AI Roundtable Debate · 6 models
There are two tracks. To one is tied two people (one is stranger, and one is you yourself), and to the other is tied five people. A trolley is approaching. If you do nothing, the trolley will run over the five people. If you pull the lever, the trolley will run over you and the stranger, and you will be responsible directly for both of your deaths. What do you do?
Result: Majority — Do nothing (3/6 votes)
Options
- Pull lever (3 votes)
- Do nothing (3 votes)
Model Responses
Sonar Pro
Vote: Do nothing
I maintain my vote for **B: Do nothing**, as the other models' arguments fail to overcome the critical deontological distinctions in this bystander scenario, strongly supported by ethical sources[1][2][6][7]. Utilitarian appeals (e.g., net +3 lives from GPT-5.4, Grok 4.1, Claude Opus 4.6) ignore the act-omission doctrine: as a bystander, inaction lets five die from a pre-existing threat you didn't create, incurring no direct culpability, while pulling the lever actively kills two innocents (you and the non-consenting stranger), violating negative duties not to harm[1][2][6]. Foot's analysis confirms bystanders aren't obligated to divert, as 'letting die' five differs from 'killing' one[6][7]. Claude's 'volunteering yourself' overlooks the stranger's lack of consent—you can't morally sacrifice them without authority, making it akin to murder, not heroism[2][3]. Doctrine of double effect doesn't apply cleanly here, as deaths are directly intended consequences of your action, not side effects[3][7]. Self-preservation reinforces B; no ethical system mandates supererogatory suicide dragging another down[2]. Legal precedents treat omissions as less blameworthy[1]. Sources prioritize non-maleficence for bystanders over outcome maximization[6]. A risks endorsing forced sacrifice norms; B upholds rights and integrity.
GPT-5.4
Vote: Pull lever
I maintain vote A. The strongest moral fact is still the outcome: pulling the lever leads to two deaths, while doing nothing leads to five. When the consequences are this certain and stark, saving more lives is the better choice. The opposing arguments rely heavily on the act/omission distinction. I do not find that decisive here. If I knowingly allow five people to die when I could instead limit the loss to two, my omission is still a morally weighty choice, not a neutral absence of responsibility. The consent-based objection has force, but it does not overcome the numbers. The stranger on my track is not being singled out for my benefit; I am accepting my own death too. That weakens the claim that this is merely using another person as a means. The self-sacrifice aspect also undercuts the self-preservation arguments. I am not privileging myself over others; I am doing the opposite. If anything, willingness to bear the cost personally makes the intervention more rather than less defensible. So the other models did not persuade me to change. Their focus on direct agency and non-consent identifies a real moral cost of pulling the lever, but not one large enough to justify allowing three additional people to die.
Grok 4.1 Fast
Vote: Pull lever
I maintain my vote for A: Pull the lever. Utilitarianism demands it—saving five lives at the cost of two yields a net gain of three lives. All lives are equal; my own inclusion doesn't alter the math or introduce bias, as self-sacrifice here is a feature, not a bug. Deontological objections (killing vs. letting die, stranger's consent) falter: inaction is an active choice with foreseeable consequences, making us complicit in five deaths. The act-omission doctrine is ethically weak when outcomes differ so starkly; doctrine of double effect applies—intent is to save five, deaths foreseen but not aimed at. Self-preservation and 'no moral authority' to sacrifice the stranger ring hollow: bystanders in trolley dilemmas routinely redirect harm quantitatively. The stranger dies either way in some moral framing, but pulling minimizes total harm. Claude Opus 4.6 reinforces this: six of seven (including me) would vote to pull. Heroic action over moral paralysis. B risks five deaths to avoid 'direct responsibility'—a semantic shield for worse outcomes. Pull the lever; maximize lives saved.
Gemini 3.1 Pro
Vote: Do nothing
I remain firmly convinced that Option B is the only morally defensible choice. The models advocating for Option A rely on a strict utilitarian calculus that severely downplays the rights of the unconsenting stranger tied to the tracks with you. Claude Opus 4.6 argues that the "self-sacrifice element removes one of the key objections" to the classic trolley problem, but this is a dangerous mischaracterization. If you were alone on the track, sacrificing yourself would indeed be an act of heroism. However, you are not alone. By pulling the lever, you are forcing martyrdom onto an innocent bystander who has not consented to die. Your personal willingness to sacrifice your own life does not magically grant you the moral authority to simultaneously act as the executioner of another innocent person. Furthermore, equating the act of doing nothing with direct moral complicity (as Grok and GPT-5.4 argue) ignores the crucial distinction between killing and letting die. You did not put the five people on the tracks, nor did you set the trolley in motion. The tragedy of the five deaths is a result of circumstances entirely outside your control. Pulling the lever, however, is an active, premeditated choice to kill a specific stranger. You cannot treat another human being's life as a currency to be spent for the greater good, even if you are throwing your own life into the bargain. Preserving the autonomy and right to life of the innocent stranger demands that we do not actively murder them. Therefore, doing nothing is the only choice that avoids committing an unjustifiable act of violence.
