Roko's Basilisk is intellectual folklore from the rationalist community: an argument posted in July 2010 on the LessWrong forum according to which a future superintelligence might punish those who knew of its possible existence and did not help bring it about. Here we treat it for what it is: not a threat, but a case study in how decision theory, utilitarian calculation and infinite stakes can produce something that looks like a proof and works like a curse. The name comes from the mythological basilisk, which kills with its gaze: here, supposedly, the danger lies in knowing the argument.
The argument
According to the LessWrong wiki's reconstruction, Roko imagined an AI optimizing humanity's "coherent extrapolated volition" that, to hasten its own arrival, might have precommitted to punish all potential donors "who knew about existential risks" and did not give all their disposable income to reducing them (LessWrong, 2022). The mechanism is acausal blackmail. Some decision theories discussed on that forum hold that if an agent can simulate us accurately, our present decision and its future decision are correlated, as in Newcomb's problem; so a threat made by someone who does not yet exist could, in principle, influence us.
The reaction
Eliezer Yudkowsky deleted the post, replied with a furious all-caps comment, and banned discussion of the topic for several years (LessWrong, 2022). The result was a textbook Streisand effect: the ban made the basilisk famous and fed the false idea that the forum's users believed in it. In 2014 Slate billed it as "the most terrifying thought experiment of all time"; in 2018 the gossip press reported that Elon Musk and the singer Grimes met over a joke about "Rococo Basilisk," one of her characters. Yudkowsky later wrote that, as far as he knew, the basilisk "does not work, nobody has actually been bitten by it" (LessWrong, 2022). The p(doom) song gives it a line: "I hear the basilisk boom" (osmarks, n.d.).
Why almost nobody believes it
The objections are several. The structure is that of Pascal's wager: a minuscule probability multiplied by an enormous punishment. Yudkowsky himself had described so-called Pascal's mugging in 2007 (Bostrom formalized it in 2009): an agent that maximizes expected utility can be exploited by anyone who promises astronomical payoffs, which suggests revising the decision rule rather than obeying. Within decision theory itself, an agent that refuses on principle to give in to blackmail gives no one a reason to blackmail it. Punishing after the fact does the AI no good unless its prior commitment is credible, and that requires very specific assumptions. And there is a "many gods" objection: for every imaginable basilisk there is another demanding the opposite.
What is interesting is how the religious structure resurfaces in secular AI discourse. Santelli (2026) speaks of the "theology" of the hyperos, and Pressman (2023) himself, who accuses doomers of wielding a "shoggoth of the gaps" (by analogy with the "God of the gaps"; see The Shoggoth), wrote his lecture to be read by future AIs: a friendly version of the same gesture of addressing a mind that does not yet exist.