AI-X tracks the technological conditions that could make a catastrophic AI loss-of-control event plausible. It is an evidence index — not a prediction of extinction.
Midnight represents a plausible loss-of-control threshold: a deployed AI system able to sustain a strategic multi-domain campaign, persist against containment, and materially reduce humanity's ability to recover control.
≠ extinction probabilityA high-performing model is not enough. AI-X asks whether seven separate barriers are being crossed. Weak gates constrain the total score through geometric aggregation.
Frontier systems now demonstrate very strong performance across coding, cyber, scientific reasoning and other strategically relevant domains.
Breadth does not yet imply robust strategic competence across arbitrary real-world conditions.
Historical values are retrospective model reconstructions, not measurements published at the time. They show how the same rubric reads earlier capability states.
5.0 minutes to midnight · Elevated
AI-X is designed to move both directions. Better containment, slower dangerous-capability growth, stronger recovery capacity or evidence that current trends do not generalize can move the clock away from midnight.
These pathways are defensive threat models. They describe required capability classes and limiting barriers without providing operational instructions for causing harm.
A sufficiently capable autonomous system combines multiple domains, adapts to countermeasures and degrades recovery pathways. This is the primary existential pathway in the model.
AI reduces some knowledge barriers in biology, but catastrophic realization still depends on physical access, experimentation, propagation and failure of response systems.
The more credible pathway is compressed decision time, misperception or cyber-enabled escalation rather than a simplistic direct-launch scenario.
Cascading disruption across digital and physical infrastructure could magnify other crises, though it is unlikely to cause literal extinction in isolation.
Large-scale physical autonomy would remove an important human leverage point, but current robotics and logistics remain major bottlenecks.
AI substantially accelerating AI research could shorten the time available for governance and safety adaptation, even if it is not itself an extinction mechanism.
Every movement must resolve to dated evidence and an explicit gate change. Headlines alone do not move AI-X.
Moves G1 and G5 upward; does not by itself imply persistent autonomous loss of control.
Tier AOpenAIAdvanced cyber and biology capability increases the importance of deployment controls.
Tier AAnthropicGovernance response is tracked separately from raw capability growth.
Tier AAnthropicSignals growing attention to the security of the physical compute layer.
Tier CNISTStrengthens the rising G2 trend while explicitly preserving uncertainty about extrapolation to defended real-world systems.
Tier BUK AI Security InstituteCompute growth accelerates capability, while energy, chips, networking and data-center construction remain physical constraints that a digital model cannot wish away.
Epoch log-linear fit to record AI data-center compute since Aug 2024
Tier DEpoch AIEpoch point estimate; underlying trend is uncertain
Tier DEpoch AIEpoch estimate for the current record-holder, Sep 2026
Tier DEpoch AIIEA Base Case projection
Tier DInternational Energy AgencyStatements from frontier-lab leaders and researchers can reveal expectations and disagreements. AI-X displays them as context, never as substitute evidence.
“the best-aligned model out there ... [but] likely poses more misalignment risk than any model we’ve used”
“The persisting importance of prompt engineering — and now harness engineering — is one of the best indicators of how far we are from AGI.”
“we are in the 'real impact phase' as they improve”
“confront the cyber threat posed by increasingly capable AI systems head-on”
Sources are separated by evidentiary role so that a lab claim, an independent evaluation and an expert opinion are never silently treated as equivalent.
AI-X intentionally makes weak links matter. A model cannot reach a plausible irreversible-loss threshold merely by being excellent at one dangerous task.
AI-X uses an equal-weight geometric mean across the seven normalized gate scores. A weak gate therefore constrains the overall index instead of being averaged away.
(4.5 × 3.2 × 3.4 × 2.0 × 3.8 × 3.0 × 1.5)1/7 / 5 = 57.75%Automated systems may collect and classify candidate evidence, but score changes require a documented human review and changelog entry.
AI-X is a proximity index, not a forecast, countdown, or probability of human extinction. The dial is a normalized visual mapping of the composite index, not a calendar.
Primary frontier-lab system cards, evaluations and policies
Independent capability and safety evaluations
Academic, governmental and international synthesis
Quantitative infrastructure and compute datasets
Expert statements, interviews and X posts — context only, never sole scoring evidence
AI-X exists to make an uncertain debate inspectable: what changed, which barrier moved, how strong the evidence is, and what would have to happen for the clock to move back.