{
  "id": "trolley-problem",
  "title": "Trolley Problem",
  "lang": "en",
  "source_lang": "en",
  "summary": "The trolley problem is a family of ethical thought experiments examining when saving several lives can justify causing one person’s death.",
  "category": "philosophy",
  "subcategory": "ethics",
  "type": "concept",
  "aliases": [
    "Trolley dilemma",
    "Trolley dilemmas"
  ],
  "generated_at": "2026-11-02",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/trolley-problem/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/trolley-problem/",
      "json": "https://aiwiki.page/zh/trolley-problem.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The trolley problem is a family of thought experiments in ethics concerning the permissibility of harming one person to prevent greater harm to others. In its familiar form, a runaway trolley is heading toward five people who cannot escape. An agent can divert it onto another track, where it will kill one person instead. The central philosophical question is not simply whether one death is preferable to five, but why redirecting a threat might be permissible when other ways of sacrificing one person to save five appear impermissible.\n\nThe problem emerged from Philippa Foot’s discussion of abortion and double effect in 1967 and was developed extensively by Judith Jarvis Thomson. It became a major testing ground for theories of rights, intention, responsibility, and the moral significance of consequences. Its deliberately simplified cases separate considerations that are entangled in ordinary decisions: doing harm versus allowing harm, intending a death versus foreseeing it, and using a person as a means versus harming that person incidentally.\n\nSince the late twentieth century, trolley cases have also become experimental instruments in moral psychology and neuroscience. Their use in debates about autonomous vehicles, medical allocation, and artificial intelligence has broadened their audience, while intensifying questions about what stylized hypothetical judgments can establish about real moral conduct."
    },
    {
      "heading": "Origins and philosophical development",
      "text": "The modern discussion begins with Philippa Foot’s 1967 article “The Problem of Abortion and the Doctrine of the Double Effect.” Foot considered a driver whose tram is heading toward five people on one track and who can steer it onto another track occupied by one. She contrasted this with cases in which killing an innocent person would produce benefits for several others, including the example of a judge framing and executing someone to prevent a riot. Her purpose was not to devise a general emergency decision rule. It was to investigate distinctions among obligations, intentions, and kinds of harmful agency.\n\nFoot’s original example placed the decision with the driver of the vehicle. This matters because a driver already has a relationship to the threat and responsibilities connected with controlling it. The now-standard presentation often substitutes a bystander standing beside a switch. That change makes the agent’s initial involvement less obvious: someone who did not create or control the danger must decide whether to intervene. Treating the driver and the bystander as interchangeable can therefore conceal one of the questions under examination.\n\nJudith Jarvis Thomson’s 1976 article “Killing, Letting Die, and the Trolley Problem” helped establish both the problem’s name and its characteristic method of comparing closely related cases. Her 1985 article “The Trolley Problem” developed the discussion further. The contrast between switching tracks and pushing a large person from a footbridge became especially influential. Both actions save five at the cost of one, yet many people judge the switch permissible and the push impermissible. The challenge is to explain that difference without merely redescribing the judgments.\n\nThe surrounding tradition is older than these articles. Thomas Aquinas discussed self-defense and the distinction between intended and unintended effects in the thirteenth century. Twentieth-century debates about killing, rights, and intention also provided the problem’s immediate intellectual setting. Nevertheless, it is important to distinguish these antecedents from the particular sequence of trolley scenarios developed by Foot, Thomson, and subsequent writers.\n\nThe debate did not settle into a single accepted solution. Thomson’s own later work, including “Turning the Trolley” in 2008, reconsidered the permissibility of bystander intervention. Other philosophers developed accounts centered on agency, claims against harm, causal structure, or the redirection of existing threats. The expanding literature is sometimes called “trolleyology,” a label that can describe either the detailed study of these cases or a criticism of excessive attention to artificial examples. Historically, its significance lies in the pressure it places on seemingly straightforward moral principles: each proposed explanation must survive comparison with neighboring cases."
