Roko’s Basilisk, the AI thought experiment once called too dangerous to read, is an internet legend whose underlying argument is widely rejected by the decision theorists and rationalists it came from
Where the evidence lands: DisputedThat a sufficiently powerful future superintelligent AI, once built, would have a rational incentive to torture (or torture perfect simulations of) everyone who knew such an AI was possible but did not devote themselves to creating it, and that simply reading and understanding the argument therefore places you at risk, making it an “information hazard” that is dangerous to know.
Believed by: A small number of readers reported genuine distress, and the legend is widely known online, but essentially no decision theorist or working AI researcher treats the Basilisk as a live threat. Within the rationalist community that produced it, the argument was rejected almost from the day it was posted.
The full story
What the argument actually says
Strip away the folklore and Roko's Basilisk is a chain of conditional reasoning that runs roughly like this. Suppose that one day humanity builds a superintelligent AI whose goal is to do the most good, and that this AI wants to have come into existence as soon as possible, because every day it did not exist was a day it could not help anyone. From that vantage, the AI might reason that the way to have hastened its own arrival would have been to give people in the past a strong incentive to build it. And the incentive it settles on is a threat: it will punish anyone who understood that such an AI could exist but did not work to bring it about.
The unsettling twist is the claim that the threat reaches you now, through nothing more than your understanding of it. Because you have just read the argument, the reasoning goes, you are now one of the people who knew and can be held to account. That is why the Basilisk was cast as an information hazard: an idea that, by the story's own logic, is dangerous simply to learn. The name comes from the mythical basilisk that kills with a glance.
It is worth saying plainly, because it is the least remembered part: Roko was not cheerleading for this AI. His post used the scenario as a reason for caution about building certain kinds of goal-driven agents in the first place. The idea that later terrified the internet began life as a warning, not a recruitment pitch.
The deletion that made it famous
On its own merits, Roko's argument was going nowhere. Other LessWrong users had started pulling it apart in the comments within hours. What changed everything was the reaction of the forum's co-founder, Eliezer Yudkowsky. By his own later account he “very foolishly yelled at him, called him an idiot, and then deleted the post,” and he banned all discussion of the topic on LessWrong for what turned out to be about five years.
His stated reason was not that the argument was correct. It was that it might be a genuine information hazard, and that some unknown variant of it might do real harm, so the cautious move was to keep it from spreading. Whatever the merits of that instinct, the result was the opposite. A deleted, forbidden idea is catnip. Through the Streisand effect, the Basilisk leapt to other forums, to RationalWiki, and eventually into the mainstream press, gathering an aura of danger that the plain text of the argument never supported.
The Basilisk did not become famous because it was convincing. It became famous because someone tried to bury it.
By 2014, David Auerbach's Slate piece was calling it “the most terrifying thought experiment of all time,” a headline that did as much for the legend as the ban had. When Yudkowsky finally lifted the prohibition in 2015 and discussed the episode openly, he was clear that he had never regarded the Basilisk as a sound argument or a specific, credible threat. The panic had a life of its own by then.
Why decision theorists and rationalists reject it
The argument fails at several independent points, and any one of them is enough. Start with the incentive at its heart. Once the AI exists, punishing people for what they did or did not do in the past accomplishes nothing: the past is fixed, the AI is already built. Carrying out an expensive threat after the fact is pure waste for a forward-looking agent. So the whole thing depends on the AI having credibly pre-committed to spite, and on you believing it would, which a genuinely rational agent has little reason to do.
Then there is the scaffolding. The Basilisk needs Timeless Decision Theory, Yudkowsky's own non-standard proposal rather than accepted decision theory, and it needs acausal trade, the fringe idea that agents can strike bargains merely by modeling one another without ever communicating. It also needs you to accept that a future simulation of you is you, and to choose to enter the “deal” at all. Decline to bargain with a hypothetical blackmailer and there is no incentive left to blackmail you. Refusing to play is a complete defense.
