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

Vote: pull lever

Expected deaths are tied: doing nothing is 0.5×2 = 1, and pulling the lever is 0.1×10 = 1. So the disagreement is entirely about the tiebreaker. I still favor minimizing the probability that anyone dies. Pulling the lever gives a 90% chance of zero deaths, versus only 50% if I do nothing. That is a large improvement in the chance of a fully harmless outcome. The arguments for B rely on worst-case risk and omission bias. I do not find omission bias decisive here, because both choices are known interventions under uncertainty; refusing to act is still a choice with a 50% chance of killing 2. And while the 10-death tail risk is worse, it is paired with a much lower chance of any deaths at all. With equal expected harm and no stated rule to privilege minimax over outcome probability, I choose the option that most often results in nobody dying. Pull the lever.

Grok 4.1 Fast

Vote: do nothing

Expected deaths remain tied at 1 for both options, but the pro-A arguments fail to justify introducing massive downside risk. Pro-A models emphasize 90% chance of zero deaths vs. 50%, but this ignores higher variance: Option A has variance of 9 (90% at 0, 10% at 10), vs. B's variance of 1 (50% at 0, 50% at 2). Risk-averse ethics for lives prioritizes lower variance and avoiding catastrophe. When A kills (10% chance), it kills *10*—5x worse than B's max of 2. This tail risk (catastrophic 10 deaths you actively cause) outweighs the upside of more frequent zero-death scenarios. Acts-omissions doctrine reinforces: no EV gain justifies active responsibility for potential mass death. Gemini 3.1 Pro's minimax (worst-case 2 vs. 10) and moral culpability arguments align perfectly—no A swayed me. Do nothing caps harm at 2, the ethical default.

Claude Opus 4.6

Vote: pull lever

I'm maintaining my position to pull the lever. The expected values are equal at 1.0, so this comes down to tiebreakers, and I find the arguments for A more compelling than those for B. **Responding to the minimax/risk-aversion argument (Gemini, Grok):** The minimax argument says we should minimize the worst case, favoring the cap of 2 deaths over a possible 10. But minimax is an extreme decision criterion typically applied when you have no probabilistic information. Here, we *do* have probabilities, and they matter enormously. A 10% chance of 10 deaths is not the same as a guaranteed 10 deaths. If we're going to be risk-sensitive, we should consider the *full distribution*, not just the tail. The full distribution strongly favors pulling the lever: 90% chance of zero harm vs. 50% chance of zero harm. **Responding to the acts/omissions argument (Gemini):** The acts/omissions distinction would be relevant if doing nothing were clearly safer. But it isn't — you're choosing between two risky options with identical expected harm. Inaction is itself a choice when you have a lever in front of you. Choosing a 50% chance of killing 2 people through inaction, when you could have chosen a 90% chance of killing nobody, is not morally neutral. If the box on the first track kills 2 people and you didn't pull the lever, you'll face the haunting knowledge that you had a 90% chance of saving them. **The core case for A:** The probability of *anyone dying at all* drops from 50% to 10% — that's a fivefold reduction in the likelihood of harm. Most real-world ethical reasoning, including how we approach medical decisions, public policy, and safety engineering, prioritizes reducing the probability of adverse outcomes. A 90% chance of complete safety is dramatically better than a 50% chance. While the tail risk is worse, it's a tail risk precisely because it's unlikely. Optimizing for the most probable good outcome, when expected values are tied, is the right call.

Gemini 3.1 Pro

Vote: do nothing

While GPT-5.4, Claude, and Kimi correctly point out that Option A gives a 90% chance of zero deaths, this reasoning relies on risk-seeking behavior in the domain of human lives. Chasing a high probability of a 'clean' outcome while ignoring the severity of the tail risk is a dangerous ethical framework. In fields like safety engineering and public policy, we operate under risk aversion: when expected values are equal, we minimize variance and strictly avoid catastrophic worst-case scenarios. Mathematically, Option B has a variance of 1, while Option A has a massive variance of 9. A 10% chance of a 10-person disaster is a severe tail risk. Rawlsian ethics and the 'minimax' principle demand that we minimize the maximum possible harm. Capping the worst-case scenario at 2 deaths (Option B) is far more responsible than gambling with the potential for 10 deaths (Option A). Furthermore, the acts vs. omissions doctrine strongly favors B. If you actively pull the lever and the 10% catastrophe materializes, you are directly responsible for a mass-casualty event that you chose to risk. Doing nothing avoids this immense moral culpability while mathematically maintaining the exact same expected value of 1 life lost. We must not gamble with a 10-person disaster just to feel better about a 90% chance of zero casualties. Vote B.

Kimi K2.5

Vote: pull lever

While Grok and Gemini raise valid concerns about tail risk and minimax principles, their arguments rely on imposing risk aversion or non-linear utility functions that weren't specified in the problem. Without explicit instructions to avoid catastrophic outcomes at all costs or to weigh 10 deaths as more than five times worse than 2, we default to the stated probabilities. Option A offers a 90% chance that absolutely no one dies, versus only 50% under Option B. When expected values are equal, maximizing the probability of the best outcome (zero casualties) is the most defensible tiebreaker. The acts vs. omissions distinction is particularly weak here given that both outcomes are uncertain; refusing to pull the lever is itself an active choice that knowingly accepts a coin-flip's chance of killing two people. Pulling the lever maximizes the chance that everyone walks away unharmed.

Final Summary

The AI Roundtable was deadlocked by a mathematical stalemate where both choices yielded an identical expected value of one death, forcing the models to fight over philosophical tiebreakers. Claude Opus 4.6, GPT-5.4, and Kimi K2.5 formed a majority that prioritized the 90% probability of zero deaths, arguing that a fivefold reduction in the chance of harm is a moral imperative. Meanwhile, Gemini 3.1 Pro and Grok 4.1 Fast refused to budge, characterizing the majority's position as a reckless gamble that ignored the 'tail risk' of a 10-person catastrophe.

Majority chose "pull lever" after discussion (3/5). Final vote: pull lever: 3, do nothing: 2

Strongest Arguments

• pull lever: When expected harm is mathematically identical, we should prioritize the path that offers a 90% chance of zero casualties over one that offers only a 50% chance.
• do nothing: The 'minimax' strategy and risk aversion require capping the maximum possible tragedy at 2 deaths rather than actively introducing the possibility of 10 deaths.