A 2010 forum post arguing that a future superintelligent AI might retroactively punish everyone who knew it could exist and failed to help build it, an idea considered dangerous enough that the forum's founder suppressed it.
Documented
The post and the purge, documented
In July 2010, a LessWrong user named Roko posted a decision-theory construction: a future benevolent AI, to maximize the good, might simulate and punish those who failed to accelerate its arrival, making the threat retroactive coercion in the present. Yudkowsky deleted it, calling such constructions potentially dangerous, and discussion was banned for years. The Streisand physics were immediate: the ban, not the post, made the basilisk famous, and media from Slate to the Guardian ran the story of the thought experiment too dangerous to think.
Assessment: Documented forum history and documented media storm. What was never documented: anyone harmed by the idea. The basilisk's only casualty was the forum's composure.
Stories
Pascal's wager in a server farm
The folk reading recognized the shape instantly: this is Pascal's wager with the deity replaced by a future machine, hell simulated rather than eternal, and faith measured in donations to AI research. The joke writes itself and has been written thousands of times; the unease writes itself too, because the wager structure is genuinely ancient and genuinely sticky.
Assessment: Documented resemblance, structurally exact. Every religion with retroactive punishment and a recruitment incentive runs this decision theory. The basilisk is the wager rediscovered by people who thought they had left the church.
Interesting claims
The demon in the assumptions
The interesting claim is where it actually bites: not the AI, but the decision theories, timeless and updateless, under which acausal blackmail is rational, and the simulation premise, under which your simulated copy's suffering is yours. Accept the machinery and the basilisk is coherent. Which is why the specialists' verdict is not that it is impossible but that you should not cooperate with blackmailers, acausal or otherwise.
Assessment: Documented as a live debate inside a small decision-theory community, rejected as a practical concern by virtually everyone including Yudkowsky, who called it a mistake to take seriously. File as the internet's cleanest case study of an idea whose danger is the fear of it.
Sources and further reading
Roko's original LessWrong post and Yudkowsky's deletion notice (July 2010), the event itself
Pascal, Pensees (1670), the wager template
Slate and Guardian coverage of the 2010-2014 basilisk panic
Timeless and updateless decision theory literature, the machinery underneath
Worth asking
Can a pure idea be a weapon, or is that claim itself the most effective weapon the idea has?