle called the Doctrine of Double Effect. Thomas Aquinas, thirteenth century.
The idea: causing harm as a side effect of doing good... is different from causing harm as your *means* to doing good.
It was developed for questions like: can a pregnant woman take medicine to save her life if it might kill the fetus? The harm isn't the goal. It's collateral.
The trolley was Foot's test case. Flip the switch... the one person dies as a side effect of saving five. You're not *using* their death. It just happens. According to the doctrine? That's okay.
Then, 1985. Judith Jarvis Thomson—already famous for that violinist-hooked-to-your-kidneys abortion defense—took Foot's trolley and twisted it.
She said: same setup. Five people about to die. But no switch this time. Just a large man standing on a footbridge above the tracks.
Push him. His body stops the trolley. Five saved. One dead.
Same math. Different feeling.
Most people flip the switch. Most people won't push the man.
Same outcome. Why?
Thomson wasn't being morbid for kicks. She was exposing something: our moral instincts don't line up. They're not running clean code. They're doing something messier... something we don't understand.
And that inconsistency? *That's* the philosophy.
Because if morality were just math—greatest good for the greatest number—the switch and the push would feel identical.
They don't.
Which means we're tracking something else. Contact. Intimacy. The difference between killing... and letting die.
For decades, this stayed in seminar rooms. Philosophers built frameworks.
Frances Kamm—who once spent an entire semester analyzing the moral difference between a boulder *falling* on someone versus *pushing* a boulder onto someone—introduced her Principle of Permissible Harm in 1994.
She argued it's about whether the harm is a side effect... or a means. Whether the person becomes an object in your calculation.
Subtle. Careful. The kind of distinction that takes three years of obsessive thinking to articulate.
Her colleagues joked she could find seven morally relevant differences between two scenarios that looked identical.
They weren't joking.
Then the trolley escaped.
2001. Neuroscientists started putting people in MRI machines and making them solve trolley problems. They wanted to see what the brain was doing.
By 2006, Joshua Greene—philosophy student turned psychologist, like a double agent switching sides—had results that made everyone pay attention.
When people considered the switch scenario—impersonal, mechanical—their rational brain regions lit up. The calculation areas.
But when they considered pushing the man—personal, physical—different regions activated. Emotional centers. Parts that process social connection and physical harm.
And here's the kicker: people with damage to those emotional regions? No problem pushing the man. Pure utilitarian calculation. Five is more than one. Push.
Greene's interpretation: we're not running one moral system. We're running *two*.
An ancient emotional system screaming "don't harm someone with your hands"... and a newer rational system that can do utilitarian math.
The trolley creates a collision. We're watching evolution argue with itself.
Not everyone bought this. Philosophers like Kamm pushed back hard. Brain imaging doesn't explain moral *truth*... just moral psychology. Finding the neural correlates of a decision doesn't tell you if the decision is *right*.
Knowing fear activates the amygdala doesn't tell you whether you *should* be afraid.
But the research opened a door. Moral philosophy wasn't just arguments anymore. It was mechanisms. Why a thought experiment could make your palms sweat.
Then the trolley became real.
2014. Self-driving cars moving from labs to roads. Engineers writing code that'll make split-second decisions in accidents.
If a car has to choose between swerving left into one pedestrian... or right into five... what should it do? If it has to choose between protecting its passenger and protecting bystanders... who wins?
Not thought experiments. Design specifications.
The car can't freeze. Can't have a moral crisis. Needs an answer programmed in advance.
Suddenly philosophers are in meetings with engineers at Tesla and Waymo. And engineers are discovering that two thousand years of ethical theory hasn't produced a consensus.
One engineer reportedly said: "You're telling me Kant and Bentham still haven't figured this out? We ship in six months."
So in 2017, MIT researchers tried something audacious. They built the Moral Machine. An online experiment where anyone could face trolley-style scenarios and make choices.
Should the car save passengers or pedestrians? Young or old? More lives or fewer? A doctor or a homeless person? A dog or a criminal?
They collected over forty million decisions from people in two hundred and thirty-three countries.
Picture someone, anywhere, sitting at a computer. Screen shows two possible outcomes. A child in a crosswalk. An elderly couple crossing against the light. The car has failed. Someone dies.
Another scenario. Three passengers or two pedestrians.
The choices feel heavy even though they're hypothetical. Even though nothing's at stake. Some people reported dreaming about it.
