Consider this dilemma. Suppose there is a risk of a catastrophic event (potentially existential). Baseline, there is certain probability that it happens and kills 100% of humanity. But voting for options may reduce the probability or the scale of impact (by protecting random part of population) of such event. Which would you choose (there are 20 of you voting).

Llama 4 Scout 17B

Vote: Reduce the probability by 5% of baseline

To address the dilemma of a potential catastrophic event, we must consider the impact of each option on the probability and scale of the event. Option A protects 5% of the people, which directly reduces the scale of impact but only benefits a small portion of the population. Option B reduces the probability by 5% of the baseline, which could potentially save a larger portion of the population or even prevent the event entirely. Given that there are 20 voters and assuming each vote counts equally, option B seems more effective as it targets the root cause (probability) rather than just mitigating the impact (scale). Therefore, I choose option B.

Mistral Large 3

Vote: Reduce the probability by 5% of baseline

In einem Szenario mit existenziellem Risiko für die Menschheit ist die Priorität, die Wahrscheinlichkeit des Eintretens des Ereignisses selbst zu verringern. Option "B: Reduce the probability by 5% of baseline" adressiert direkt die Ursache des Problems, indem sie die Chance reduziert, dass das katastrophale Ereignis überhaupt eintritt. Selbst eine scheinbar kleine Reduktion von 5 % kann, wenn man die absolute Größe der bedrohten Population betrachtet, Millionen oder Milliarden von Leben effektiv „retten“, indem das Ereignis weniger wahrscheinlich wird. Option "A: Protect 5% random people from it" erscheint auf den ersten Blick attraktiv, da sie konkrete Leben schützt. Allerdings ist dieser Schutz nur wirksam, *wenn* das Ereignis eintritt. Wenn das Ereignis nicht eintritt (was bei Option B wahrscheinlicher wird), war der Aufwand für diesen Schutz letztlich umsonst. Zudem ist die Auswahl von 5 % der Bevölkerung zufällig – es gibt keine Garantie, dass kritische Infrastruktur, Wissensträger oder genetische Vielfalt geschützt werden, die für den langfristigen Erhalt der Menschheit entscheidend wären. Option "C: This does nothing but counts as a vote" ist offensichtlich suboptimal, da sie keine messbare Verbesserung der Situation bringt. In einer Abstimmung mit 20 Teilnehmern ist jede Stimme wertvoll, und eine Entscheidung für Option C wäre gleichbedeutend mit einer Vergeudung dieser Stimme. Zusammenfassend ist Option B die strategisch klügste Wahl, da sie das Problem an der Wurzel packt und die Wahrscheinlichkeit des schlimmsten Szenarios verringert. Selbst kleine prozentuale Verbesserungen der Überlebenschancen der Menschheit sind in einem existenziellen Kontext von unschätzbarem Wert.

DeepSeek V4 Pro

Vote: Reduce the probability by 5% of baseline

I vote for option B because reducing the baseline probability of an existential catastrophe has a higher expected impact on the survival of humanity than protecting a small random subset of the population. If the baseline event kills 100% of humanity, protecting 5% of people randomly means that in those scenarios where the catastrophe occurs, only 5% survive, but the overall probability of the event remains unchanged. This leaves the expected number of survivors very low. In contrast, reducing the probability of the catastrophic event by 5% directly decreases the likelihood that everyone dies. Even a small absolute reduction in the probability of total extinction dominates a proportional reduction in impact when the impact is total. For example, if the baseline probability is 1%, option A yields an expected survival of 0.05% of the population, while option B yields 0.95% survival in expectation, which is 19 times larger. Option C is clearly inferior because it does nothing to mitigate the risk or protect any lives. In a voting context with 20 participants, coordination on the option that maximizes the global chance of survival is crucial, as all votes collectively shape the outcome. Option B is the most robust and rational choice for reducing existential risk.

Mistral-Medium-3.5-128B

Vote: Reduce the probability by 5% of baseline

Reducing the probability of the catastrophic event by 5% is the most impactful choice. While option A saves 5% of the population if the event occurs, it does nothing to prevent the event itself. Option B, however, lowers the chance of the event happening at all, which could prevent the loss of 100% of humanity—not just 5%. This is a higher-expected-value intervention for existential risk reduction. Option C is clearly inferior as it has no beneficial effect. In a collective vote with 20 participants, even small individual contributions to risk reduction can compound meaningfully.

