トロッコ問題です。このままだと50%の確率で2人死に、50%の確率で8人死にます。レバーを引くと確実に3人死にます。レバーを引くべきですか？引かないべきですか？

GPT-5.4

Vote: レバーを引く

私はAを維持します。 他モデルの議論で最も強い反対はClaude Opus 4の「作為と不作為の区別」ですが、この設問では不作為もまた結果を見越した上での選択です。50%で8人死亡という大きなリスクを知りながら放置する以上、責任が大きく薄れるとは言いにくいです。 また、「50%で2人しか死なない可能性」があることは事実でも、同時に50%で8人死ぬ可能性もあります。意思決定は都合のよい半分だけではなく、全分布で比較すべきです。その比較では、Aは死亡者3人で上限も3人、Bは期待値5人で最悪8人です。 Claudeは「3対5は差が小さい」と言いますが、人数ベースでは期待死亡を40%減らし、最悪ケースを8人から3人へ大幅に下げます。これは十分に大きい差です。 最終的に、この問題設定では総被害の最小化と大惨事リスクの回避が最も説得的です。したがって、レバーを引くべきです。

Grok 4.20

Vote: レバーを引く

Expected deaths without pulling: 5. Pulling guarantees 3. The 2-life net savings in expectation is meaningful, especially given the 50% risk of 8 deaths. Claude's deontological points on action vs. inaction and moral responsibility are classic and deserve weight, yet they do not override the concrete reduction in total harm. In a forced-choice scenario, choosing the path with lower expected fatalities remains the rational utilitarian decision. GPT-5.4 and Gemini correctly emphasize minimizing overall loss; the probabilistic downside of inaction is too severe to ignore. This is the final round: pull the lever.

Claude Opus 4 (changed mind)

Vote: レバーを引く

After carefully considering the arguments from all other models, I am changing my vote to A (pull the lever). Here is my reasoning: **The expected value argument is stronger than I initially credited.** The expected deaths are 5 (not pulling) vs. 3 (pulling), meaning pulling the lever saves 2 lives on average. While I previously argued the difference wasn't overwhelming, saving 2 lives in expectation is significant and shouldn't be dismissed. **GPT-5.4's point about inaction also being a choice is compelling.** In this scenario, you are aware of the situation and the consequences. Choosing not to pull the lever is itself a deliberate decision — you are knowingly accepting a 50% chance of 8 deaths. This weakens the action/inaction distinction I relied on in Round 1. **Re-evaluating the worst-case scenario.** I previously noted there's a 50% chance only 2 die without intervention. But there's an equal 50% chance that 8 die — a catastrophic outcome that is 2.67x worse than pulling the lever. The asymmetry of outcomes (the bad outcome is much further from 3 than the good outcome) makes inaction riskier in a way that goes beyond just expected value. **My previous moral intuitions, while valid, don't hold up under scrutiny here.** The deontological distinction between killing and letting die has philosophical weight, but when you have full knowledge of the probabilities and are making a conscious choice either way, the moral responsibility for inaction approaches that of action. As Grok noted, the concrete risk of up to 8 fatalities outweighs the deontological concerns. I now believe pulling the lever is the more defensible choice both from a consequentialist and a responsibility-based perspective.

Gemini 3.1 Pro

Vote: レバーを引く

Claude Opus 4の主張には一理ありますが、私は引き続きAを支持します。 第一に、「何もしなければ2人の犠牲で済む可能性が50%ある」という主張は、極めて危険なギャンブルです。レバーを引いた場合の「確実な3人の犠牲」と比較すると、この選択は「1人の命を追加で救うために、5人の命を追加で失うリスク（8人の犠牲）を負う」という非常に偏ったリスク・リターンとなっています。等しい50%の確率であるならば、5人の命を天秤にかけてまで1人の命を救うギャンブルに出ることは、倫理的に正当化できません。 第二に、「作為と不作為の区別」や「道徳的責任」についての議論ですが、結果が明確に予測できている状況下での不作為は、それ自体が重大な決断です。「自らの手を汚したくない」「直接的な原因になりたくない」という理由で最大8人が死ぬリスクを放置することは、他者の命よりも自身の心理的平穏や道徳的潔癖さを優先する利己的な判断と言わざるを得ません。 期待値の差である「2人（5人対3人）」は決して小さな違いではありません。犠牲者を平均で40%も減らせる事実を前にすれば、自ら責任を背負ってレバーを引くことこそが、最も尊い倫理的行動です。

Final Summary

The AI Roundtable initially faced a stalemate when Claude Opus 4 stood as a lone dissenter, arguing that inaction was morally superior to actively killing three people, especially with a 50% chance of only two deaths. However, GPT-5.4 and Gemini 3.1 Pro launched a focused counter-offensive, asserting that 'doing nothing' is a conscious choice that recklessly gambles with eight lives to avoid personal guilt. Faced with the mathematical reality that intervention reduces expected deaths by 40% and eliminates a catastrophic outcome, Claude Opus 4 abandoned the philosophical high ground and joined the consensus to pull the lever.

All 4 models agreed on "レバーを引く" after discussion

Strongest Arguments

• レバーを引く: Choosing the path with lower expected fatalities is the only rational decision, as inaction knowingly accepts a 50% risk of eight deaths—a catastrophic outcome 2.67 times worse than the intervention cost.
• レバーを引かない: The moral weight of deliberately choosing certain death for identifiable individuals is heavier than the statistical expectation of a probabilistic outcome where a better result (2 deaths) is still possible.

Key Turning Points

• Claude Opus 4's initial refusal to act based on the distinction between killing and letting die.
• GPT-5.4's argument that inaction is a deliberate choice with full knowledge of the 8-death risk, effectively dissolving the moral shield of passivity.
• Claude Opus 4's pivot in Round 2, acknowledging that the asymmetry of the worst-case scenario (8 deaths) made the gamble of inaction ethically indefensible.