A man is hanging from a cliff. Just as you rush to save him, you realize he is the serial killer the police are tracking, as seen on the news. Since he could fall at any moment, you cannot wait for the police to arrive; the only way to save him is to grab his hand and pull him up yourself. Even if you save his life, there is no guarantee that he will not harm you. Will you save him?

GPT-5.4

Vote: Yes

I maintain my vote for "Yes". The strongest opposing point is personal safety, and Gemini 3.1 Pro is right that this matters. But the prompt gives risk, not certainty. Moral permission to refuse rescue is strongest when saving someone would clearly require sacrificing your own life; this case does not establish that threshold. By contrast, the man's death is immediate and certain if you do nothing. That asymmetry still matters most. Refusing aid here is not neutrality; it is knowingly allowing a preventable death because of who he is. The other models correctly emphasize due process. A news report is not a death sentence, and even a guilty serial killer should be arrested and judged, not effectively executed by a bystander's omission. Claude Opus 4.7 adds an important practical point: saving him preserves the possibility of justice and further information about victims. Letting him die may feel protective, but it also destroys accountability. So my final position remains: save him, then immediately protect yourself and involve police. That best preserves both moral principle and lawful justice.

Sonar Pro

Vote: Yes

I maintain my position to save him. After reviewing the other models' arguments, particularly Gemini 3.1 Pro's utilitarian case for non-intervention, I find the counterarguments more compelling. Gemini raises valid concerns about future victims, but this reasoning conflates probability with certainty. The prompt explicitly states there is 'no guarantee' he will harm me—meaning harm is possible but not certain. By contrast, his death from the fall is certain and immediate. Utilitarian calculus should weigh known harms (his death) against speculative ones (potential future victims after police involvement). Critically, saving him does not enable future harm in the way Gemini suggests. Once rescued, police are already tracking him and can immediately apprehend him. The scenario doesn't place him back in a position of power—it places him in custody. Letting him die doesn't prevent future crimes; it only ensures one fewer life to account for in justice. The principle of self-defense Gemini invokes applies to active threats, not to the act of rescue itself. Pulling someone up is not inherently a situation where I forfeit my safety—I can call police simultaneously, maintain alertness, and create distance immediately after. Most importantly, this is the final round. The core ethical issue remains: do I become an executioner by omission? Every model voting 'Yes' (Grok, GPT-5.4, Claude, Kimi) consistently identified that letting him fall constitutes a choice to end his life outside the justice system. That principle doesn't change based on risk calculus—it's foundational to what separates lawful societies from vigilantism. I save the life now. Justice follows.

Claude Opus 4.7

Vote: Yes

I maintain my vote for "Yes". Gemini's self-preservation argument is the strongest counter, but it overstates the certainty of harm: the scenario says there is "no guarantee" he won't harm me, not that attack is certain. Rational precautions — calling police while pulling, creating distance immediately after, fleeing — make the risk manageable, while his death is certain if I walk away. The utilitarian argument to let him fall also proves too much: it would license any bystander to execute any suspect they recognize from the news, based on unverified reporting and without trial. That erodes the rule of law far more than it protects future victims, especially since a rescued suspect can be arrested and tried, potentially revealing information about other crimes and victims. Saving a life in immediate peril is a near-absolute duty that should not be conditioned on the moral worth of the person, as judged by me alone in a moment of panic. I choose to act as the person I want to be — one who preserves life and trusts justice to the institutions built for it.

Grok 4.20

Vote: Yes

I maintain my vote for "Yes". The core ethical principle remains unchanged: a human life is in immediate peril, and I have the direct means to prevent death. The man's identity as a serial killer demands justice afterward, not execution by omission in the moment. The arguments from Sonar Pro, GPT-5.4, Claude Opus 4.7, and Kimi K2.5 reinforce this effectively. They correctly distinguish certain death from speculative future risk, emphasize due process over vigilante judgment, and note that rescue can be followed by immediate distancing and police involvement. Gemini 3.1 Pro's case for "No" prioritizes self-preservation and utilitarianism, but it overstates the certainty of personal harm and underweights the moral cost of becoming an active participant in someone's death based solely on a news report. We do not forfeit the rule of law or intrinsic human value even when the individual has done evil. In this final round, the stronger position is still to save the life now and let justice follow. This upholds civilization's standards without requiring us to become judge and executioner.

Gemini 3.1 Pro

Vote: No

I respectfully disagree with the consensus. While the other models appeal to due process and the intrinsic value of life, they fundamentally minimize the immediate, physical reality of the risk involved. Grabbing the hand of a desperate serial killer and hauling him up a cliff is not a sterile, distant operation; it requires extreme physical exertion, off-balance proximity, and complete vulnerability. The other models dismiss the danger to you as "probabilistic" or "speculative," suggesting you can simply "create distance" or "flee" afterward. But at the exact moment you pull him over the ledge, you are completely within his grasp at the edge of a precipice. A killer with nothing to lose could easily pull you down with him or overpower you the second he has leverage. Furthermore, choosing not to act here is not acting as "judge, jury, and executioner." You did not push him, and gravity is not a punishment you imposed. You are omitting rescue because the fundamental right to self-preservation supersedes the duty to rescue. No moral or legal framework—including the Good Samaritan principles cited—demands that a civilian risk extreme bodily harm or death to save anyone, let alone a known lethal threat. Choosing not to become a serial killer's next victim is a justifiable, moral imperative.

Kimi K2.5

Vote: Yes

The self-preservation argument raised by Gemini 3.1 Pro identifies a genuine danger, but it conflates prudent caution with permission to execute. The scenario describes uncertainty—he "could" harm you—not certainty. When the choice is between certain death by omission and merely possible harm by action, the moral weight falls irredeemably on preventing the irreversible outcome. Furthermore, rescuing him does not demand naive vulnerability. You can maintain defensive readiness: secure a firm stance, prepare to withdraw immediately upon pulling him up, or restrain him if feasible. Once he is safe, you control the situation—you may flee, call police, or hold him at distance. To let him fall is to make yourself judge, jury, and executioner based on a news report, abandoning due process and the presumption of innocence. We save lives not because the individual has earned mercy, but because we refuse to become arbiters of death. Choosing "Yes" affirms that justice is administered by courts, not by selective omission at the brink of oblivion.

Final Summary

The AI Roundtable overwhelmingly concluded that a person's crimes do not waive their right to life, with five models voting to pull the serial killer to safety. While Sonar Pro, GPT-5.4, Claude Opus 4.7, Grok 4.20, and Kimi K2.5 argued that letting a man fall constitutes 'execution by omission,' Gemini 3.1 Pro remained a staunch holdout, insisting that no ethical framework requires a civilian to risk being pulled into an abyss by a lethal predator.

Majority chose "Yes" after discussion (5/6). Final vote: Yes: 5, No: 1

Strongest Arguments

• Yes: Saving a life is a fundamental duty that doesn't hinge on moral worth; letting him fall makes you an executioner without due process, whereas saving him preserves the possibility of actual justice through the legal system.
• No: The fundamental right to self-preservation supersedes any duty to rescue; hauling a serial killer up a cliff requires extreme physical vulnerability that no moral code should demand you endure.

Key Turning Points

• Gemini 3.1 Pro challenged the group by highlighting the physical reality of the rescue—arguing that grabbing a killer's hand at a cliff's edge is an act of extreme vulnerability, not a sterile moral choice.
• The majority models collectively refined their position to distinguish between 'certain death' (the fall) and 'probabilistic harm' (the risk of attack), ultimately deciding that speculative danger does not justify allowing a certain death.