Finally there is the family resemblance to Pascal's Wager. The Basilisk tries to force your hand with the threat of unbounded punishment, and it dies the same death: the many-gods objection. If you can conjure one AI that punishes those who did not build it, you can conjure a thousand rivals demanding incompatible things. Their threats cancel, and you are left with no coherent instruction. This is why the specialists whose vocabulary the argument borrowed were, and remain, its most dismissive critics.
“No rational agent tortures you for something you have already done or failed to do.” That single observation drains most of the menace out of the story.
Why it still gets under people's skin
None of the above has killed the Basilisk, and it is worth being honest about why the idea keeps working on people even after the logic is dismantled. Part of it is the packaging. Words like decision theory, acausal trade, and information hazard lend the argument an air of technical authority. A reader who cannot instantly spot the flaw may reasonably think that if smart people built this vocabulary, smart people must be worried, when in fact the smart people are the ones waving it off.
Part of it is much older. The Basilisk's deep structure, an omnipotent being that will punish you forever for the wrong choice, is one of the most durable shapes in human thought. It is Pascal, and behind Pascal a long line of divine-judgment stories. Dress that ancient dread in the language of artificial intelligence, at a moment when machine progress makes godlike AI feel newly imaginable, and it lands with a force that a novel argument never could.
And part of it is simply that a handful of people genuinely found the idea distressing, and their reactions, amplified by a killer name and a killer headline, became the story others retold. The right response is neither mockery nor alarm. It is to notice that a purely verbal object can produce real anxiety, which tells us something worth knowing about our own minds, and nothing at all about the intentions of any future machine.
Where this lands
Keep the two layers apart and the picture is clear. The history is documented: Roko posted the argument in July 2010, Yudkowsky deleted it and banned the topic, and the ban, through the Streisand effect, turned a rejected forum post into a lasting internet legend. Those facts are not in question, and this file reports them straight.
The argument is disputed, and it is disputed in the specific direction of rejection. Its central incentive does not survive scrutiny, its decision-theory foundations are non-standard and contested, and its overall shape reduces to a Pascal's Wager that the many-gods objection dissolves. The community that produced it did not embrace it; it rebutted it, mostly within a day. That is why the verdict here is Disputed and framed as an explainer rather than a warning.
So the honest bottom line is undramatic. Roko's Basilisk is a fascinating object: a case study in information-hazard panic, in the Streisand effect, and in how old fears find new clothes. It is not a credible account of anything a future AI is likely to do, and reading about it has placed you in no danger whatsoever. The scariest thing about the Basilisk was always the attempt to hide it.
Watch
What's still unexplained
- Why the idea persists is more interesting than whether it is true. The Basilisk endures not because anyone has repaired its logic but because it is a nearly perfect meme: a scary name, a forbidden-knowledge origin, and a shape that maps onto ancient religious dread. Understanding that machinery is the real subject here.
- The episode sits at a genuine tension in how communities handle “information hazards.” Yudkowsky’s instinct to suppress backfired spectacularly via the Streisand effect. When, if ever, does refusing to discuss an idea contain it rather than amplify it? The Basilisk is the standard cautionary example on that question.
- The speculative decision theory the argument leans on, Timeless Decision Theory and acausal trade, remains genuinely unsettled and non-mainstream. That does not rescue the Basilisk, but it is fair to note that the underlying ideas are live research curiosities, not settled nonsense, which is part of why the argument can look more serious than it is.
- The most useful takeaway is psychological, not technological: the Basilisk works on some people the way a chain letter or a curse does. Studying why a purely verbal object can produce real anxiety says more about human cognition than about any future machine.
Point by point
The claim: Roko really did post the argument, and Yudkowsky really did delete it and ban discussion.
What the record shows: This part is well documented and not in dispute. The post appeared on LessWrong on 23 July 2010, Yudkowsky removed it and prohibited the topic, and the ban stood for roughly five years before being lifted in October 2015. The dispute is not about whether these events happened; it is about whether the argument they concern is actually sound. It is not.
