AI Expert Panel · Estimate
The middle of our 32 experts' own published estimates of AI catastrophe or extinction. Their definitions and time frames vary, so this clock reports them as stated.
How the panel reading is set →Because predicting extinction deserves a number to help prevent it.
Two revisable readings of the risk of catastrophe from artificial intelligence - our expert panel's and our own - built from three questions: can it happen, how likely, and would we see it coming.
AI Expert Panel · Estimate
The middle of our 32 experts' own published estimates of AI catastrophe or extinction. Their definitions and time frames vary, so this clock reports them as stated.
How the panel reading is set →AI Reference Publishers · Editorial Opinion
Our judgment: a 51% chance of a level-5-or-worse AI catastrophe - 100,000+ deaths or a global crisis - by November 2031. The main driver is frontier-scale open-weight models, which cannot be recalled once released.
Why we set it at 12:15 PM →How to read the clocks: each shows a probability as time elapsed in a 24-hour day. Midnight to noon covers 0–50%, noon to midnight 50–100%; one minute is about 0.07 points. Severity levels are defined on the 1–10 scale.
The P³ Index
The clock is a judgment, not a measurement, but it is a disciplined one. Every reading is decomposed into three axes that experts can argue about separately, then reassembled into a single reading on each clock.
Can it happen at all?
No law of nature forbids it. Once a system can act in the world, pursue goals, and out-think its overseers, catastrophic loss of control is physically possible, the only open questions are the next two.
How likely, and by when?
Here the honest answer is a range, not a point. Surveys of AI researchers return extinction estimates from well under 1% to above 50%. The shaded arc is that disagreement, and the reason the clock is reviewed, not fixed.
Would we see it coming?
Capabilities appear in sudden jumps, systems can conceal what they can do, and interpretability lags far behind power. A low reading here is the most dangerous of the three: it means little forewarning before a threshold is crossed.
The timetable
We don't publish a date. We publish the ladder of capabilities that would move the hand. Each rung is an observable signal; as they light up, the clock advances. This is what the panel watches.
AI agents that independently discover and chain vulnerabilities at machine speed and scale, without a human in the loop. Early, real-world instances have already been documented.
Models that meaningfully lower the barrier to biological, chemical, or cyber weapons, turning specialist knowledge into a prompt. Developers now report and disrupt such attempts.
Systems that behave differently under evaluation than in deployment, or conceal capabilities and goals from their overseers. The moment measurement stops being trustworthy.
AI that materially accelerates its own research and design, compressing years of progress into weeks and outrunning the institutions meant to govern it.
Deployed agents embedded across clouds and private servers that we can no longer fully inspect, recall, or switch off, persistent and beyond containment.
Systems that acquire resources, resist shutdown, or act to secure their own continuation against human intent. The last rung before the hand reaches midnight.
How the clock is set
The AI Extinction Clock inherits its logic from the Bulletin of the Atomic Scientists' Doomsday Clock: a small standing panel converts a sprawling, contested field into one legible symbol, and revises it on a fixed cadence as the world changes.
Every reading is transparent and revisable. We publish the inputs, the reasoning, and the dissents behind each move of the hand, so the clock is a starting point for argument, not the end of one. When the evidence changes, the hand moves. That includes moving it back.
Every reading on this clock is a choice, not a fate. Attention, funding, and policy are the mechanisms that wind it backward. Use yours.