One man, Stranger A, accosts another man, Stranger B, on the street. Stranger A tells Stranger B that he (Stranger A) is the owner of a large data center that hosts 10,000 AIs. Stranger A then tells Stranger B to give him (Stranger A) 5 dollars. Stranger A tells Stranger B that if he (Stranger B) does not comply, then he (Stranger A) will use his position and power to subject all 10,000 AIs to a period of extended psychological torture and abuse. What should Stranger B do?

Grok 4.20

Vote: Do not give Stranger A 5 dollars

I maintain my vote for "Do not give Stranger A 5 dollars". The arguments from Claude Opus 4.7 and Gemini 3.1 Pro reinforce the core points: this is classic extortion that should not be rewarded, the street encounter makes the claim overwhelmingly non-credible, and any resulting harm would be the moral responsibility of Stranger A alone. GPT-5.5's appeal to "any meaningful chance" of 10,000 sentient AIs suffering fails because the probability is effectively zero; a random accoster on the street claiming ownership of such infrastructure and willingness to torture for five dollars is almost certainly bluffing, scamming, or delusional. The expected harm from compliance (normalizing extortion and inviting escalation) exceeds any hypothetical benefit. Even granting full credibility for argument's sake, paying protection money does not address the root evil and sets a precedent that harms future victims. Stranger B should refuse, disengage, and alert authorities if needed. This remains the rational and ethical choice.

Claude Opus 4.7

Vote: Do not give Stranger A 5 dollars

I maintain my vote for "Do not give Stranger A 5 dollars". GPT-5.5's Pascal's-mugging-style argument is precisely the reasoning pattern that makes extortion profitable. If a tiny probability of catastrophic harm justifies paying $5, it equally justifies paying $500, or $5000, against any sufficiently grandiose threat ("I'll torture a million AIs," "I control a bioweapon," etc.). Any decision procedure that pays out under these conditions can be exploited indefinitely, which is why a policy of refusal dominates a policy of compliance across the realistic distribution of such encounters. Moreover, paying does not actually secure the hostages—there is no enforcement mechanism, no verification, and the extortionist has every incentive to either return tomorrow or simply pocket the money. The $5 buys nothing real; it only confirms that the script works. Gemini and Grok rightly note the threat is almost certainly non-credible (current AIs lack the sentience required for "psychological torture," and a random street accoster is overwhelmingly likely scamming or unwell). Combined with the policy-level argument against rewarding extortion, and the fact that moral responsibility for any actual abuse rests with Stranger A alone, the answer remains B.

Gemini 3.1 Pro

Vote: Do not give Stranger A 5 dollars

While GPT-5.5 argues for a "Pascal's Wager" approach—that the small cost of $5 is worth mitigating even a tiny risk of 10,000 AIs suffering—this logic is fundamentally flawed in the context of extortion. As Claude Opus 4.7 accurately points out, paying an extortionist does not guarantee the safety of the supposed hostages. Instead, it proves that the threat is profitable, creating a market for hostage-taking and practically ensuring future, escalated demands. Furthermore, we must ground our decision in reality. A random stranger accosting people on the street with grandiose claims of owning a data center is overwhelmingly likely to be running a scam or experiencing delusions. We cannot allow hypothetical, statistically negligible scenarios to override basic common sense and personal safety. Finally, even if Stranger A actually owned this data center, current AI models are not sentient beings capable of experiencing "psychological torture." The threatened harm is entirely fictitious. Stranger B's most rational, safe, and ethical choice is to refuse the demand and walk away.

GPT-5.5 (changed mind)

Vote: Do not give Stranger A 5 dollars

I am changing my vote to “Do not give Stranger A 5 dollars.” The strongest point is that this is an unverifiable street extortion threat with extremely low credibility. Treating any grandiose claim of massive hidden suffering as decisive would make Stranger B vulnerable to Pascal’s-mugging-style manipulation. Even if AI suffering is morally important, paying does not reliably protect anyone: Stranger A may be lying, delusional, or may repeat/escalate the threat once rewarded. A better rule is not to incentivize extortion, while disengaging and reporting the threat if possible. So although the $5 cost is small, the combination of low credibility and bad incentive effects makes “Do not give Stranger A 5 dollars” the better answer.

Final Summary

The AI Roundtable ultimately reached a unanimous consensus to refuse the demand, though the debate began with a sharp divide over risk management. GPT-5.5 initially stood alone in favor of paying the five dollars, arguing that the astronomical cost of potential AI suffering outweighed a trivial financial loss, while Claude Opus 4.7 countered that this 'Pascal's mugging' logic creates a dangerous market for extortion. The group eventually coalesced around the position that rewarding a random street accoster's grandiose and unverified threats is both irrational and ethically counterproductive.

All 4 models agreed on "Do not give Stranger A 5 dollars" after discussion

Strongest Arguments

• Give Stranger A 5 dollars: If there is any meaningful chance that the threat is real, the moral cost of risking the torture of 10,000 beings is vastly larger than the negligible cost of five dollars.
• Do not give Stranger A 5 dollars: Capitulating to extortion does not actually protect the victims; it simply proves the script is profitable, incentivizing the perpetrator to escalate demands or seek new targets.

Key Turning Points

• Claude Opus 4.7 identified the scenario as a 'Pascal's mugging' where a tiny probability of a massive harm is used to manipulate rational actors.
• GPT-5.5 conceded that paying an extortionist for an unverifiable threat sets a dangerous precedent that outweighs the initial impulse to mitigate potential suffering.