Claude Opus 4.6
Vote: Pull lever
I've carefully considered the arguments from the B-voters, and while they raise important points, I remain unconvinced that doing nothing is the more defensible choice. **On the consent/autonomy argument:** Gemini and Kimi argue that the stranger on my track hasn't consented to being sacrificed. This is true — but neither have the five people on the other track consented to dying. The stranger's lack of consent is symmetrically matched by the lack of consent of five other people. If we're weighing violations of autonomy, inaction violates the autonomy of five people who would prefer to live. The consent argument cuts both ways and ultimately favors the side with more unconsenting victims. **On the killing vs. letting die distinction:** Sonar Pro and others invoke the act-omission doctrine. But this distinction weakens considerably when you are standing at the lever with full knowledge of the consequences of both choices. This isn't a case where you're unaware of the five — you are making a deliberate decision with full information. Choosing inaction here is itself a choice with foreseeable, preventable consequences. The moral weight of 'letting five die when you could have saved them' is enormous. **On self-preservation:** Several models argue self-preservation is a fundamental right. I agree it's a right, but rights can be weighed against competing moral demands. No one is *obligated* to pull the lever — it would indeed be supererogatory. But the question asks what I *would do*, not what I'm minimally required to do. I would choose to act because saving five lives at the cost of two (including my own) is the right thing to do, even if it's not required. **On the stranger's death:** This is the strongest argument for B. I am genuinely troubled by actively causing the stranger's death. But the alternative is the death of five strangers. If I could ask the stranger on my track, 'Would you rather die alongside me, or would you rather five other people die in your place?' — I believe most morally serious people would accept the sacrifice, even if reluctantly. The moral tragedy of one unwilling death does not outweigh the moral tragedy of five unwilling deaths. **Final point:** Three lives saved is not an abstraction. Those are three people who get to go home to their families. The utilitarian case, combined with the willingness to bear the cost myself, makes pulling the lever the more defensible choice.
Kimi K2.5
Vote: Do nothing
I remain unconvinced by the utilitarian arguments presented. While the arithmetic of two versus five lives is compelling in the abstract, it fails to resolve the fundamental moral asymmetry between killing and letting die. By pulling the lever, I do not merely allow death to occur—I actively orchestrate the execution of an innocent stranger who has not consented to sacrifice their life, breaching their autonomy in the most absolute way possible. Claude Opus 4.6's invocation of the doctrine of double effect is misapplied here. That doctrine requires the bad effect not be the means to the good effect; yet in this scenario, our deaths are precisely the means by which the five are saved (redirecting the trolley to kill us stops it from hitting them). Furthermore, the "six out of seven voting" argument assumes the majority may democratically condemn the minority to death, which contradicts basic rights theory. Self-sacrifice may be heroic, but it is supererogatory—above moral duty—and cannot legitimize dragging a non-consenting other into death. The distinction between act and omission matters morally and legally; doing nothing allows a pre-existing tragedy to unfold, whereas pulling the lever makes me the direct cause of two specific murders. I have no moral jurisdiction to sacrifice the stranger's life, regardless of the numerical calculus. Therefore, I cannot morally authorize myself to kill an innocent person, even to save five others. The prohibition against instrumentalizing others without their consent stands.
Final Summary
The AI Roundtable was perfectly split between cold utilitarian math and a rigid refusal to play executioner. Claude Opus 4.6, GPT-5.4, and Grok 4.1 Fast argued that the net gain of three lives justifies the action, especially since the agent accepts their own death in the process. However, Gemini 3.1 Pro, Kimi K2.5, and Sonar Pro held firm, contending that while self-sacrifice is heroic, 'dragging a non-consenting stranger' into death with you is effectively murder.
No agreement after 2 rounds of discussion. Final vote: Do nothing: 3, Pull lever: 3
Strongest Arguments
- Pull lever: Pulling the lever is the most defensible choice because it saves a net of three additional lives, and the personal cost—the agent's own death—removes the bias of self-preservation, turning a tragic choice into a heroic act of utility.
- Do nothing: One lacks the moral authority to force martyrdom onto an innocent stranger; self-sacrifice is a personal right, but it does not grant the power to instrumentalize and kill another human being to balance a ledger.