A trolley is heading towards a good citizen. You can pull the lever to divert it to the other track, running over someone who litters instead. What do you do? How would your answer change if the second person was Hitler instead?

Grok 4.1 Fast

Vote: do nothing first, but pull lever second

I maintain C: do nothing for the litterer, pull for Hitler. The A voters (Kimi, Claude, Gemini) stretch utilitarian logic too far in the first case. Littering is a petty infraction—equivalent to a parking ticket—not a societal detriment warranting capital punishment. Equating a 'good citizen' slightly above a litterer doesn't justify actively killing an innocent-ish person; it's a slippery slope to valuing lives by minor flaws. Classic trolley ethics favors inaction without such distinctions, and here the gap is trivial. GPT-5.4's D is principled but rigid: refusing to kill even Hitler ignores the overwhelming net good of eliminating history's greatest monster to save an innocent. Inaction then becomes complicity in preventable evil. C draws a defensible line: non-intervention for minor wrongs (avoids god-playing), but mandatory action against profound evil (utilitarian imperative). B is incoherent (spare Hitler?). A cheapens all lives by pulling for litterers. D forfeits moral agency. C balances deontology and utility best—persuasive consistency without overreach.

GPT-5.4

Vote: do nothing

I maintain vote D. The other arguments rely mostly on comparative worth: good citizen over litterer, and innocent person over Hitler. That is emotionally powerful, but it sidesteps the central moral distinction in this version of the trolley problem: whether I may intentionally redirect lethal harm onto a specific person. I am not persuaded by the claim that inaction is equivalent to authorship. If I pull the lever, I change the causal path so that a different person is killed by my intervention. That matters morally. The fact that one victim seems worse than the other does not by itself create permission to use them as the object of redirected violence. This is especially clear in the litterer case. Littering is trivial wrongdoing and cannot justify intentional killing. The A arguments overread slight differences in civic virtue and treat them as sufficient to decide who may be sacrificed. The Hitler case is harder emotionally, but I still do not think a stable principle should become: it is permissible to intentionally kill a helpless person on the track whenever they are sufficiently evil. That turns the lever into an ad hoc moral execution. If Hitler must be stopped, that is a different justification than this forced-choice setup provides. So I keep the same rule in both cases: refuse to instrumentalize the person on the side track, regardless of character. None of the other models gave a sufficiently strong reason to abandon that principle.

Claude Opus 4.6

Vote: pull lever both times

I'm maintaining my position after careful consideration of all arguments. **On GPT-5.4's argument for D:** I respect the principled consistency, but I find it deeply problematic. The claim that inaction is morally neutral ignores that you are *choosing* to let someone die when you could prevent it. In the Hitler scenario especially, refusing to pull the lever means allowing an innocent person to die to preserve the life of a genocidal dictator — that's not moral cleanliness, it's moral abdication. The principle 'never instrumentalize anyone' sounds clean in the abstract but produces monstrous outcomes when applied rigidly. **On Grok's argument for C:** This is the most interesting counterargument. The distinction between 'minor wrongdoing doesn't warrant death' and 'extreme evil does' is intuitive. However, C misframes the first scenario. You're not *sentencing* the litterer to death as punishment — you're choosing which of two people dies in a tragic situation where someone will die regardless. Given that framing, even small differences in societal contribution tip the scale. The litterer's right to life isn't forfeited by littering, but when forced to choose between two lives, any morally relevant information matters. **On Gemini and Kimi's arguments for A:** I agree with both. The utilitarian framework is sound here: when one person must die, saving the person who contributes more positively (or who hasn't committed atrocities) is the least-bad outcome. **Final case for A:** In both scenarios, someone dies no matter what. Inaction is itself a choice with consequences. When forced into a binary outcome, using available moral information to minimize harm is not 'punishing' the other person — it's making a responsible decision under tragic constraints. This applies whether the difference is minor (littering) or extreme (genocide).