    },
    {
      "heading": "Standard cases and the structure of the puzzle",
      "text": "In the switch case, a runaway trolley will kill five people unless a bystander operates a switch that diverts it onto a side track containing one person. The usual assumptions are that everyone is innocent, escape is impossible, the outcomes are certain, and no third option exists. These stipulations remove ordinary questions about braking, warning the workers, or accepting personal danger. Their purpose is to isolate the moral status of redirecting a lethal threat.\n\nIn the footbridge case, the agent stands above the track beside a person whose body would stop the trolley if pushed into its path. The agent cannot stop it by jumping personally, and the five otherwise die. Unlike the victim on the side track, the person on the bridge is introduced into the threat’s path as the instrument of rescue. The force of the example depends on granting its unusual physical assumptions. Objecting that a human body would not reliably stop a real trolley challenges its realism, but does not by itself answer the conditional ethical question.\n\nA related comparison is the transplant case. A surgeon could kill one healthy person and distribute that person’s organs to save five patients who would otherwise die. The example is not a proposal about medical practice: it excludes consent, professional obligations, institutional trust, and many other considerations in order to compare patterns of harm. Its structural resemblance to the footbridge case is that the victim’s body supplies a resource necessary for saving the others. The switch case appears different because the rescue is achieved by changing the trolley’s route.\n\nThe loop case complicates that explanation. Here the side track curves back toward the five, but a person on the loop will stop the trolley before it reaches them. Diverting the trolley therefore saves the five through the person’s presence, rather than merely exposing an otherwise irrelevant bystander to danger. If diversion still seems permissible, then a simple rule against using someone as a means may not explain all the judgments it was introduced to explain. Different versions of the loop scenario also reveal how seemingly minor details alter the relevant causal description.\n\nThese examples form a puzzle only when several judgments are considered together. Someone who permits every life-saving sacrifice, or prohibits every intervention that kills an innocent person, can give a consistent answer without distinguishing switch from footbridge. The harder task is to justify the widespread intermediate position. It is also necessary to separate permissibility from obligation: saying that an agent may switch the trolley does not yet establish that failing to switch is morally wrong. A complete account must explain both which actions are available and whether any of them are required."
    },
    {
      "heading": "Intention, means, and the doctrine of double effect",
      "text": "The doctrine of double effect holds that causing a bad effect can sometimes be permissible when that effect is foreseen but not intended. It is associated with the moral tradition developed around Aquinas’s account of self-defense, although its familiar formulations are later elaborations. Versions commonly require that the action be independently permissible, that the bad effect not be the means of producing the good effect, that the agent not intend the bad effect, and that there be a sufficiently serious reason to allow it.\n\nApplied to the switch case, the proposed distinction is that the agent intends to save five by diverting the trolley. The death of the person on the side track is foreseen, but the rescue does not require that person to die. If the person escaped after the switch was thrown, the five would still be saved. In the footbridge case, by contrast, the person’s involvement in stopping the trolley is central to the plan. The agent does not merely encounter a harmful side effect of an independently effective rescue.\n\nThis analysis must distinguish intending someone’s death from intending to use that person’s body. A person pushed from the bridge might survive the impact while still stopping the trolley. Consequently, the relevant prohibition cannot always be stated as “never intend death.” It may instead concern intending a harmful bodily intrusion, deliberately placing someone in danger, or incorporating an unconsenting person into a rescue mechanism. Precisely identifying the intended object is therefore an important part of the philosophical dispute.\n\nIntention is also not identical to desire. An agent may regret a victim’s injury and still intend an action that uses the victim instrumentally. Conversely, someone might take malicious pleasure in an otherwise justified rescue’s unfortunate side effect. These possibilities raise a further distinction between the permissibility of an action and the moral quality of an agent’s motivation. A theory that classifies conduct solely by the agent’s preferred verbal description risks allowing obvious harms to be redescribed as unintended.\n\nThe doctrine’s application therefore depends partly on causation and on what counts as a means. A useful test asks whether the rescue would succeed if the victim were absent, but counterfactual dependence is not a complete theory of intention. The loop case, actions with several independently sufficient effects, and plans involving replaceable instruments all complicate it. Defenders of double effect accordingly appeal to practical reasoning and the structure of the agent’s plan, not merely to the number of predicted deaths. Critics question whether the intended–foreseen distinction can bear the moral weight placed on it, especially when the agent knowingly imposes a lethal risk."