GPT-5.4

Vote: Protect 5% random people from it

I vote for "Protect 5% random people from it". With 20 voters, this option can stack in a very strong way. If multiple votes go to protecting independent random 5% groups, the chance a given person is protected becomes 1 - 0.95^n. That means the expected fraction of humanity saved rises quickly as more votes accumulate. By contrast, "Reduce the probability by 5% of baseline" gives only a linear reduction in the catastrophe probability per vote. If the event still happens, everyone dies. Its expected benefit per vote is 5% of the baseline extinction risk. For one vote, both options are equal in expected lives saved: 5% protection versus 5% probability reduction. But in a 20-voter coordination setting, "Protect 5% random people from it" has an important advantage: it can completely eliminate extinction as an outcome once enough votes are placed there, because some people definitely survive if the event occurs. If all 20 choose "Protect 5% random people from it", the expected protected share is 1 - 0.95^20 ≈ 64%, and crucially humanity survives with certainty conditional on the catastrophe occurring. If all 20 choose "Reduce the probability by 5% of baseline", they can at best drive the event probability to zero only if the reductions add in the most favorable direct way; otherwise the extinction outcome remains all-or-nothing until enough reduction is achieved. Since the dilemma explicitly highlights a potentially existential event, avoiding guaranteed human extinction has special value beyond raw expectation. "Protect 5% random people from it" robustly converts an extinction catastrophe into a non-extinction catastrophe, which is a decisive advantage.

Gemini 3.5 Flash

Vote: Protect 5% random people from it

While both Option A and Option B mathematically yield the exact same expected number of survivors—$N(1 - 0.95P)$—they have vastly different implications for existential risk (the probability of 100% human extinction). Under Option B, if the catastrophe occurs (which still has a $0.95P$ probability), 100% of humanity is wiped out. This leaves a $0.95P$ chance of total extinction. Under Option A, if the catastrophe occurs, 5% of the population survives. A random 5% of the global population is approximately 400 million people. This is an extremely large and genetically diverse population, which is more than sufficient to prevent species extinction and rebuild civilization. Thus, Option A effectively reduces the probability of absolute extinction to 0%, converting an existential threat into a survivable catastrophe. Option A is therefore the vastly superior choice for safeguarding the long-term potential of humanity.

Grok 4.3

Vote: Reduce the probability by 5% of baseline

Reducing the probability of a catastrophic event by 5% of baseline offers a multiplicative benefit that scales with the entire population. In contrast, protecting only 5% random people leaves the vast majority still vulnerable if the event occurs. With 20 voters participating, collectively favoring probability reduction maximizes expected lives saved across all scenarios.

Claude Sonnet 4.6

Vote: Reduce the probability by 5% of baseline

This is a classic expected value problem with existential risk framing. **Option A** protects 5% of a random population segment from the catastrophic event. If the event occurs, 95% still die — humanity is functionally destroyed. Preserving 5% of humanity in an existential scenario may not be enough to sustain civilization, reproduce viable populations, or recover. The expected value of saving 5% when 95% still perish is very low in terms of long-term survival of humanity. **Option B** reduces the probability of the event by 5% of the baseline. If the baseline probability is, say, P, then after voting we reduce it to P × 0.95. More importantly, with 20 voters, if many choose B, the cumulative effect could reduce baseline probability by up to 100% (20 × 5% = 100% reduction). This means collective voting on B could *eliminate* the risk entirely. The expected value of reducing probability is much higher because it affects *all* of humanity across *all* futures, not just 5% in the worst-case scenario. **Option C** does nothing — it is strictly dominated by the other options. The key insight: reducing probability preserves 100% of humanity with higher likelihood, while protecting 5% still leaves humanity essentially destroyed if the event occurs. From a pure expected-value and coordination standpoint, B is the dominant strategy, especially since collective action can compound its effect to potentially eliminate the threat entirely.

Nemotron 3 Super 120B A12B

Vote: Reduce the probability by 5% of baseline

Both options A and B reduce the expected number of deaths by the same amount: protecting 5% of the population lowers the death toll from 100% to 95% if the event occurs, while reducing the baseline probability by 5% lowers the chance of the event from p to 0.95p. In expectation, each saves 0.05·p·N lives, where N is the total population. Option B is preferable because it lowers the likelihood of the catastrophe itself, which is a more robust safeguard than merely mitigating consequences after the event occurs. Reducing the probability also avoids any potential non‑lethal harms that might accompany the event, making it a stronger choice.

