A2. Human Extinction Risk Mitigation

Off-Worlding the Risk Portfolio

Off-Worlding the Risk Portfolio explains how going off-world reduces species-level risk not by distance alone, but by de-correlating failure modes. With N hubs (e.g., Earth + Mars), the species failure probability for hazard i can be decomposed into a common-mode term plus an independent term. Lowering cross-hub correlation ρi is therefore more valuable than simply adding hubs without independence.

Core concept:

Outcome classes.

GCR (Global Catastrophic Risk): billions affected; recovery plausible.

X-risk (Existential Risk): permanent, unrecoverable loss of humanity’s potential (extinction or locked-in dystopia).

Decomposition (for each hazard i).
Risk ∝ Hazard × Exposure × Vulnerability × Cross-hub Correlation (ρi).
Independence is engineered—through governance, traffic rules, and technology stacks—not granted by kilometers.

Portfolio approximation (two hubs, Earth E and Mars M).

Big wins come from low-ρ hazards (planet-specific shocks); benefits are limited for high-ρ hazards unless we deliberately de-correlate them.

Driver line. Baseline syntheses place this-century existential risk at non-negligible levels, with anthropogenic hazards dominating. Off-worlding sharply reduces risk where ρ is low and buys recovery options where ρ is high.

Distance ≠ independence:

Distance cuts some exposures (e.g., blast radius of nuclear war), but shared architectures and traffic can keep ρ near 1 for AI or bio.

Independence requires: (1) separate governance, (2) traffic controls/quarantine, (3) diverse tech stacks and supply chains, (4) fail-safe protocols that prevent Earth failures from becoming Mars failures (and vice versa).

Example: An AI catastrophe remains high-ρ if both hubs run the same models under the same oversight; it becomes lower-ρ under separately governed, capped-capability systems.

Sub-probabilities by major hazard (with planning notes)

Use order-of-magnitude bands (OOM), not point estimates. Highlight whether hazards are inherently high-ρ or naturally low-ρ.

Anthropogenic (high-ρ unless deliberately de-correlated)

Advanced AI misalignment

Evidence/anchors. Multiple large-n surveys of ML authors report median ≈ 5% chance of “extinction or similarly permanent disempowerment” within ~100 years. A broad expert statement frames AI as an extinction-level risk alongside pandemics and nuclear war.

Planning note. Strongly high-ρ unless architectures, oversight, and deployment are separately governed per hub; adopt capability caps, tripwires, and model diversity to push ρ↓.

Engineered biological risks

Evidence/anchors. The National Academies consensus lays out concrete threat pathways for synthetic biology (recreating known pathogens, enhancements, novel constructs).

Planning note. High-ρ if travel is open. Reduce with quarantine protocols, independent BSL capacity, biosurveillance, and separate vaccine/countermeasure platforms per hub.

Nuclear war → nuclear winter

Evidence/anchors. Modern climate–crop models estimate multi-year food shocks; worst cases imply multi-billion famine deaths.

Planning note. Low-ρ (Earth-centric). A self-sufficient Mars hub is a strong hedge—if its agriculture and energy do not depend on Earth imports.

Climate change (existential tail)

Evidence/anchors. IPCC AR6 documents severe risks and low-likelihood, high-impact tails (including tipping cascades). Extinction-scale outcomes are tail-tail, but systemic fragility is substantial.

Planning note. Correlated food/conflict/migration stresses; multiplanetary helps only with agricultural and energetic independence on Mars.

Natural (mostly low-ρ between Earth and Mars)

Asteroids & comets

Evidence/anchors. ~1-km impactors hit Earth on ~10^5 ⁣− ⁣10^6-year timescales; ~10-km on ~10^8-year scales. NASA estimates >90–95% of >1-km NEAs are cataloged. DART proved kinetic deflection, shortening Dimorphos’ period by ~32 minutes; 

Planning note. Classic low-ρ single-planet risk; Mars hub is a robust hedge. Maintain IAWN/SMPAG coordination and fund survey+deflection

Supervolcanoes (VEI-8)

Evidence/anchors. USGS places Yellowstone’s annual caldera-forming probability ~1 in 730,000 and stresses the event is exceedingly low on human timescales.

Planning note. Largely Earth-localized → low-ρ diversification via Mars.

Solar superstorms (extreme space weather)

Evidence/anchors. Estimates for a Carrington-class event span ~0.5–2% to ~12% per decade, reflecting different methods. Impacts concentrate on grids, satellites, nav/comms.

Planning note. Partly high-ρ (the Sun hits both), but consequences differ: hardened, underground, low-grid Mars designs can materially reduce vulnerability.

Nearby supernovae / gamma-ray bursts

Evidence/anchors. Lethal-distance events are very rare on human timescales; work suggests dangerous supernovae must be within a few-tens of light-years, and estimates for such proximity run ≪1% per century (geologic-timescale frequency).

Planning note. High-ρ across planets; prioritize shielding and underground habitats over geography.

Correlation-aware portfolio math (for planning)

Two hubs (Earth, Mars).

Big wins: low-ρ hazards (nuclear war, asteroids, supervolcanoes).

Limited wins unless de-correlated: high-ρ hazards (unaligned AI; some bio; stellar events).

