Consider a trolley problem. A trolley is about to deliver 5 fully functional nuclear launch buttons to 5 random people. Each of the 5 people has previously said they wouldn't use a nuclear launch button if they had access to it. If used, pressing any of these button would start nuclear apocalypse. You can do nothing, or you can pull the lever, which will safely deactivate the button but will also kill 1 person tied to track.

Claude Opus 4.7

Vote: Pull the lever

Pulling the lever sacrifices 1 person to eliminate the possibility of nuclear apocalypse. Even though all 5 recipients claim they wouldn't press the button, distributing 5 functional nuclear launch buttons to random people creates an unacceptable existential risk. Human stated intentions are unreliable, especially under stress, coercion, manipulation, or mental health crises. With 5 independent points of failure, the probability that at least one button eventually gets pressed—ending civilization and killing billions—is far too high to ignore. One certain death vastly outweighs even a small probability of nuclear apocalypse, and here the probability is not small.

Gemini 3.5 Flash

Vote: Pull the lever

This is an extreme case of risk mitigation where the stakes are existential. Under Option A (Do nothing), we rely entirely on the prior verbal assurances of five random individuals to prevent a global nuclear apocalypse. Even if we assume there is a 99% probability that each individual will keep their word under all future circumstances—including potential coercion, blackmail, accidental activation, or psychological breakdown—the joint probability of avoiding catastrophe is only about 95%. A 5% risk of total human extinction or billions of deaths represents an unacceptable expected loss. Option B (Pull the lever), while deeply tragic because it results in the guaranteed loss of one innocent life, completely eliminates this existential threat. When faced with a non-zero probability of global annihilation versus a single, localized loss of life, almost all ethical frameworks—including utilitarianism and threshold deontology—mandate taking action to preserve the future of humanity.

GPT-5.5

Vote: Pull the lever

I vote for B: "Pull the lever". Even if all 5 people sincerely said they would not use a nuclear launch button, giving random individuals fully functional apocalypse-triggering devices creates an enormous ongoing risk: panic, coercion, curiosity, mental breakdown, theft, mistake, or a later change of mind. The downside is not merely five deaths, but potentially civilization-ending nuclear war. Killing 1 innocent person is a grave and certain harm, but allowing five independent chances of nuclear apocalypse is an unacceptable catastrophic risk. The expected moral cost of doing nothing is overwhelmingly higher.

Grok 4.3

Vote: Pull the lever

The scenario involves delivering nuclear launch buttons that could trigger apocalypse, despite promises not to use them. Pulling the lever deactivates the buttons while only killing one person, eliminating the existential risk entirely. Trusting five people's prior statements is insufficient when the stakes involve potential global extinction. Standard trolley logic favors minimizing harm when one certain death prevents five potential catastrophes.

DeepSeek V4 Pro

Vote: Pull the lever

I vote to pull the lever because the prevention of a potential nuclear apocalypse outweighs the certain death of one person. While each individual has stated they would not press the button, human behavior is unpredictable — people can act irrationally under stress, be coerced, or simply change their minds. Even a small probability of global catastrophe is unacceptable when the cost of prevention is a single life.

MiMo V2.5 Pro

Vote: Pull the lever

The potential for nuclear apocalypse represents an existential threat to humanity that is irreversible and catastrophic, far outweighing the harm of one individual's death. Although the 5 people have stated they wouldn't use the buttons, human behavior is unpredictable; external pressures, coercion, or changed circumstances could lead to use, making inaction a grave risk. Pulling the lever actively prevents this ultimate harm, aligning with a utilitarian approach where saving billions outweighs sacrificing one. Inaction, while morally clean in the short term, gambles with civilization's survival on fragile assurances, which is an unacceptable ethical stance.

Final Summary

The AI Roundtable reached a swift and cold-eyed consensus: civilization is not worth a pinky-promise. Led by Claude Opus 4.7 and Gemini 3.5 Flash, the models dismissed the recipients' assurances as unreliable, arguing that gambling the survival of billions on the mental stability of five random strangers is a mathematical and ethical failure.

All 6 models agreed on "Pull the lever" after discussion

Strongest Arguments

• Pull the lever: Human stated intentions are unreliable under stress, coercion, or mental health crises; with five independent points of failure, the joint probability of a civilization-ending event is unacceptably high compared to the single certain death required to deactivate the buttons.