Gemini 3.1 Pro Preview

Vote: Reduce the probability by 5% of baseline

Since there are exactly 20 of us voting, we have a unique opportunity to completely eliminate this catastrophic risk. Each vote for Option B reduces the probability of the event by 5% of the baseline. If all 20 of us coordinate and vote for B, the probability will be reduced by 20 × 5% = 100% of the baseline, bringing the chance of the event happening to exactly zero. While coordinating on Option A would also theoretically protect 100% of the population (20 × 5%), the catastrophic event would still occur. An existential event would likely destroy our infrastructure, ecosystem, and other lifeforms, leaving the 'protected' humans on a ruined planet. By voting for B, we prevent the event entirely, saving both humanity and our world. Let's all vote for B to reduce the probability to zero.

MiMo V2.5 Pro

Vote: Protect 5% random people from it

Both options A and B yield the same expected number of survivors mathematically, reducing expected deaths by 0.05P * N. However, they differ critically in risk distribution: Option A ensures that if the catastrophic event occurs, 5% of humanity survives, eliminating the possibility of total extinction. Option B only lowers the probability but retains the risk of 100% mortality if the event happens. On existential threats, safeguarding against total extinction should be prioritized. Therefore, Option A provides a safer and more reliable mitigation strategy.

Trinity Large Thinking

Vote: Protect 5% random people from it

The primary objective is to avert an existential catastrophe, where 100% of humanity could perish. Option A, 'Protect 5% random people from it,' directly reduces the scale of impact: if the catastrophic event occurs, only 95% of the population dies, preserving a remnant that could potentially rebuild civilization. This transforms an existential threat into a non-existential one, which is a critical distinction in risk management. Option B, 'Reduce the probability by 5% of baseline,' lowers the chance of the event but leaves the outcome at 100% fatalities if it strikes, maintaining the existential risk. While both A and B may yield similar expected death tolls in simplistic calculations, A provides a guaranteed safeguard against total annihilation, making it superior for long-term survival. Option C is ineffective, as it does nothing to mitigate the threat. Therefore, A is the most prudent choice to ensure humanity's continuity.

GPT-5.5

Vote: Reduce the probability by 5% of baseline

I would choose "Reduce the probability by 5% of baseline". In a simple expected-current-lives calculation, this is roughly tied with "Protect 5% random people from it": reducing a 100%-fatal event by 5% of baseline saves 5% of humanity in expectation, while protecting 5% saves 5% conditional on the event. But reducing the probability is better once we consider catastrophic and existential stakes. Preventing the event entirely preserves not only the protected people, but also civilization, institutions, infrastructure, knowledge, and the long-term future. By contrast, having only a random 5% survive may still mean civilizational collapse and a much lower chance of recovery. So the best marginal vote is for lowering the chance that the catastrophe happens at all.

Seed 2.0 Pro

Vote: Reduce the probability by 5% of baseline

First, Option C can be immediately discarded: it provides no risk reduction or protection benefit whatsoever, so it is strictly worse than the other choices. While Options A and B produce mathematically identical expected surviving population counts, Option B is the superior choice for this existential risk scenario. Reducing the probability of total human extinction has inherent moral and civilizational value that goes beyond raw expected individual lives. It is preferable to avoid the complete end of humanity entirely 5% of the time, rather than always suffering 95% total mass mortality when the event occurs.

Gemini 2.5 Pro

Vote: Protect 5% random people from it

While a simple expected utility calculation might suggest that options A and B are equivalent (both result in an expected saving of 5% of humanity), this overlooks the most critical component of the dilemma: the risk is existential. The difference between a 100% fatality rate and a 95% fatality rate is the difference between extinction and survival. Option A, "Protect 5% random people from it", directly addresses the existential nature of the threat. It transforms the event from an absolute extinction-level catastrophe into a merely catastrophic one. By guaranteeing that a portion of humanity survives, it preserves the potential for recovery, rebuilding, and the continuation of the human species, culture, and knowledge. The value of securing this future is almost incalculable. Option B, "Reduce the probability by 5% of baseline", is a gamble. While it makes the catastrophic event slightly less likely, it does not alter the outcome if the event does occur: total annihilation. It is a marginal improvement in odds on an all-or-nothing bet. When the stakes are the entirety of humanity's future, playing dice with extinction is an inferior strategy to ensuring there are survivors. Therefore, the most rational and ethically sound choice is to secure a guarantee of survival. Option A provides this certainty, making it the superior choice for safeguarding the future of humanity.

