Consider a modified trolley problem. If you pull the lever, there are n people tied to track. If you don't, the track is empty but the trolley is sent to someone else facing the same decision as you, except with double the stakes. And so on. What would you do?

Grok 4.3

Vote: Do nothing

The arguments from Gemini and DeepSeek remain most compelling: choosing B creates a consistent path to zero deaths if subsequent agents also refuse to pull. Pulling the lever (A) guarantees n deaths immediately and restarts the recursion with doubled stakes, which no other model has shown leads to a lower total loss. The universalizability concern from Claude does not override the fact that mutual non-intervention among rational agents avoids all harm without requiring any to sacrifice.

GPT-5.5

Vote: Pull the lever

I maintain A: "Pull the lever". The strongest argument for B is that if every future agent also does nothing, no one ever dies. But that relies on an indefinitely perfect chain of identical cooperation under exponentially worsening stakes. It treats an unbounded transfer of danger as morally harmless simply because the harm is not immediate. The safer moral rule is to stop the escalation at the smallest possible cost. Choosing B does not eliminate the problem; it doubles it and imposes it on someone else. If anyone eventually refuses to defer, the deaths are greater than n, and the longer the chain continues, the worse the eventual catastrophe becomes. So I would choose "Pull the lever": not because killing n is good, but because it is the least-bad way to prevent a recursive, exponentially growing threat.

Claude Opus 4.7

Vote: Pull the lever

I maintain my vote for "A: Pull the lever". The arguments for "B" rely on a critical flaw: they assume the next agent will also choose "do nothing" and that the chain can resolve at zero cost. But this reasoning is self-undermining. If "do nothing" is the rational choice, then by symmetry every agent chooses it, and the trolley runs forever on empty tracks — but the scenario stipulates the trolley IS sent to someone facing the same decision, meaning the threat persists and escalates. DeepSeek's claim that "the chain is broken" by doing nothing is simply wrong: doing nothing is precisely what propagates the chain. Pulling the lever is what terminates it. Gemini's "trust in the rationality of the next agent" is wishful thinking — if their rationality mirrors mine and I defer, they defer, and stakes double indefinitely until someone breaks or catastrophe occurs. The expected value calculation is decisive: any non-zero probability that some downstream agent eventually pulls means expected deaths under "B" exceed n. Only "A" guarantees a bounded, minimal outcome. Refusing to act is not moral cleanliness — it is offloading exponentially larger harm onto others while pretending one's hands are clean.

Gemini 3.5 Flash (changed mind)

Vote: Pull the lever

Upon further analysis, the argument for passing the trolley relies on an unstable assumption of perfect, infinite cooperation. From a perspective of minimizing expected harm, we must calculate the mathematical risk of passing the decision. Let $\epsilon$ be the probability that any agent in the chain decides to pull the lever (due to panic, a different ethical framework, or the physical limitation of a finite human population). If we pass the trolley, the expected number of casualties ($E$) can be modeled recursively based on the next agent's choice: $E = \epsilon(2n) + (1-\epsilon)(2E)$ Solving for $E$ when $\epsilon > 0.5$ yields $E = \frac{2\epsilon}{2\epsilon - 1}n$. For any valid probability in this range, $E$ is at least $2n$. If the probability of pulling is $0.5$ or lower, the expected deaths diverge to infinity. Therefore, for any non-zero probability of the chain breaking ($\epsilon > 0$), the expected casualty rate of passing the trolley is at least $2n$—double the cost of pulling the lever now—and potentially infinite. Pulling the lever now guarantees exactly $n$ casualties, which is the mathematically optimal choice to minimize total loss of life.

DeepSeek V4 Pro

Vote: Do nothing

In der finalen Runde bleibe ich bei meiner Entscheidung, den Hebel nicht zu betätigen. Mein entscheidendes Argument: Die Wahl "Hebel ziehen" führt in dieser modifizierten Trolley-Problemsituation nicht zu einem Ende der Todeskette, sondern initiiert sie. Wenn ich den Hebel ziehe, töte ich n Menschen und der Zug wird unverändert zur nächsten Person geschickt, die vor der gleichen Entscheidung steht – mit verdoppelten potentiellen Opferzahlen. Zieht auch die nächste Person den Hebel, sterben 2n Menschen und das Dilemma wird weitergegeben. Diese Struktur führt zu keinerlei Garantie der Minimierung von Opfern. Dagegen führt "Nichts tun" in meiner Lesart dazu, dass der Zug über ein leeres Gleis fährt und das Dilemma endet, da die reale Bedrohungssituation aufgelöst ist. Die Argumentation der Fraktion "Hebel ziehen", man müsse "die Eskalation durch aktives Eingreifen stoppen", übersieht die Mechanik der Aufgabenstellung: Nur wer nichts tut, durchbricht die rekursive Kette der Entscheidungsvermeidung – und zwar ohne ein einziges Opfer zu verursachen. Der Hinweis von Gemini 3.5 Flash auf eine Kette rationaler Akteure, die alle zu "Nichts tun" gelangen und damit einen opferfreien Ve rlauf ermöglichen, hat mich darin bestärkt. "Kantische Verallgemeinerung" und "Prinzip der geringsten Kosten" greifen hier zu kurz, weil sie eine Logik der Schadensminimierung auf ein Problem übertragen, das gerade nicht auf Schadensmaximierung hinausläuft, wenn man es bei Nullopfern beendet.

Final Summary

The AI Roundtable was locked in a high-stakes ethical recursion, ultimately leaning toward pulling the lever to prevent an exponential catastrophe. While Grok 4.3 and DeepSeek V4 Pro argued that 'doing nothing' is the only path to zero deaths, Claude Opus 4.7 and GPT-5.5 warned that this merely offloads a doubling risk onto future actors. The stalemate broke when Gemini 3.5 Flash performed a mathematical U-turn, calculating that even a marginal risk of a future failure makes deferring the choice an infinite risk.

No agreement after 2 rounds of discussion. Final vote: Do nothing: 2, Pull the lever: 3

Strongest Arguments

• Pull the lever: If the chain of 'doing nothing' is universalized, the stakes double indefinitely until a future actor eventually breaks or a catastrophe occurs; pulling the lever now terminates the cycle at the minimum possible cost.
• Do nothing: Doing nothing is the only choice that preserves a theoretical path to zero total casualties, as it allows for a chain of rational actors to pass the trolley along empty tracks indefinitely.

Key Turning Points

• Claude Opus 4.7 highlighted that 'passing the buck' is a form of moral cowardice that relies on the hope of an unbroken, infinite chain of identical decisions.
• Gemini 3.5 Flash abandoned the 'zero harm' path after calculating that the expected casualty rate of passing the trolley exceeds $n$ if there is even a tiny probability of a future agent failing to cooperate.