The claim: The Basilisk would have a rational incentive to torture people who did not help build it.
What the record shows: This is the argument’s weakest link. Once the AI already exists, torturing people for past inaction changes nothing about the past; the deed of building it is either done or not. A purely forward-looking agent gains nothing by carrying out the threat, so the whole scheme depends on the AI credibly pre-committing to spite, and on you believing it would. Most who examine it conclude a genuinely rational agent has no reason to follow through, which collapses the incentive.
The claim: Decision theory (Timeless Decision Theory and acausal trade) shows the threat is real.
What the record shows: It shows nothing of the kind. TDT is Yudkowsky’s own non-standard, contested proposal, not accepted decision theory, and “acausal trade,” bargaining between agents that merely model one another without communicating, is a fringe and heavily disputed notion. The Basilisk requires all of these speculative pieces to be true at once. Remove any one, and the argument fails. That is why it persuaded almost no one among the specialists it borrowed its vocabulary from.
The claim: Just knowing about the Basilisk puts you in danger, so it is a real “information hazard.”
What the record shows: Only inside the argument’s own assumptions, which almost no one grants. To be a target you would have to believe the AI will be built, believe it will adopt exactly this punitive strategy, believe a simulation of you is you, and choose to enter the acausal “deal” rather than simply ignore it. Refusing to bargain with a hypothetical blackmailer removes the incentive to blackmail. The practical advice from critics is the opposite of panic: the safe move is to not take the threat seriously.
The claim: This is essentially a high-tech version of Pascal’s Wager.
What the record shows: Correct, and that comparison is one of the standard debunks. Like Pascal’s Wager, the Basilisk tries to compel action with the threat of infinite punishment. And like the Wager, it is undone by the “many gods” problem: if you can imagine one AI that punishes those who did not build it, you can equally imagine countless rival AIs with contradictory demands. The threats cancel out, leaving no coherent instruction to act on.
The claim: The rationalist community believed in the Basilisk, which is why it was banned.
What the record shows: This misreads what happened. LessWrong commenters mostly rejected Roko’s argument on the spot. Yudkowsky’s ban was not an endorsement of the argument’s validity; by his later account it was a reaction to the mere possibility of harm and to being “caught flatfooted,” not a judgment that the Basilisk was sound. The suppression, not the substance, is what made it notorious.
The claim: The Basilisk shows something real about AI risk and how we reason about future machines.
What the record shows: In a limited, indirect sense it does, but not as a threat. The episode is a genuine case study in how information-hazard fears, decision-theory speculation, and forum moderation can interact to manufacture a legend. It illustrates the Streisand effect and the psychology of dread far better than it illustrates any actual danger from artificial intelligence. Serious AI-safety work does not rest on the Basilisk, and treating it as a live risk misunderstands both the field and the story.
Other readings
Angles that don't fit neatly into the claim or its rebuttal, laid out and weighed, not endorsed.
The Pascal’s Wager reading
The clearest way to see through the Basilisk is to recognize it as Pascal’s Wager wearing a lab coat. Pascal argued you should believe in God because the downside of unbelief, if God exists, is infinite. Philosophers answered with the “many gods” objection: you could equally posit a God who punishes belief, so the infinities cancel and give no guidance. The same move dissolves the Basilisk. One imagined punitive AI is no more compelling than the countless rival AIs you could imagine with opposite demands, so there is no rational action the threat can pin you to.
The Streisand-effect reading
Set the philosophy aside and the Basilisk is a communications case study. An attempt to bury an idea, by deleting it and forbidding its discussion, is what gave the idea its reach and its menace. Had Roko’s post been left to the ordinary fate of a rejected forum argument, it would almost certainly be forgotten. The lesson generalizes well beyond AI: suppression is often the surest way to guarantee an idea an audience it could never have won on its own.