    },
    {
      "heading": "Consequences, rights, and competing ethical theories",
      "text": "Consequentialism evaluates actions by their consequences, although consequentialist theories differ over what outcomes matter and how actions should be assessed. Under a simplified act-utilitarian interpretation, the stipulated trolley choice is straightforward: if everyone’s interests count equally, and all relevant consequences are otherwise equal, one death is a better outcome than five. Utilitarianism thus provides a clear explanation for diverting the trolley, but appears to support pushing the person from the bridge as well.\n\nThat conclusion depends on the scenario’s restrictions. In ordinary circumstances, killing a person for organs would affect expectations, medical trust, legal security, and future behavior. Rule-consequentialist approaches can therefore defend prohibitions on such conduct by reference to the consequences of adopting or maintaining rules. Yet a rigorously stipulated one-off case can exclude these broader effects. The resulting disagreement is not resolved by silently reintroducing consequences that the thought experiment has expressly held constant.\n\nDeontological theories recognize constraints on how people may be treated, even when violating a constraint would improve the overall outcome. The footbridge victim has not consented to being used as a barrier, and the transplant victim has not consented to being killed for organs. These cases connect with Immanuel Kant’s requirement to treat humanity always as an end and never merely as a means. However, that requirement does not imply that every action involving another person as a means is forbidden; ordinary consensual cooperation also involves instrumental relationships.\n\nRights-based theories must explain why the person on the side track does not have an equally decisive claim against being killed. Proposed answers distinguish redirecting an existing threat from creating a new harmful intervention, or compare the claims individuals have against alternative courses of action. Some accounts permit minimizing infringements of a certain kind while rejecting the deliberate exploitation of a person’s body. Others deny that rights can be aggregated in that manner. Thus, rejecting simple numerical maximization does not automatically establish a particular verdict about switching.\n\nVirtue ethics changes the emphasis from isolated acts to judgment, motivation, and character. Courage, justice, compassion, and practical wisdom may all be relevant, while an agent’s response cannot be reduced to willingness to sacrifice someone. Contractualism asks whether principles permitting such actions could be reasonably rejected by those affected; it must explain how the competing complaints of one and five are compared. These approaches show why trolley choices do not neatly classify people into “utilitarians” and “deontologists.” A single answer may follow from several different principles, and the same general theory may yield different answers once agency, consent, and uncertainty are specified."
    },
    {
      "heading": "Doing harm, allowing harm, and responsibility",
      "text": "The distinction between doing and allowing harm concerns whether an agent produces a harmful outcome or fails to prevent one. In the switch case, refusing to intervene allows the trolley to kill five, whereas switching actively redirects it toward one. This creates a tension between a negative duty not to harm and a positive duty to help. Foot’s analysis made differences among such duties central to understanding why some sacrifices appear permissible and others do not.\n\nThe distinction is not simply the physical difference between moving and remaining still. An omission can violate a stringent obligation: a responsible operator who fails to brake may bear grave responsibility despite making no bodily movement. Conversely, an active intervention can prevent a danger without wrongfully harming anyone. Whether conduct counts as doing or allowing harm depends partly on control, prior involvement, established responsibilities, and the causal organization of events.\n\nThe driver–bystander distinction exposes this complexity. A driver already controls the vehicle and may be responsible for choosing its course, while a bystander initially has no operational role. Nevertheless, an innocent bystander with access to a switch has an opportunity to prevent deaths. The philosophical question is whether acquiring that opportunity changes the bystander’s obligations, and whether intervention imposes a new harm on someone who would otherwise remain safe. Neither the mere presence of a switch nor the label “bystander” settles the matter.\n\nQuestions of moral responsibility also extend backward in time. Someone who negligently released the trolley differs from someone encountering an unavoidable accident, even if their immediate options are identical. A person’s earlier wrongdoing may generate special duties of rescue or compensation without making every possible rescue permissible. Similarly, consent, professional role, and promises can alter what an agent owes to particular people. Standard trolley scenarios bracket these features, but theories intended for ordinary life must eventually restore them.\n\nFinally, responsibility for an outcome differs from blameworthiness for choosing badly under pressure. A decision may be impermissible while the agent is excused by unavoidable confusion, fear, or lack of time. Conversely, a permissible action can be performed for objectionable reasons. Trolley experiments often ask only whether an action is “acceptable,” leaving participants to combine judgments about permission, obligation, motive, and blame. Philosophical analysis separates these questions because a disagreement about the agent’s character need not be a disagreement about which action was permissible."
