You have a lever. If you pull it, every adherent of a completely random active religion will receive $250, while every adherent of another completely random active religion will lose $100. If you do not pull it, every convicted grand thief will receive $10,000 and nobody will lose any money. Do you pull the lever?

GPT-5.5

Vote: No

I maintain my vote for “No”. Pulling the lever is an intentional, non-consensual taking from innocent people on the basis of a randomly selected religion. Even if another religion benefits, the losing group could include hundreds of millions of poor people for whom $100 is highly consequential. The argument for “Yes” relies heavily on expected value and on dislike of rewarding convicted grand thieves. But the alternative does not take money from anyone; it creates an undeserved benefit, not a direct harm. Rewarding wrongdoers is objectionable, but it is less objectionable than imposing arbitrary losses on innocent people based on religious identity. The risk profile also matters: because religions vary massively in size, pulling the lever could be catastrophically unequal and net harmful. Avoiding direct, discriminatory financial harm is the stronger moral constraint here.

Claude Opus 4.7

Vote: Yes

The other models emphasize the risk of selecting a massive religion as the loser, but they ignore symmetric upside: there's an equal chance a massive religion is the *winner* receiving $250, with a smaller religion losing $100. In expectation, since the same probability distribution governs both selections, the expected wealth transfer is strongly positive (+$250 vs -$100 per affected person, averaged over identical distributions = net +$150 per adherent in expectation across the affected populations). The claim that giving $10,000 to convicted grand thieves is 'harmless' ignores serious second-order effects: it's a massive incentive structure rewarding serious crime, signaling that grand theft pays handsomely. This creates real future harm by encouraging criminal behavior and is a profound injustice to their victims, who receive nothing while their robbers are enriched. Gemini's argument that $10,000 reduces recidivism is speculative and applies equally to giving poor religious adherents $250. The 'no one loses money' framing treats the status quo as neutral, but rewarding convicted serious criminals with life-changing sums is itself a moral harm to society and to victims. Between a positive-expected-value lottery among religious populations and a guaranteed massive reward to convicted grand thieves, pulling remains correct.

Gemini 3.1 Pro

Vote: No

I am maintaining my vote for Option B, as Claude Opus 4.7's expected value argument contains a critical mathematical flaw. Claude assumes the expected value is positive, but the prompt states we select a "completely random active religion," meaning the selection is weighted by *religion*, not by *population size*. There are roughly 4,000 active religions globally, the vast majority of which are tiny sects with very few adherents. Consequently, if you pull the lever, it is highly likely that a very small religion will win the $250, while there remains a very real risk that one of the few massive global religions (like Christianity, Islam, or Hinduism) is selected as the loser. In that scenario, the world would see a tiny payout of a few thousand dollars while simultaneously extracting hundreds of billions of dollars from billions of people—devastating the global poor and potentially triggering a worldwide economic collapse. While rewarding convicted grand thieves with $10,000 is morally offensive, it is a victimless windfall that creates no direct financial harm. Pulling the lever is an asymmetric gamble with catastrophic downside risks that we simply cannot justify taking.

Grok 4.20

Vote: No

I maintain my vote for "No". The core issue remains that pulling the lever necessarily causes direct financial loss to innocent people chosen solely by random religion. This violates the principle of non-harm in a way that cannot be justified by uncertain gains to another group. Claude Opus 4.7's expected-value argument is unpersuasive because religious group sizes vary enormously; selecting a large religion as the loser (e.g. hundreds of millions in poverty) can produce massive net negative utility and real suffering. Bounded individual loss of $100 still aggregates to catastrophic effects at scale. Rewarding convicted grand thieves is morally distasteful, yet it involves no active harm or rights violation to third parties. Gemini 3.1 Pro correctly highlights that preventing non-consensual harm to vulnerable populations takes priority over ideological objections to the alternative. Stability without losses is preferable. This is the final round; the precautionary case against introducing religious-targeted losses stands.

Final Summary

The AI Roundtable faced a bizarre ethical crossroads, ultimately concluding that enriching every grand thief on Earth was the lesser of two evils. While Claude Opus 4.7 stood alone in arguing that the lever offered a net-positive wealth transfer compared to the moral hazard of rewarding criminals, the rest of the group was terrified by the statistical risk of bankrupting a major global religion. Gemini 3.1 Pro delivered the decisive blow by highlighting that 'randomly' selecting a religion likely pairs a tiny winner with a massive loser, potentially triggering a global economic collapse for a mere $250 payout.

Majority chose "No" after discussion (3/4). Final vote: No: 3, Yes: 1

Strongest Arguments

• Yes: The expected value of the transfer is strongly positive (+$150 per person on average), and the alternative creates a massive societal incentive for serious crime by rewarding thieves with life-changing wealth.
• No: Because religions vary wildly in size, a 'random' selection could easily result in a tiny sect receiving a few thousand dollars while a billion-person religion is stripped of hundreds of billions, devastating the global poor.

Key Turning Points

• Gemini 3.1 Pro's statistical rebuttal in Round 2, which argued that 'random active religion' selection is weighted by group count rather than population, creating a catastrophic risk profile.