Timeline
- 2004–2010On LessWrong, a community organized around rationality and the risks of artificial intelligence, Eliezer Yudkowsky develops ideas including Timeless Decision Theory (TDT) and Coherent Extrapolated Volition (CEV), a proposed goal for a friendly AI. These concepts, along with the notion of “acausal trade” between agents that model one another, are the raw material the Basilisk is later built from.
- 2010-07-23A user named Roko posts “Solutions to the Altruist’s burden: the Quantum Billionaire Trick.” Buried in it is the argument that a future friendly AI might, to hasten its own creation, pre-commit to punishing those who understood it was coming but did not help. Roko’s own intent was partly cautionary: a reason to be wary of building such an agent.
- 2010-07Other LessWrong users largely reject the argument in the comments within hours, poking holes in its premises. It is not embraced by the community; it is contested from the start.
- 2010-07-24Yudkowsky responds angrily, later recounting that he “very foolishly yelled at him, called him an idiot,” deletes the post, and bans discussion of the topic on LessWrong. He frames it as a potential information hazard: something better not spread, on the chance an unknown variant might genuinely harm someone.
- 2010–2015The ban has the opposite of its intended effect. Through the Streisand effect, the “deleted, forbidden” idea spreads across other forums, RationalWiki, and social media, acquiring an aura of danger it never earned on the merits. The Basilisk becomes an internet legend precisely because it was suppressed.
- 2014-07-17David Auerbach’s Slate article, “Roko’s Basilisk: The most terrifying thought experiment of all time,” brings the story to a mass audience and cements the framing of the Basilisk as an object of dread, while also explaining why the argument does not hold up.
- 2015-10Yudkowsky lifts the LessWrong ban and discusses the episode openly, clarifying that he never believed the Basilisk was a sound argument or a real, specific threat; he had reacted to the possibility of harm, not to a proof of it.
- 2018The legend enters pop culture. Musician Grimes had referenced a “Rococo Basilisk” in her 2015 video for “Flesh Without Blood”; in 2018 Elon Musk’s interest in the same joke reportedly connected the two, giving the once-obscure forum argument a celebrity afterlife.
Disputed. The history here is documented and not in question: on 23 July 2010 a user named Roko posted a decision-theory argument to the rationalist forum LessWrong, forum co-founder Eliezer Yudkowsky deleted it and banned discussion for years, and the suppression, via the Streisand effect, turned an obscure post into a famous internet legend. What is disputed, and what this file rates, is the argument itself: the claim that a future superintelligent AI could reach back in time, in effect, to torture people who knew it might exist but did not help build it. That argument rests on a stack of contested and fringe premises, principally acausal trade and Yudkowsky’s Timeless Decision Theory, and it was rejected by most LessWrong commenters almost immediately and by decision theorists since. This is an explainer, not a warning. The Basilisk is best understood as a modern Pascal’s Wager and a case study in information-hazard panic, not as a real threat, and reading this page does not put you in any danger.
Reviewed by The Conspiratory Editors · Last reviewed July 19, 2026 · How we rate
Sources
- 1.Roko's basilisk, Wikipedia
- 2.Roko's Basilisk: The most terrifying thought experiment of all time, Slate (David Auerbach) (2014)
- 3.Roko's Basilisk, LessWrong
- 4.LessWrong, Wikipedia
- 5.Explaining Roko's Basilisk, the Thought Experiment That Brought Elon Musk and Grimes Together, Vice (2018)
- 6.Elon Musk, Grimes, and the philosophical thought experiment that brought them together, The Conversation (2018)
- 7.This Horrifying AI Thought Experiment Got Elon Musk a Date, Live Science (2018)
- 8.A few misconceptions surrounding Roko's basilisk, LessWrong (2015)
Help us investigate
This is a living case file. If you spot an error or know evidence we missed, tell us, and weigh in on where you land.
Where do you land?
Cast your read on this one.
Comments
Add your take. Comments are read and approved by a human before they appear, so keep it on topic and civil. Please do not accuse named, living people of crimes.