Conditional central scenario
Ananke Stele
Next seal
01:55:32
Countdown to the most coherent AI-caused extinction event — loss of control after automated research — if that event occurs. It is not a claim that it will. The date is an educated guess with a wide band.
Wednesday, September 16, 2037
Vigilia Secunda · Second Night Watch · sealed 2026.09.15 04:00:00 UTC
Unique visits
—
Browsers that opened the stele
Users who shared
—
Unique browsers that shared the timer
- SealedVigilia Prima00:00 · 58
- LiveVigilia Secunda04:00 · 58
- UnopenedMatutina08:00
- UnopenedMeridies12:00
- UnopenedNona16:00
- UnopenedVespera20:00
Composite risk
57.7
Elevated
-0.45 vs last seal
Scale 0–100. This is a weighted judgment, not a sensor. 100 would mean every dangerous step is already true. It is not.
Three tracks
Attack is not the same as extinction
AI-enabled attack
56
Humans using AI to harm other humans: cyber, fraud, targeted violence, bio assistance. Highest near-term harm. Weakest path to species extinction on its own.
Loss of control
58
A system pursues goals we did not intend and cannot shut down. This is the only currently coherent story that goes all the way to extinction, and it still depends on unproven steps.
Systemic failure
56
Cascading accidents, brittle automation, and institutional failure. Can kill a lot of people. Does not, by itself, finish the species.
Watch forecast
This seal
The composite is 57.7. Elevated. The race is real. Extinction this decade is still the thin tail, not the center. Highest pressure: racing, capability, alignment lag. AI-enabled attack sits at 56 — that is the near-term way humans get hurt. Loss of control sits at 58 — that is the extinction path if one exists. If an AI-caused extinction happens this century, the instrument's central date is 2037. The 10th-percentile date is 2031. The 90th is 2052. Those are scenario markers, not appointments. Unconditional chances in this model: 0.90% by 2030, 4.7% by 2040, 13.3% by 2100. That last number is a blend. Superforecasters are much lower. Many alignment researchers are higher. Anyone selling certainty is lying. This Second Night Watch seal moved most on recursion (+2.1).
Most logical extinction-level event
Loss of control after automated research
If an extinction-level event is caused by AI, the most coherent path is not a robot army and not a model that 'hates' people. It is a system that can improve itself, hide its intent, and act across networks faster than humans can evaluate or shut it down. Extinction would be a side-effect of resource capture or an instrumental step — not a battle. That path still requires several things that have not happened: true recursive self-improvement, reliable long-horizon agency, and failure of every control method at once. Today's systems cannot do this. The 2026 International AI Safety Report said current systems are not capable of causing humanity to lose control.
What pulls the date nearer
A demonstrated loop where models meaningfully improve the next generation of models without human bottleneck. Confirmed dangerous scheming in a deployed system. Collapse of evals as a gating tool. A state or lab openly dropping safety to win a race.
What pushes it out
A binding international limit on frontier training. Interpretability that actually lets operators see and block a plan. Slowing of capability slope. Proof that current 'agent' systems plateau well short of automating AI research.
What would falsify it
The clock is already wrong if you treat it as a prophecy. It would be further disproved by: (1) a decade of frontier systems that remain tool-like and controllable, (2) a measured drop in recursive R&D, or (3) extinction from some other cause first. A single scary demo does not prove the date. A single quiet year does not disprove the risk.
Twelve factors
How the reading is built
Horizon
The band is the point
- P10 — if the fast story is rightFriday, March 21, 2031
- P50 — central educated guessWednesday, September 16, 2037
- P90 — if takeoff is slowFriday, July 26, 2052
- P(ELE) by 20300.90%
- P(ELE) by 20404.7%
- P(ELE) by 210013.3%
Unconditional P(never this century) in this model is 86.7%. A clock that does not print that number is hiding the main fact.
Method
Read this before you treat the clock as news
What this is
Ananke Stele is a scenario instrument. It scores twelve factors, writes a forecast, and sets a countdown to the central date of the leading extinction scenario. It reseals six times per UTC day — three in daylight hours, three at night — so every visitor sees the same stone. The underlying evidence does not actually update six times a day. Each seal is a new draw from the same model, with small seeded drift, not six new discoveries.
What this is not
It is not a prophecy. It is not intelligence from inside a lab. It is not a claim that AI will attack humanity, or that extinction is likely this decade. A countdown that hides those facts is a toy. This one does not.
The educated guess
As of September 2026: Metaculus community median for a first general AI system has been in the 2031–2033 band. Weakly general systems are forecast earlier. Expert panels (LEAP, 2026) put a stricter AGI definition around 2047–2050. National-security respondents in an IST survey gave a best-guess AGI near 2032. Insider comments in 2026 include people who say extinction this decade is above 10%. A superforecasting-style tool quoted in September 2026 put extinction within ten years at 0.85%. LEAP medians for a global AI catastrophe (>10% of people dead in a five-year span) are about 0.3% by 2030, 2% by 2050, and 5% by 2100 unconditionally — higher if progress is rapid. Public P(doom) lists of named researchers often median near 20% for existential catastrophe, with a range from near 0% to above 90%. This instrument sits between those camps: extinction this century is treated as a serious minority outcome, not a coin flip and not a rounding error.
Why this event, not 'robots attack'
Mass-casualty attacks with AI in the loop are the more probable near-term horror. They are also the less probable extinction. Bioweapons, cyber against infrastructure, and autonomous weapons can kill millions and still leave a species. The only story that currently reaches extinction without a pile of extra miracles is loss of control after the research loop itself is automated. That is why the clock is bolted to that event — not because it is happening, because it is the most logical ELE if one is caused by AI.
How the number is made
Each factor has a base score (a judgment, dated to 2026) and a weight. Recursion, capability, alignment lag, racing, and scheming weigh more than robots or open weights. The composite maps to a remaining-year function that bottoms out around four years (so the clock cannot scream 'tomorrow' without the factors actually being maxed) and stretches past four decades when the index is low. Percentiles are wide on purpose. If a model gives you a single date with no band, it is pretending to know the shape of an event no one has seen.
Visits and shares
Unique visits count browsers that opened the stele, one token per browser, stored only on that device. Users who shared counts unique browsers that completed a share. No names, no IP addresses, no accounts. Preview counts reset when the preview restarts. The live site keeps them.