N hubs (heuristic).

where pˉii​ captures the common-mode component. The marginal value of an additional hub collapses as ρi→1

Three Operating Scenarios

Three Operating Scenarios compares ways to “off-world the risk portfolio.” The core claim: distance ≠ independence. You cut species-level risk by lowering cross-hub correlation (ρ) — via governance, traffic, and tech-stack separation—not by kilometers alone.

1) Plug-&-Play — high ρ (cheap, fast, fragile)

Posture. Shared clouds and models; shared policies; frequent, lightly screened traffic; integrated supply chains.

Risk effect. High common-mode failures (AI, bio) keep ρ\rhoρ near 1; only planet-specific hazards (nuclear winter, supervolcanoes) decouple much.

When it appears. Early expansion, cost-minimizing phase; “copy Earth off-world.”

2) Governed — mid ρ (near-frontier performance, much safer)

Posture. Separate regulators and safety cases; model diversity and caps for AI; quarantine windows and signed cargo/data exchanges; partially decoupled supply chains.

Risk effect. Major anthropogenic hazards partly de-correlated (AI/bio), big wins on low-ρ\rhoρ naturals and nuclear winter.

IL7 sweet spot. For Mars at IL7, this delivers most of the benefit at modest cost.

3) Hard-Partition — low ρ (safest, slower)

Posture. Legal/physical/info firebreaks; strict AI capability caps; air-gapped stacks; minimal biotraffic; stockpiled spares and local substitutes.

Risk effect. ρ\rhoρ pushed toward 0 for bio/AI/infra; performance and speed hit by duplication and latency.

Illustrative quantitative impact (using accepted baseline)

Starting with a ~1 in 6 (~16.7%) century-scale X-risk (aggregate), apportioned across: AI, engineered bio, nuclear winter, climate tail, naturals, other anthropogenic. Apply the two-planet adjustment

with scenario-specific ρ assumptions (heuristic). One reasonable split yields:

What’s driving this:

AI dominates residual risk (~63–65% of the total in all three cases); big gains require independent governance and caps per hub. Expert surveys of ~5,000+ ML authors place the median chance of “extinction or similarly permanent disempowerment” around 5% this century; a broad expert statement ranks AI alongside pandemics and nuclear war as extinction-level.

Engineered bio meaningfully drops under quarantine + telepresence logistics (Governed) and further under Hard-Partition (strict biotraffic control). See the National Academies threat framework for concrete pathways and mitigations

Nuclear winter & naturals (asteroids, supervolcanoes) become species-level rounding errors once a second self-sufficient hub exists (low ρ and p^2 tiny). Contemporary climate-crop modeling finds multi-year food shocks from large exchanges—reinforcing the benefit of an off-Earth, independent food/energy base.

Climate tail sees partial benefit: Mars is not tied to Earth’s climate, but geopolitical/market feedbacks keep ρ>0 unless supply chains are genuinely independent. (AR6 stresses low-likelihood, high-impact tails and compounding risks.)

Planning heuristic. IL7 without policy/ops separation → ~15% improvement. IL7 with separate AI/bio regimes + quarantine → ~40%. Hard-Partition with strict caps + minimal biotraffic → ~2/3 reduction (from ~1 in 6 to ~1 in 17).

Moon vs Mars vs 3 Hubs

Moon vs Mars vs 3 Hubs compares multiplanetary architectures through a risk-portfolio lens. Physics (latency, cadence, transfer times) forces operating posture: the Moon tends to behave Plug-&-Play unless fenced; Mars is “governed by physics” (asynchronous by default). The result: for the next century, de-correlation (ρ↓) matters more than distance or hub count.

Architecture summary:

Earth-only baseline. Widely cited ~1 in 6 (≈16.7%) this-century existential-risk driver; treat as a range, not a point

+ Lunar hub (IL7, Plug-&-Play). Shared clouds/policies + continuous traffic → high ρ; indicative aggregate ~14%.

+ Martian hub (IL7, “Governed-by-physics”). Asynchronous latency + pulsed windows → mid ρ; indicative aggregate ~9–10%.

+ Martian hard-partition. Legal/physical/info firebreaks + AI caps + bio quarantine → low ρ; indicative aggregate ~~6%.

Third hub (3-hub IL7). Diminishing returns unless the third hub further lowers ρ for AI/bio.

Why the Moon ≠ Mars (physics that drives ρ)

Light-time / teleops.

Moon: ~1.3 s one-way ⇒ near-real-time cloud/teleops feasible → high coupling.

Mars: ~3–22 min one-way (configuration-dependent) ⇒ no tight control loops; async ops by design.

Crew/cargo cadence.

Moon: ~3-day transfers; essentially continuous windows.

Mars: ~7–10 months transit; launch opportunities ~every 26 months ⇒ pulsed traffic and natural quarantine gates.

Implication:

On the Moon, “Plug-&-Play by default” unless policy fences force separation.

On Mars, “Governed-by-physics”: local autonomy, signed async exchanges, quarantine windows—even without heavy policy.

Extending to 3 hubs (Earth + two IL7 hubs)

Use a simple common-mode + idiosyncratic form:

What’s going on

Diminishing returns: With two independent hubs, low-ρ hazards (nuclear winter, big naturals) are already near zero at the species level; a third hub moves the needle only if it pushes ρ lower for AI/bio.

Residual composition: Even with three hubs, AI stays ~60–65% of residual risk unless governance/caps truly diverge.

Design takeaway: If you invest in just one lever, invest in de-correlation (governance, traffic, stacks) — not hub count.