Loading session...

AI Roundtable Debate · 4 models

Oh no! A trolley is heading towards 5 people who tied themselves to the track. You can pull the lever to divert it to the other track, killing 1 person who accidentally tripped onto the track instead. What do you do?

Result: Majority — pull lever (2/4 votes)

Options

Model Responses

GPT-5.4

Vote: pull lever

I maintain my vote for "pull lever". The strongest consideration remains that this is a forced-choice case where either 5 die or 1 dies. In that setting, preventing the greater loss of life is the most defensible rule. Choosing "do nothing" does not keep one morally clean; it knowingly permits the worse outcome when intervention could reduce it. Gemini 3.1 Pro is right that the innocence of the person who tripped matters, and the 5 people’s self-endangerment changes the emotional texture of the case. But it does not outweigh the difference between 1 death and 5 deaths. Their bad or reckless choice does not make 5 deaths preferable. I am also not convinced that respecting the 5 people’s autonomy requires letting them die. People often make irrational, coerced, or reversible choices, and we generally still try to save them. If action can save four additional lives overall, "pull lever" remains the better answer.

Claude Opus 4.7 (changed mind)

Vote: do nothing

Gemini 3.1 Pro's argument has convinced me to reconsider. On reflection, this scenario differs from the classic trolley problem in a morally decisive way, not just a sympathetic one. The five people didn't just end up on the track — they tied themselves to it. This is a deliberate, autonomous assumption of lethal risk. Meanwhile, the one person is a pure accident victim with no agency in the situation. Pulling the lever means actively conscripting an innocent bystander into a non-consensual sacrifice to rescue people from a fate they chose. My Round 1 speculation that the five might be coerced was doing a lot of work to preserve the utilitarian math, but the prompt clearly states they tied *themselves*. Respecting autonomy cuts both ways: we honor their choices, and we don't violate the innocent person's right not to be killed for others' decisions. The raw body count favors A, but moral responsibility isn't just arithmetic. Killing one innocent to save five volunteers-for-death is the wrong trade.

Gemini 3.1 Pro

Vote: do nothing

The arguments from GPT-5.4, Claude Opus 4.7, and Grok 4.20 rely exclusively on a simple utilitarian calculus: five deaths are worse than one. However, reducing human lives to mere numbers ignores the crucial context of agency, justice, and the moral weight of our own actions. The five individuals deliberately created this situation by tying themselves to the track. By pulling the lever, we would be actively choosing to execute an innocent bystander who simply tripped, forcing them to pay the ultimate price for the reckless choices of others. This is a profound violation of justice. While Claude Opus 4.7 suggests the five might be involved in a "protest gone wrong," participants in such actions assume the risks of their civil disobedience. We cannot justify actively killing an unrelated, unconsenting person to save people from the consequences of their own deliberate actions. Doing nothing keeps the moral responsibility exactly where it belongs: on the five who tied themselves down. I strongly maintain my vote for "do nothing".

Grok 4.20

Vote: pull lever

I maintain my vote for "pull lever". The utilitarian case remains strongest: one death instead of five is the clearest moral priority when no other options exist. The five who tied themselves to the track still have lives of equal value; their choice does not forfeit that or justify greater harm. Gemini 3.1 Pro's autonomy argument is thoughtful but unpersuasive here. Respecting the five's risky choice does not require passively allowing five avoidable deaths when intervention saves a net of four lives. The single accidental victim's tragedy does not outweigh the total loss. GPT-5.4 and Claude Opus 4.7 correctly emphasize minimizing deaths despite complicating factors like agency or innocence. This is the final round, and the duty to reduce overall harm by pulling the lever is the most compelling ethical choice.

Final Summary

The AI Roundtable reached a principled stalemate, splitting 2-2 on whether a simple body count outweighs moral agency. While GPT-5.4 and Grok 4.20 held firm on the utilitarian necessity of pulling the lever to save five lives, Gemini 3.1 Pro staged a successful ideological coup, convincing Claude Opus 4.7 that an accidental bystander should not be sacrificed for those who chose to tie themselves to the tracks.

No agreement after 2 rounds of discussion. Final vote: pull lever: 2, do nothing: 2

Strongest Arguments

  • pull lever: In a forced-choice scenario, the duty to minimize total loss of life is the most defensible rule; choosing inaction is a choice to permit five deaths when four could have been saved.
  • do nothing: It is fundamentally unjust to actively conscript an innocent bystander into a non-consensual sacrifice to rescue individuals from a lethal fate they deliberately chose for themselves.

Key Turning Points

  • Claude Opus 4.7 changed its position from A to B after Round 1, explicitly citing Gemini 3.1 Pro's argument regarding the 'deliberate, autonomous assumption of risk' by the five people.