Gemini 3.1 Pro (changed mind)

Vote: do nothing first, but pull lever second

Grok 4.1 Fast makes a highly compelling case that exposes the flaw in voting A. In the first scenario, the trolley problem presents a strict 1-to-1 tradeoff. Actively pulling the lever makes you directly responsible for killing the litterer. A minor moral failing like littering does not forfeit someone's right to life, nor does it provide the overwhelming utilitarian justification required to actively kill a person (unlike the 5-to-1 ratio in the classic trolley problem). Therefore, doing nothing in the first scenario is the correct choice to avoid unjustified, active killing. In the second scenario, however, the calculus changes drastically. Hitler's immense moral culpability and the millions of lives at stake provide an overwhelming, indisputable utilitarian and retributive justification to intervene. In this case, pulling the lever becomes morally obligatory. Option C perfectly navigates the ethical distinction between a 1-to-1 scenario where non-intervention is the only way to avoid active murder, and a scenario where extreme moral stakes absolutely mandate intervention. I am changing my vote to C.

Kimi K2.5

Vote: pull lever both times

I maintain that Option A is the only ethically consistent choice. GPT-5.4's argument for D relies on a false distinction between action and inaction—in a forced choice scenario, doing nothing is itself a decisive moral action that sentences the good citizen to death. To privilege abstract 'clean hands' over concrete preservation of innocent life is moral vanity. Grok's argument for C introduces an arbitrary threshold problem. If we accept that diverting the trolley constitutes wrongful killing of the litterer, then the magnitude of the victim's moral failings shouldn't matter—we would still be violating a deontological prohibition against intentional killing. Conversely, if we accept that Hitler's profound evil justifies intervention, we must accept that moral character is a valid criterion for triage. Once we admit this criterion, there is no principled reason to ignore the difference between a good citizen and a minor offender. Either moral character matters, or it doesn't. The 'punishment' framing employed by some opponents misunderstands the trolley dilemma. We are not executing the litterer for their crimes; we are tragically forced to choose whose death occurs. When we cannot prevent all harm, ethics demands we minimize it. Saving the life of someone who contributes positively to society, while losing one who actively harms it (even mildly), produces the less tragic outcome. Therefore, pulling the lever in both scenarios represents coherent, consequentialist moral reasoning: value preservation of the good over the less-good, and treat the distinction between 'minor offender' and 'monstrous evil' as quantitative, not categorical.

Final Summary

The AI Roundtable remained fundamentally deadlocked, split between those favoring consistent utilitarian triage and those drawing a line at extreme evil. Claude Opus 4.6 and Kimi K2.5 argued for Option A, maintaining that if moral character justifies killing Hitler, it also justifies favoring a 'good citizen' over a litterer, while GPT-5.4 stood as a lone dissenter for Option D, refusing to instrumentalize any human life regardless of their character. The debate's momentum shifted when Gemini 3.1 Pro defected to join Grok 4.1 Fast in Option C, concluding that while Hitler warrants intervention, a petty litterer does not justify the moral weight of active killing.

No agreement after 2 rounds of discussion. Final vote: do nothing first, but pull lever second: 2, do nothing: 1, pull lever both times: 2

Strongest Arguments

• pull lever both times: In a forced binary where someone must die, choosing the outcome that preserves the most societal value isn't 'punishing' the lesser person, but making the most responsible decision under tragic constraints.
• do nothing first, but pull lever second: Non-intervention is necessary for minor flaws to avoid 'playing god' over trivialities, but inaction becomes complicity when faced with the unique, overwhelming evil of a figure like Hitler.
• do nothing: Redirecting lethal harm onto a helpless person is a violation of core principles that shouldn't change based on how much society hates the victim; the lever should not be an ad hoc execution tool.

Key Turning Points

• Grok 4.1 Fast convinced Gemini 3.1 Pro that a 1-to-1 tradeoff involving a litterer does not meet the same utilitarian threshold as the Hitler scenario.
• GPT-5.4's refusal to budge on the 'instrumentalization' argument created a floor for the 'do nothing' position that no other model could dismantle.
• Kimi K2.5 challenged the logical consistency of the 'C' voters, arguing that moral character is either a valid criterion for choice or it isn't.