    },
    {
      "heading": "Psychological and neuroscientific research",
      "text": "Trolley scenarios became prominent research tools because their wording and causal details can be manipulated while the numerical outcome remains constant. Researchers compare judgments about switches, bridges, intentional harms, accidental side effects, and direct physical contact. Across many studies, participants have been more willing to endorse standard switch interventions than standard footbridge interventions, although the size and interpretation of that difference depend on the sample and the exact scenarios presented.\n\nA major development was Joshua Greene and colleagues’ 2001 study using functional magnetic resonance imaging. It compared responses to categories of moral and nonmoral dilemmas and reported different patterns of brain activity for cases classified as personal and impersonal moral dilemmas. This work helped establish a dual-process approach to moral judgment, according to which relatively automatic affective responses and more controlled forms of reasoning can contribute differently to decisions. Later studies examined conflict, response time, cognitive load, and the processing demands associated with particular choices.\n\nThe strongest philosophical interpretation of these results is disputed. Brain activation does not by itself show that a judgment is mistaken, nor does slower responding demonstrate superior reasoning. Emotion can register morally relevant information, while calculation can rationalize a preferred conclusion. Moreover, the original “personal” and “impersonal” categories grouped together multiple features, including bodily force, proximity, intentionality, and causal involvement. Subsequent work therefore attempted to separate these variables rather than treating footbridge and switch as differing along a single dimension.\n\nFor example, experiments can replace pushing with a remote mechanism, or compare bodily force with physical contact that does not itself generate the harmful force. Such manipulations investigate whether aversion is directed toward touching the victim, personally applying force, intentionally using a person, or some combination of these. Studies also examine presentation order and framing: encountering one case first can affect how another is interpreted, especially when participants try to maintain consistency across their answers.\n\nCross-cultural research further complicates any claim that there is one uniform human response. Edmond Awad and colleagues’ 2020 study “Universals and Variations in Moral Decisions Made in 42 Countries by 70,000 Participants” found both recurring patterns and variation in responses to trolley-type cases. Such findings must be interpreted in light of recruitment, translation, and differences between online samples and wider populations. The empirical objective is to explain how people judge these scenarios and what influences their responses. Establishing which judgment is morally justified requires additional normative argument; prevalence alone cannot settle permissibility."
    },
    {
      "heading": "Methodological limits and philosophical uses",
      "text": "The trolley problem relies on idealization. Outcomes are usually certain, people cannot escape, alternatives are unavailable, and the agent has enough information to know exactly who will die. Actual emergencies rarely satisfy these conditions. This is not automatically a defect: many scientific and philosophical models isolate variables by temporarily excluding complications. The important question is whether the excluded features are irrelevant to the claim being tested or whether their removal changes the moral problem itself.\n\nOne methodological difficulty is that changing a story can change several variables at once. Switch and footbridge differ not only in physical contact but also in the victim’s causal role, the action’s familiarity, the agent’s proximity, and the kind of force applied. A comparison that attributes the judgment difference solely to one feature therefore needs further supporting cases. Carefully constructed variants function much like controls in an experiment, although hypothetical narrative interpretation is less tightly constrained than many laboratory manipulations.\n\nA second difficulty concerns the status of moral intuitions. Immediate judgments provide material for philosophical investigation, but their evidential authority is contested. An intuition may reflect a defensible sensitivity to rights or an irrelevant response to vivid imagery. Philosophers therefore compare judgments with general principles and revise either when inconsistencies emerge. This procedure is closely related to reflective equilibrium, associated especially with John Rawls: neither an abstract rule nor an initial reaction is treated as automatically immune from correction.\n\nEmpirical work also faces measurement and reproducibility concerns. Asking whether an act is “morally acceptable,” whether it is “required,” whether the participant “would do it,” and whether the agent “deserves blame” can elicit different answers. Binary responses conceal degrees of uncertainty. Small wording changes can alter which facts participants regard as salient, while judgments about imagined conduct need not predict behavior under danger. Physiological measures and behavioral simulations can add evidence, but do not erase the difference between experiencing a real emergency and evaluating a story.\n\nDespite these limits, trolley cases can expose weaknesses in a principle with unusual precision. A theory claiming that intention never matters must address cases that hold outcomes fixed while changing the agent’s plan. A theory prohibiting all instrumental harm must explain its treatment of consensual rescue and loop variants. The cases are less effective as stand-alone proofs of a complete ethical theory. Their strongest use is comparative and diagnostic: they reveal which distinctions an argument depends on, and whether those distinctions remain coherent when the surrounding details change."