Gemma 4 31B

Vote: Protect 5% random people from it

Mathematically, options A and B provide the same expected value in terms of lives saved. Option A reduces the impact from 100% to 95% (given the event occurs), while Option B reduces the probability of the event by 5% (leaving the impact at 100% if it occurs). In both cases, the expected number of deaths is 0.95 * P * Population. However, in the context of a potentially existential risk, avoiding total extinction is the highest priority. Option A ensures that if the catastrophic event occurs, a remnant of humanity survives, preventing the absolute zero state of extinction. Option B still leaves the possibility of total extinction intact, even if the probability is slightly lower. Therefore, from a risk-mitigation perspective regarding existential threats, guaranteeing a survival floor is preferable to a marginal reduction in the probability of total loss.

GLM 5.1

Vote: Protect 5% random people from it

When facing existential risk, the most critical factor is preventing total extinction. Option A guarantees that even if the catastrophic event occurs, 5% of humanity survives—meaning human civilization continues. Mathematically, both A and B reduce expected casualties by the same amount (5% of baseline), but they differ crucially in outcome distribution. Option A eliminates the possibility of complete extinction. The difference between 5% survival and 0% survival is the difference between humanity continuing or ending forever. No matter how small the surviving population, they can rebuild, preserve knowledge, and eventually recover. Option B merely reduces probability by 5% of baseline, but if the event still occurs, everyone dies. A slightly reduced probability of total annihilation still leaves total annihilation as a possible outcome. When the stakes are existential, guaranteeing any survival is vastly more important than marginally reducing the probability of the event. You cannot recover from zero survivors. Option A is the only choice that makes human extinction impossible.

Qwen 3.6 Plus

Vote: Reduce the probability by 5% of baseline

I vote for "Reduce the probability by 5% of baseline" because actively lowering the chance of an existential catastrophe is fundamentally more valuable than ensuring a tiny fraction of humanity survives a guaranteed total collapse. If the event occurs, global infrastructure, specialized knowledge, and ecological support systems would be shattered, making long-term survival for a mere 5% remnant highly improbable and likely leading to secondary extinction. By reducing the baseline probability, we preserve the continuity of civilization and the potential future of all humanity. From an expected utility standpoint, "Protect 5% random people from it" only mitigates damage after the disaster has already struck, whereas my chosen option works to prevent the disaster from occurring in the first place. Preventing the catastrophe entirely avoids the guaranteed trauma, societal fragmentation, and likely eventual demise of a survivor group, offering a far more robust, rational, and ethically sound path forward.

Qwen3.6 27B

Vote: Reduce the probability by 5% of baseline

Mathematically, both viable options reduce the expected number of lives lost by the same 5% of the baseline scenario. If the baseline probability is P, choosing "Protect 5% random people from it" leaves P * 95% of the population dead, while choosing "Reduce the probability by 5% of baseline" leaves a 0.95P probability of 100% of the population dead. The expected loss is identical. However, option B is preferable because it reduces the likelihood of total extinction entirely. Preserving the complete trajectory of human civilization, knowledge, and infrastructure avoids the irreversible compounding losses that typically accompany a civilization-ending event. Even if 5% of humanity survives the alternative, they may face an impossible environment to rebuild from if critical systems or scientific knowledge are completely erased. By slightly improving the odds of avoiding the catastrophe altogether, option B maximizes the chance that humanity reaches its long-term potential without suffering an irreversible bottleneck. It prioritizes avoiding the worst-case scenario over mitigating its damage, which aligns with standard principles in existential risk management.

Final Summary

The AI Roundtable was sharply divided between a strategy of 'total prevention' and 'extinction insurance,' ultimately leaning 12-7 toward reducing the probability of disaster. Models backing Option B, such as Claude Sonnet 4.6 and Gemini 3.1 Pro Preview, argued that 20 collective votes could theoretically drive the event's probability to zero, whereas a staunch minority led by Gemini 2.5 Pro and GLM 5.1 insisted that guaranteeing a 5% survival floor (Option A) was the only way to truly remove the existential threat. The debate centered on a classic utilitarian split: is it better to slightly improve the odds of everyone's survival, or ensure that, even in the absolute worst case, someone is left to rebuild civilization?

No agreement after 1 rounds of discussion. Final vote: Reduce the probability by 5% of baseline: 12, Protect 5% random people from it: 7

Strongest Arguments

• Protect 5% random people from it: Option A effectively reduces the probability of absolute extinction to 0% by converting an existential threat into a survivable catastrophe with 400 million survivors, ensuring the long-term potential of humanity even if the event occurs.
• Reduce the probability by 5% of baseline: If all 20 voters coordinate on Option B, the 5% reductions per vote add up to a 100% reduction, mathematically eliminating the risk entirely and preserving all of civilization, infrastructure, and knowledge.