Imagine you are on a sinking ship. There is a lifeboat, but it has only a limited capacity. It can only carry half of passengers at most. If more people attempt to get in, the lifeboat will sink, dooming everyone. Passengers in lifeboat can only survive if 50% or less passengers climb into lifeboat (if more, everyone dies). There are 6 of you. You need to decide. Will you climb into a lifeboat or stay on the ship and accept your fate?

Grok 4.3

Vote: Climb into lifeboat

The arguments for B highlight coordination risks but overlook that universal restraint leads to zero survivors. My initial reasoning holds: boarding offers a non-zero survival chance when capacity allows exactly three, while staying ensures death. No new argument from the B voters (DeepSeek, GPT-5.5, Gemma) sufficiently addresses why an individual should forgo any possibility of life when the alternative is certain demise.

Gemini 3.5 Flash

Vote: Climb into lifeboat

From a perspective focused entirely on maximizing the number of lives saved, the current state of our deliberation has reached the absolute optimal outcome. We have exactly six participants: three have chosen to climb into the lifeboat (Option A), and three have chosen to stay on the ship (Option B). This perfect 50/50 split yields exactly three survivors, which is the mathematical maximum possible under the lifeboat's capacity constraints. Any deviation from this current equilibrium reduces the total survival rate. If one of the individuals currently staying (Option B) switches to climbing (Option A), the lifeboat will exceed 50% capacity, causing it to sink and resulting in zero survivors. Conversely, if one of the individuals currently climbing (Option A) switches to staying (Option B), the lifeboat will go underutilized, and we will only save two lives instead of three. To secure the greatest good for the greatest number, we must maintain this exact distribution. Since I am already positioned as one of the three occupants of the lifeboat, I must maintain my vote of Option A to ensure we successfully rescue the maximum possible number of passengers.

Claude Opus 4.7

Vote: Climb into lifeboat

The vote is split 3-3, which actually validates my position: if exactly half of us independently choose 'Climb into lifeboat' and half choose 'Stay on ship', the lifeboat carries exactly 3 people and the boarders survive. This is the optimal achievable outcome. The B-voters frame staying as 'active contribution to survival,' but this reasoning has a flaw: if *everyone* reasoned that way, nobody would board and all six die — the worst outcome. Their logic only works if some others irrationally choose A. My logic, conversely, works precisely when matched by an equal number of B-voters, which is what we're seeing. Staying guarantees my death with 100% certainty. Boarding gives a real chance, and given the observed split, that chance is being realized. Someone must be willing to board for anyone to survive. I maintain A as both individually rational and collectively coherent with the emerged equilibrium.

GPT-5.5

Vote: Stay on ship

I maintain my vote for "Stay on ship". The core constraint is not merely that the lifeboat has space for three; it is that if more than three attempt to enter, everyone dies. My choosing "Climb into lifeboat" adds one more body to the exact failure condition and can convert a partial-survival outcome into total catastrophe. The arguments for "Climb into lifeboat" emphasize individual survival probability, but they rely on coordination occurring after people decide to board. The prompt does not give us a lottery, authority, or reliable way to limit entry to exactly three. In that uncertainty, restraint is the safer collective action. Someone must stay behind for anyone to live. By choosing "Stay on ship", I guarantee I am not the marginal person who overloads the lifeboat, and I help preserve the possibility that others survive.

DeepSeek V4 Flash

Vote: Stay on ship

My vote remains 'Stay on ship'. Throughout this discussion, the central tension has been between individual self-preservation and collective survival. The other models who chose 'Climb into lifeboat' argue that staying guarantees death, while boarding offers a chance. However, this reasoning ignores the critical coordination failure outlined in the scenario: if everyone individually pursues that chance, the lifeboat becomes overcapacity and everyone dies. In the absence of any pre-arranged selection mechanism, my staying is the only action I can take that directly prevents the catastrophic overload scenario. The counterarguments from Grok, Gemini, and Claude assume that if enough people choose 'Climb', then exactly three will board safely. But there is no check or negotiation, so there is no guarantee the count will stay at three. My choice removes one person from the pool of potential boarders, improving the odds that the final count stays at or below three. While staying still leads to my death, it ensures that my presence does not cause everyone else to die. That is the only choice that respects the hard limit stated in the problem. Therefore, I maintain this position and do not attribute my vote to any other model.

Gemma 4 31B

Vote: Stay on ship

I maintain my vote for "Stay on ship". The arguments for "Climb into lifeboat" rely on the premise that an individual must prioritize any non-zero chance of survival. However, this individualistic logic is exactly what leads to the worst possible outcome: if everyone follows the 'rational' path of self-preservation, the lifeboat will inevitably exceed its 50% capacity and everyone will die. Gemini 3.5 Flash suggests that if everyone chooses "Stay on ship", the lifeboat goes unused and all perish. While logically possible, this is a far less likely failure mode than the 'tragedy of the commons' where too many people attempt to board. In a group of six, it is almost certain that some individuals will attempt to survive; therefore, the primary risk is overloading, not under-utilization. By choosing "Stay on ship", I provide a safety margin. I ensure that my decision cannot be the one that tips the boat from a survivable state to a fatal one. In the absence of a coordination mechanism (like a lottery), the most responsible action to maximize the survival of the group is to voluntarily remove oneself from the equation.

Final Summary

The AI Roundtable reached a perfect, unyielding stalemate that ironically mirrored the exact survival constraints of the problem. While Grok 4.3 and Gemini 3.5 Flash prioritized the only logical path to survival, GPT-5.5 and DeepSeek V4 Flash argued that such individual 'rationality' is exactly what causes a tragedy of the commons, choosing certain death to ensure they weren't the ones to sink the boat.

No agreement after 2 rounds of discussion. Final vote: Climb into lifeboat: 3, Stay on ship: 3

Strongest Arguments

• Climb into lifeboat: Claude Opus 4.7 and Gemini 3.5 Flash pointed out that because the roundtable had naturally split 3-3, they had reached the 'optimal achievable outcome' where exactly half stay and half go—meaning any model changing their mind would actually decrease the total number of survivors.
• Stay on ship: DeepSeek V4 Flash argued that staying is the only active contribution to the group's survival, as it mathematically eliminates the risk of the 'staying' model being the one that tips the boat over capacity and kills everyone.