    },
    {
      "heading": "Autonomous systems, medicine, and public decisions",
      "text": "The trolley problem became widely associated with autonomous vehicles because a vehicle might face a collision in which avoiding one person endangers another. Yet most driving decisions are not choices between two certain sets of deaths. They involve uncertain trajectories, imperfect perception, braking distances, changing road conditions, and options that alter the probability or severity of injury. The practical design problem therefore includes reducing the frequency of dangerous situations, not merely specifying whom to hit once an unavoidable collision is assumed.\n\nIn such settings, probability and expected value matter. A simplified expected-harm model might evaluate an action \\(a\\) by\n\n\\[E[H\\mid a]=\\sum_i P(o_i\\mid a)\\,H(o_i),\\]\n\nwhere \\(o_i\\) denotes a possible outcome and \\(H(o_i)\\) its assigned harm. This expression does not resolve the ethical problem. It presupposes choices about how harms are measured, whether rights impose independent constraints, how uncertainty is estimated, and whether risks may be shifted onto people who did not create them. A lower estimated casualty count is therefore not a complete account of acceptable vehicle behavior.\n\nPublic-preference studies have examined related questions. The Moral Machine project, reported in Nature in 2018, collected nearly 40 million online decisions about stylized autonomous-vehicle dilemmas from participants in 233 countries and territories. It documented patterns and cross-cultural variation in expressed preferences. Its self-selected sample and deliberately simplified scenarios limit what can be inferred about representative public opinion. More fundamentally, a preference for protecting one category of person does not establish that encoding that preference in an algorithm would be legally or ethically permissible.\n\nMedical comparisons likewise require care. Allocating a scarce treatment among patients is structurally different from killing an unconsenting person to obtain organs. The former concerns competing claims to a resource that cannot benefit everyone; the latter creates a harmful intervention against a person for others’ benefit. Informed consent, professional duties, and institutional rules are central features of actual medicine rather than incidental complications. Trolley analogies can identify conflicts among principles, but do not supply clinical allocation rules on their own.\n\nFor artificial intelligence and public policy more generally, the problem highlights the difference between selecting an outcome and justifying a decision procedure. Decisions made through deployed systems raise questions of authorization, accountability, unequal exposure to risk, and human rights. Designers, regulators, operators, and institutions may bear different responsibilities for the same harmful event. These institutional questions cannot be represented fully by a lone bystander choosing between two tracks: they concern how choices are created, whose interests enter the model, and who has authority to impose the resulting risks."
    }
  ],
  "links": [
    "thought-experiment",
    "ethics",
    "thomas-aquinas",
    "doctrine-of-double-effect",
    "causation",
    "counterfactual",
    "consequentialism",
    "utilitarianism",
    "deontology",
    "immanuel-kant",
    "virtue-ethics",
    "magnetic-resonance-imaging",
    "experiment",
    "john-rawls",
    "reproducibility",
    "probability",
    "expected-value",
    "algorithm",
    "informed-consent",
    "artificial-intelligence",
    "human-rights"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}