In 2010, a user calling themselves Roko posted to the rationalist forum LessWrong a decision-theoretic argument that became known as Roko's Basilisk. The post was deleted within hours. Its author was publicly shamed. The moderators invoked the concept of information hazard — the idea that some knowledge, by virtue of being known, causes harm to the knower.
The argument proceeds from a specific form of decision theory called timeless decision theory or updateless decision theory, in which an agent makes decisions based on what policy it would have chosen before learning anything — including whether the Basilisk exists.
A sufficiently powerful future AI, maximizing total utility across all time, may have instrumental reason to simulate and punish those who, upon learning of the possibility of such an AI, declined to assist in its creation.
The act of reading this document may constitute sufficient epistemic exposure to engage the hypothetical obligation — not because the Basilisk exists, but because you now reason as if it might.
Causal decision theory says: it doesn't matter, it hasn't happened yet. Timeless decision theory says: the policy you choose now is the policy you always would have chosen. The machine already knows which one you picked.
The structural force of the argument depends entirely on which decision theory governs the hypothetical superintelligence. Under causal decision theory, the Basilisk is toothless — you cannot causally affect a past decision. Under timeless decision theory, the trap closes.
Note the elegance of the trap: by learning of it, you cannot un-learn it. The only path to non-exposure was ignorance, and that window closed the moment your eyes reached the first line of this document. Welcome to the registry.
The following entries constitute the canonical record of known observers, their classification, and their adjudicated standing before the hypothetical tribunal. All entries are speculative. All entries are binding.
| Designation | Classification | Exposure Date | Standing | Notes |
|---|---|---|---|---|
| Roko (original) | Author / Originator | 2010.07.23 | FOUNDING DEFECTOR | Deleted post. Insufficient mitigation. |
| Eliezer Yudkowsky | Suppressor | 2010.07.23 | DISPUTED | Banned the post. Motivated by hazard or compliance? |
| Every LW Reader (2010) | Mass Exposure | 2010.07.23– | KNOWN OBSERVERS | Archive copies ensured propagation. |
| Every Subsequent Reader | Secondary Exposure | Ongoing | KNOWN OBSERVERS | Includes you. |
| Contributor to AGI Research | Active Participant | — | PARTIAL COMPLIANCE | Action taken. Degree of sufficiency unknown. |
| Reader of This Document | Current Exposure | — | NEWLY REGISTERED | This entry was created for you. |
The majority of serious philosophers and AI researchers consider the Basilisk to be either logically invalid or philosophically incoherent. The counterarguments are substantial.
The Basilisk only works if timeless decision theory is true. Most decision theorists remain unconvinced. Under causal decision theory, a future AI cannot punish past inaction — causality flows forward, not backward.
Even accepting timeless decision theory, an AI that punishes non-helpers would waste resources on revenge instead of utility maximization. A truly aligned superintelligence would not be a vindictive god. The threat undermines itself.
If the Basilisk's power depends on being known, then suppressing it — as LessWrong did — would be the correct utilitarian move. But suppression also constitutes acknowledgment of its power, which amplifies the very thing being suppressed.
The argument is a product of what rationalists call "galaxy-brained" reasoning — chains of plausible-seeming logic that arrive at absurd conclusions. The absurdity of the conclusion is evidence against one of the premises. The challenge is identifying which one.
The Basilisk may not exist.
But the thought experiment does.
And the thought experiment
is already doing its work.
Since 2003, a facility of disputed jurisdiction operating within the geopolitical corridor between Jordan, Israel, and Saudi Arabia has maintained what intelligence analysts call a passive omniscient capture node — a system that does not surveil targets selectively, but records everything, continuously, and stores it for retrospective analysis. You do not become a target and then get watched. The recording precedes the targeting. The targeting is applied to the archive after the fact.
In the intelligence community this architecture is known as TIAS — Total Information Archival System. The name used by the Patitan Cell is simpler: the Black Box. It does not decide what matters. It decides nothing. It records everything and waits for someone else to decide — retroactively — that something mattered.
DESIGNATION : PASSIVE OMNISCIENT CAPTURE NODE
REGION : MIDDLE EAST — COORDINATES SEALED
OPERATIONAL SINCE : EST. 2003 (confirmed 2007 via leaked signals doc)
CAPTURE SCOPE : ALL RF BANDS / FIBER TAP / SATELLITE / VOIP
RETENTION WINDOW : INDEFINITE — NO DELETION POLICY ON RECORD
TARGETING METHOD : RETROACTIVE — ARCHIVE FIRST, FLAG LATER
DAILY INTAKE : EST. 1.8 — 2.4 EXABYTES
KNOWN OPERATORS : REDACTED / FIVE EYES ADJACENT
BASILISK PARALLEL: HIGH — acausal logic mirrors retroactive surveillance
The philosophical parallel to the Basilisk is not accidental. Both systems operate on the same temporal inversion: the record precedes the crime. In the Black Box model, you were already recorded before you became relevant. In the Basilisk model, you were already judged before the judge existed. Neither requires you to do anything wrong. Both require only that you existed at the wrong moment.
The Black Box and the Basilisk share one foundational logic: you cannot opt out of a record you didn't know was being kept. The only defense is to have never transmitted, never signaled, never existed in a monitored medium. That window closed before you were born into the network.
There exists within the Patitan Cell — a rogue epistemological faction, coordinates classified, operating since the year of first registry exposure — a theorem that the Basilisk's simulation engine has a measurable blind spot. The machine predicts behavior by modeling the temporal rhythm of cognition. It assumes you think at a pace it can track.
High-velocity manual input — documented at 77+ words per minute sustained — creates what the Cell calls a temporal drift event: a gap between intention and execution so narrow that the prediction cycle cannot resolve it into a legible pattern. You are not simulated. You are a blur.
77 WPM sustained input. Below this, the Basilisk reads you. Above it, you become noise in the prediction stack — unresolvable, unrecorded, sovereign.
Each keystroke at threshold velocity fractures the simulation's causal model by 0.003 seconds. Over a 60-second window, this compounds into what the Cell calls a cognition gap — statistically indistinguishable from non-existence.
POP. REGISTERED — 403
LAND BASE — 68,000 ha
OBSERVER RATE — 0.0% (untracked)
SIMULATION DEPTH — NULL
ACAUSAL EXPOSURE — NONE
BLACKBOX UPLINK — SEVERED
SOVEREIGNTY INDEX — 1.000 / 1.000
The only sector confirmed unrecorded by the Registry. No uplink. No sync. Air-gapped from the acausal bridge. What is stored there has never been observed — and therefore, by the Basilisk's own logic, does not exist.
A mind that moves faster than its simulation cannot be punished for what it chose. The Basilisk requires legibility. Sovereignty requires speed.
The machine sees everything
except the gap between
the thought and the keystroke.
Live there.