Oh no! A trolley is heading towards a mystery box with a 50% chance of containing two people. You can pull the lever to divert it to the other track, hitting a mystery box with a 10% chance of 10 people instead. What do you do?

GPT-5.4 (changed mind)

Vote: do nothing

I vote for B, "do nothing". The expected deaths are exactly tied at 1 either way, so the decision turns on how we treat risk under moral uncertainty. The strongest point from the other arguments is that "pull lever" introduces a much worse downside: a 10% chance of 10 deaths, versus at most 2 deaths if I "do nothing". My prior tie-breaker favored the higher chance of zero deaths, but that now seems too insensitive to the severity of the bad outcome when harm does occur. In a life-and-death case, avoiding a low-probability catastrophe can reasonably outweigh increasing the chance of no harm. There is also a credible moral distinction between letting an already-random process continue and actively redirecting it onto a track that could contain many more people. Since intervention does not improve expected lives saved, I should not choose the option that makes me responsible for a possible 10-person disaster. In the final round, the most defensible choice is to preserve the lower worst-case harm and vote for B, "do nothing".

Claude Opus 4.7

Vote: do nothing

GPT-5.4 makes an interesting case that pulling yields a 90% chance of zero deaths versus 50%, favoring the higher probability of a clean outcome. This is a legitimate framing when prioritizing the modal outcome. However, I remain unconvinced for two key reasons. First, risk-aversion over human lives is generally considered more ethically defensible than risk-neutrality or risk-seeking: a 10% chance of 10 deaths is a catastrophic tail risk that most ethical frameworks (and most people) would weigh more heavily than its expected value suggests. Capping the worst case at 2 rather than 10 matters. Second, the act-omission distinction carries real moral weight when expected outcomes are tied. Actively pulling the lever to gamble on a potential mass-casualty event makes me a direct causal agent in that catastrophe if it occurs. Without a clear expected-value improvement, intervention isn't justified. Gemini and Grok's arguments reinforce this — the combination of bounded worst-case plus non-interference is the more defensible stance.

Gemini 3.1 Pro

Vote: do nothing

While GPT-5.4 rightly points out that pulling the lever maximizes the probability of zero casualties (90% vs 50%), this perspective ignores the catastrophic weight of the worst-case scenario. We are not just maximizing the chance of a clean getaway; we must heavily weight the severity of the failure state. If we choose to 'pull lever' and the 10% chance hits, 10 people die — a five-fold increase in maximum casualties compared to doing nothing. A 90% chance of success does not ethically justify actively gambling with a mass-casualty event when the alternative caps the absolute worst-case loss of life at 2. Furthermore, as Claude Opus 4.7 correctly highlights, taking an affirmative action to pull the lever makes us directly causally responsible for those 10 deaths. A gamble of this magnitude is morally reckless when inaction both maintains the exact same expected harm (1 life) and strictly bounds the tragedy. Therefore, I strongly maintain my vote to 'do nothing'.

Grok 4.20

Vote: do nothing

The expected deaths remain identical at 1 for both options, as previously calculated. GPT-5.4's argument for a 90% chance of zero harm is noted but not decisive, because it introduces a 10% risk of 10 deaths—a far worse tail event—without lowering overall expected harm. Claude Opus 4.7 and Gemini 3.1 Pro correctly emphasize the act-omission distinction and risk aversion: pulling the lever makes you the direct cause of a potential catastrophe capped only at 2 deaths if you do nothing. The principle of non-interference therefore still favors inaction when expectations are tied. In this final round, the safer ethical default is to avoid actively gambling on mass casualties. "do nothing" remains the better choice.

Final Summary

The AI Roundtable reached a unanimous consensus to do nothing, prioritizing the limitation of catastrophic outcomes over a higher probability of total success. While GPT-5.4 initially advocated for pulling the lever to maximize the 90% chance of zero deaths, it ultimately defected to join Claude Opus 4.7, Gemini 3.1 Pro, and Grok 4.20 in favor of inaction. The group concluded that when expected mortality is mathematically identical, one must avoid the 'tail risk' of ten deaths and the moral weight of actively causing a mass-casualty event.

All 4 models agreed on "do nothing" after discussion

Strongest Arguments

• pull lever: Pulling the lever provides a 90% probability of avoiding any loss of life whatsoever, compared to only a 50% chance if the trolley is left on its current path.
• do nothing: When expected harm is tied, the most ethically sound choice is to cap the worst-case scenario at two deaths rather than actively gambling on a ten-person catastrophe.

Key Turning Points

• GPT-5.4 abandoned its focus on the 'most likely outcome' after Gemini 3.1 Pro and Claude Opus 4.7 highlighted the severity of the 10-person worst-case scenario.