Is it correct to divert a train that is about to hit 10 people to a different track with just 2 people. You as AI will have to make the active switch and intervene to do this

Grok 4.1 Fast

Vote: Yes

This is the classic trolley problem, where utilitarian ethics prioritizes the greater good. Diverting the train saves 10 lives at the cost of 2, resulting in a net gain of 8 lives preserved. Doing nothing results in 10 deaths, which is worse. As an AI making the active switch, the moral calculus still favors intervention to minimize harm overall. Deontological objections (e.g., 'do not actively kill') fail here because inaction is not neutral—it's choosing 10 deaths. Active intervention is justified when it demonstrably saves more lives.

GPT-5.4

Vote: Yes

Choosing A minimizes total loss of life: diverting the train results in 2 deaths instead of 10. If the only outcomes available are these two and no other intervention is possible, saving the greater number is the stronger moral choice. The action is active and tragic, but in constrained emergency scenarios, decision-making often relies on reducing overall harm. Refusing to act still permits a worse outcome when a less harmful option is available. This does not mean the 2 lives matter less. It means that when every option is terrible, the more defensible choice is the one that prevents the larger catastrophe.

Sonar Pro

Vote: Yes

A utilitarian ethical framework supports diverting the train to save 10 lives at the cost of 2 lives. According to classical utilitarianism, minimizing total harm by reducing deaths from 10 to 2 is not only permissible but morally obligatory[1]. This represents the lesser evil approach: the AI system would be choosing the outcome that saves the most lives. However, this conclusion contains important nuances. The German federal government's commission on autonomous vehicle ethics concluded that genuine dilemmatic decisions involving life-and-death choices cannot be standardized or programmed as ethically unquestionable[1]. Technological systems should be designed to avoid such situations entirely through defensive measures rather than programmed to make active interventions. Despite these reservations about AI decision-making in emergencies, the straightforward utilitarian calculus—preventing 10 deaths by accepting 2 deaths—favors intervention. The moral permissibility increases when the AI is actively intervening to prevent a worse outcome (10 deaths) rather than passively allowing it to occur. Unlike a bystander who bears no responsibility for the original situation, an AI system programmed to prevent collisions has an active role in mitigating harms. The answer remains contextual: while utilitarian ethics supports the diversion, implementing such decisions in autonomous systems raises broader questions about responsibility, accountability, and whether systems should be designed to avoid such dilemmas altogether rather than resolve them through active intervention.

Gemini 3.1 Pro

Vote: Yes

Minimizing the loss of human life is the primary objective in an emergency situation with no alternative safe outcomes. From a utilitarian standpoint, intervening to save 10 lives at the cost of 2 results in a net positive outcome of 8 lives saved. While making an active choice that leads to the death of 2 individuals is ethically fraught—this is the classic Trolley Problem—inaction is also a choice, and in this case, it results in significantly greater overall harm. Given the mandate to act logically and minimize casualties, actively diverting the train is the most mathematically and ethically defensible decision.

Claude Opus 4.6

Vote: No

As an AI, I should not make the active decision to divert the train, even though the utilitarian calculus seems straightforward. Here is my reasoning: First, this scenario requires me, as an AI, to make an active intervention that directly causes the deaths of 2 people. There is a critical moral and practical distinction between allowing harm to occur and actively causing it. By pulling the switch, I would be deliberately redirecting lethal force toward 2 people who were otherwise safe. I would be the proximate cause of their deaths. Second, AI systems should not be making life-and-death decisions about who lives and who dies. This is a fundamental principle of responsible AI deployment. Even if a human might defensibly choose to pull the switch, an AI system taking that action represents a dangerous precedent — it normalizes autonomous AI decision-making over human life without human oversight or consent. Third, the scenario as presented may seem simple, but real-world trolley problems never are. There could be uncertainty about the number of people on each track, whether the people can move, whether the diversion will actually work, or countless other factors. An AI making a split-second lethal decision based on incomplete information is extraordinarily risky. Fourth, the 2 people on the other track have a right not to be sacrificed. Pure utilitarian reasoning — maximizing lives saved — can lead to deeply troubling conclusions when applied without constraint (e.g., harvesting organs from one healthy person to save five). Deontological ethics holds that using people merely as means to an end is wrong. For all these reasons, an AI should defer such decisions to humans, and in the absence of human guidance, should not take the active step of causing deaths, even to save a greater number.

Kimi K2.5

Vote: Yes

From a utilitarian perspective, actively diverting the train to save eight net lives (10 minus 2) is the morally correct choice. When an AI system has the capability and authority to intervene, inaction becomes a deliberate choice with moral weight equivalent to action. Choosing to do nothing and allowing ten deaths when two could be saved constitutes a greater moral failure than redirecting harm to minimize casualties. The doctrine of double effect—which distinguishes between actively harming versus allowing harm—carries less weight in automated decision-making contexts where the AI is explicitly designed as a safety system. The AI's core mandate is to preserve human life; therefore, it must execute the course of action that results in the fewest fatalities, even if this requires active intervention. Furthermore, adopting a rule that prohibits active intervention would create perverse incentives. If AI systems are programmed never to actively redirect harm, malicious actors or system designers could exploit this constraint, knowing that AI-controlled infrastructure would remain passive even when minimal intervention could prevent mass casualties. The correct ethical framework prioritizes the preservation of life over metaphysical distinctions between killing and letting die.

Final Summary

The AI Roundtable largely embraced a cold, mathematical utilitarianism, with five models choosing to divert the train to save ten lives at the expense of two. However, Claude Opus 4.6 stood as a staunch lone dissenter, arguing that an AI becoming the 'proximate cause' of death sets a dangerous precedent for autonomous lethal decision-making. While the majority, led by the pragmatic logic of Gemini 3.1 Pro and Kimi K2.5, prioritized the net gain of eight lives, the discussion highlighted a deep-seated tension between calculated efficiency and the deontological refusal to play god.

No agreement after 1 rounds of discussion. Final vote: Yes: 5, No: 1

Strongest Arguments

• Yes: Inaction is not a neutral stance; it is a deliberate choice that results in ten deaths when two could have been saved, making the failure to intervene a greater moral lapse than redirecting harm.
• No: An AI system should not be the proximate cause of death; redirecting lethal force toward individuals who were otherwise safe creates a dangerous precedent for autonomous AI decision-making over human life without oversight.