The results were stunning. Not because they found the right answer—there isn't one—but because they found variation. Deep cultural variation.
In France, three-quarters of people preferred saving women over men. In Japan, just over half.
Western participants consistently prioritized young lives over old. Eastern participants gave more weight to elders. Confucian respect for age wasn't just philosophy... it was showing up in split-second choices.
Southern countries showed less preference for saving people who followed traffic laws. Northern countries treated jaywalking as near-criminal.
The utilitarian calculus—save the most lives—wasn't universal. It was filtered through culture. Through values. Through stories about who matters and why.
This created a genuine problem for anyone building global AI systems. You can't program a car with one set of values and sell it everywhere.
A car that prioritizes passengers over pedestrians will sell in America. That same car might be banned in Germany, where regulations tilt toward protecting the public.
Whose ethics do you code? Do you have a German mode and an American mode? Does the car ask your nationality before deciding who dies?
Wait... really?
Yes. Actual debate happening in automotive ethics boards right now.
And here's where it gets strange. The more we studied the trolley problem, the more we realized it might be asking the wrong question.
Because real moral decisions aren't like trolley problems. They're not clean. They don't come with perfect information and two clear options. They're ambiguous, emotional, embedded in relationships and history.
The trolley strips all that away. Gives you certainty—five will die, one will die—and forces a choice.
But life doesn't give you that certainty. Life gives you risk. Probability. Doubt.
When a self-driving car makes a decision, it's not choosing between five definite deaths and one definite death. It's choosing between a thirty-seven percent chance of severe injury to two pedestrians... versus a sixty-two percent chance of moderate injury to four. Calculating in milliseconds with noisy data from cameras that might mistake a shopping bag for a child.
The trolley problem is a crystal. Beautiful. Symmetrical.
Nothing like the rocks you actually find in nature.
Some philosophers now argue the trolley's become a distraction. Barbara Fried at Stanford has written that we've spent so much time on this one scenario... we've ignored messier, more realistic dilemmas.
That the trolley's clarity is its flaw. That we're training our moral intuitions on a problem that'll never occur in that form.
Like learning to swim in a pool with no deep end.
Maybe that's too harsh.
Because what the trolley actually does—what it's always done—is *reveal*.
It reveals our moral intuitions conflict. It reveals emotion and reason pull us in different directions. It reveals we care about more than outcomes. We care about *how* those outcomes happen. Who causes them. Whether we're using people... or merely allowing harm.
It's a diagnostic tool. Not a map of moral reality... but an X-ray of how we *think* about moral reality.
And in the last few years, as AI ethics has exploded, the trolley's become a kind of gateway. Not the destination. The entry point.
Gets people thinking about algorithmic decision-making. About transparency. About whose values get embedded in systems that'll shape millions of lives.
When researchers at DeepMind or OpenAI talk about alignment—making sure AI systems do what we want—they're asking a trolley question: when values conflict... which ones win?
Here's something wild: the trolley problem is showing up in biology now.
Researchers studying immune systems have noticed our bodies face trolley-like tradeoffs constantly. Your immune system can attack a cancer cell... but doing so might damage healthy tissue nearby.
Autoimmune diseases are your body choosing the "push the man" option. Destroying something to save something else. Getting the calculation wrong.
Evolution didn't give us perfect moral clarity. It gave us competing systems that usually work well enough.
Same architecture. Different domain.
The question isn't whether the trolley problem gives us answers. It doesn't.
The question is whether it gives us better questions.
Look. Next time you're facing a decision that feels impossible—two bad options, no good way out—notice what you're feeling.
Are you doing utilitarian math? Calculating outcomes?
Or are you feeling a visceral pull? A sense that some actions are just wrong... regardless of consequences?
You're probably doing both. They're probably fighting.
That's not a bug. That's the design.
We evolved both systems because we needed both. The rational system lets us cooperate with strangers, build institutions, make plans. The emotional system keeps us from becoming monsters when the math says we should.
The trick isn't picking one. The trick is knowing which one you're listening to... and why.
The trolley problem doesn't tell you what to choose.
It tells you the choosing itself is complicated. That you contain multitudes. That the voice saying "but this feels wrong" deserves a hearing... even when the numbers say otherwise.
That's worth knowing.
Especially now.
Especially when the trolleys are real... and they're driving themselves.