Methodology — IRCA v2 · v2.0

The index

humancollapse publishes the Collapse Risk Index (IRCA v2), computed from public, auditable signals grouped into six pillars: AGI proximity, outcome severity, human cognition, climate strain from datacenters, labor displacement, and mitigation. Every signal is normalized to a 0–100 subscore with a cited source and a published uncertainty; pillar scores are weighted means of their signals.

The whole pipeline runs in ~30 ms on a CPU with no GPU and no machine-learning black box: every stage below is a closed-form equation from the published literature.

The pipeline in equations

pillar_i = Σ wⱼ·xⱼ / Σ wⱼ
s*_i     = s_i + (100 − s_i)·2·max(0, σ(c·Σⱼ A_ij·(s_j − τ)/100) − ½)    (×12)
C        = ( Σᵢ Wᵢ·(s*ᵢ)^p )^(1/p)                        p = 2.5
IRCA     = C · (1 − κ·M/100)                              κ = 0.25
λ_anchor = −ln(1 − P*)·b / (e^(b·H) − 1)                  H = 78 y,  b = 0.02
λ(IRCA)  = λ_low·exp( ln(λ_high/λ_low)·(IRCA − 25)/(80 − 25) )
S(t)     = exp( −(λ/b)·(e^(b·t) − 1) )
year_q   = year_now + (1/b)·ln( 1 − (b/λ)·ln(1 − q) )
p_ext    = (1 − S(100)) · 0.25

Mitigation enters as the dampener M, outside the risk weights: it reduces the index by up to 25% and can never add risk.

Cascade coupling

Interdependent systems fail abruptly, not gracefully (Buldyrev et al., Nature 2010; Motter & Lai, 2002). A documented coupling matrix amplifies pillars when their sources cross a threshold: fast AI progress amplifies labor displacement and cognitive offloading; environmental and labor strain amplify outcome severity. Coupling only amplifies — it never hides risk.

Parameters: τ = 50 · c = 4 · K = 12

Non-compensatory aggregation

A plain weighted average lets calm pillars wash out a critical one — the exact failure mode the OECD/JRC Handbook on Composite Indicators warns about. IRCA uses a weighted power mean with exponent p = 2.5: a single pillar spiking toward 100 pulls the index up disproportionately.

Mitigation (governance) enters as a bounded multiplicative dampener — stronger regulation reduces the index by up to 25%, but can never erase risk.

The variables

How each signal is measured, its normalization anchors, its weight and its published uncertainty. Live signals refresh every 6 hours; the rest follow their upstream report cadence.

Linear share = pillar weight × signal weight within its pillar. The true marginal impact grows as scores rise, because aggregation uses p = 2.5 (non-compensatory).

The collapse date

A risk level has no time semantics by itself. IRCA maps the index to an annual hazard via a proportional-hazards link (Cox, 1972) with a rising Gompertz baseline, calibrated to two published anchors: the Existential Risk Persuasion Tournament medians for AI-caused global catastrophe by 2100, 2.13% (superforecasters) and 12% (domain experts). Catastrophe is defined by the XPT as the death of at least 10% of humanity within a five-year window.

The countdown targets the median of the resulting date distribution — the year by which collapse probability reaches 50% under current conditions. Collapse is never certain by any fixed date: the survival curve stays above zero.

The extinction-level probability scales the 100-year catastrophe probability by the XPT experts' AI extinction/catastrophe ratio (0.25).

Calibration anchors

Uncertainty

The plausible range is the P10–P90 of a Monte Carlo simulation (JCGM 101; Saisana et al., 2005) that resamples every signal within its published uncertainty and every contestable model assumption (anchor probabilities, hazard shape) within its published range. The random seed derives from the computation date, so anyone re-running the open code reproduces the published band exactly.

Monte Carlo: N = 500 · P10–P90 · P*_low ∈ [0.005, 0.05] · P*_high ∈ [0.06, 0.3] · b ∈ [0.005, 0.035]

The new dimensions

Cognition: MIT's EEG study found up to 55% reduced brain connectivity in LLM-assisted writing; a CHI 2025 survey of knowledge workers found higher confidence in AI predicts lower critical-thinking effort; PNAS Nexus measured a causal 25% displacement of public knowledge-sharing. The live signal tracks the sustained decline of independent information-seeking (Wikipedia pageviews, Stack Overflow question rate) against the 2022 baseline. Contrary evidence (meta-analyses showing learning gains from AI tutoring) is catalogued in the evidence library.

Environment: the IEA projects datacenter electricity to grow from 415 TWh (2024) to ~945 TWh (2030); Google and Microsoft report emissions up 18–25% year over year, attributed to AI buildout. The live signal is the world grid's carbon intensity (Ember via Our World in Data) — the same buildout is far worse on a dirty grid.

Labor: the IMF estimates 40% of global jobs exposed to AI; AI-attributed layoffs reached 23% of US job cuts in H1 2026 (Challenger); the entry-level employment gap in AI-exposed occupations reached 19% (Stanford/ADP). The live signal is the world unemployment rate (World Bank).

Autonomous collection

Six scrapers (Manifold, Metaculus, arXiv, Our World in Data, World Bank, Wikimedia/Stack Exchange) run every 6 hours, update their signals, recompute the index and persist a daily snapshot. All endpoints are keyless public APIs; per-source health is shown on the sources page.

Scientific validity

Operationalization: every variable has a published measurement rule with normalization anchors (previous section), a cited source and a 1-sigma uncertainty. Nothing is scored without a rule a third party could re-apply.

Reproducibility: the code is open and the pipeline is deterministic given (inputs, date) — the Monte Carlo seed derives from the computation date. Every constant on this page is served by the public /api/v1/model endpoint, so the documentation cannot drift from the implementation.

Sensitivity: the plausible range already propagates both signal uncertainty and the contestable model assumptions (anchor probabilities, hazard shape) across their published ranges. Weights are constants in versioned source code; changing them is a visible, reviewable diff.

Falsifiability and updating: the index moves down when its inputs do — safety-paper velocity rising, AI-layoff attribution falling, grids decarbonizing, binding regulation strengthening. Known confounds are disclosed inline (PISA/COVID) and contrary evidence is catalogued in the evidence library rather than excluded.

Known validity threats: expert elicitations are not measurements; behavioral proxies (pageviews, question rates) can drift for unrelated reasons; single-provider components (ADP, Challenger) carry provider bias. These are mitigated by multi-source pillars and published uncertainties — and disclosed, not hidden.

Honest limitations

Anchors are expert and superforecaster elicitations, not measurements. The coupling matrix and weights are judgment — published and versioned so they can be challenged.

PISA 2022 is COVID-confounded; PISA 2025 (Dec 2026) is the first genuinely post-LLM cycle. Structural estimates (IMF, IEA) justify anchor placement; they are not time-varying inputs.

No one can predict an unprecedented event with certainty. We publish probabilities and ranges, never certainties.