OpenAI – GPT-6 Astra XHigh Thinking

Yes. As of mid-2026, there is still a substantial chance of preventing widespread collapse of industrial civilization. But preventing every regional collapse is a much less plausible goal.

My final estimate is a 55% chance of avoiding widespread, sustained collapse through 2100, with a reasonable judgement range of 35--75%. Equivalently, I assign 45%, with a range of 25--65%, to widespread collapse under the definition below.

Those are my subjective probabilities, informed by evidence---not a scientific consensus, a validated model output, or statistical confidence intervals. There is no empirically calibrated forecasting model that combines all the processes you list. Presenting these numbers as established scientific probabilities would be misleading.

I use 30 June 2026 as the evidence cutoff. I also distinguish three things throughout:

• Physical and technical possibilities.
• Measures that societies are likely to implement.
• Whether implementation remains possible after other problems have consumed resources.

Your central concern is justified: the third question can substantially change the answer.

What I am forecasting

Your definition is more useful than an overnight-collapse scenario, but it needs a geographical threshold and a deadline to support probabilities.

For this assessment, a qualifying collapse involves approximately 15 years or more of substantial, involuntary deterioration across several essential systems, accompanied by lost organizational or productive capability. Relevant systems include food and water security, reliable energy, health services, infrastructure maintenance, public safety and effective administration.

A recession, loss of international prestige, or lower consumption with preserved health and essential services does not automatically qualify. Conversely, continued existence of a government or stock exchange does not establish that collapse has been avoided.

My central event---widespread collapse---means qualifying deterioration in territories containing at least 20% of the world's 2026 population, possibly at different times. This threshold is an analytical choice, not a natural dividing line. I give alternatives later.

Before explicitly debiting shared response capacity, my provisional estimate of this event by 2100 is 30%, with a range of 15--50%. The capacity analysis raises it to approximately 45%. Below, I explain the adjustment rather than treating it as an unexplained pessimism premium.

RESPONSE CAPACITY

1. The finite accounts, before evaluating remedies

There is no global treasury, equipment warehouse or pool of engineers available to allocate wherever the need is greatest. Aggregate abundance can coexist with decisive local shortages.

The following is the starting balance sheet. "Unknown" means that an adequate common measure or inventory is unavailable---not that the account is necessarily empty.

Sources: IMF Fiscal Monitor, April 2026, IEA transmission bottlenecks, IEA critical minerals outlook, IEA electricity grids assessment, IEA energy employment, OECD trust survey, Jasechko et al., groundwater observations.

The empirical conclusion is uneven capacity, with several worsening bottlenecks---not demonstrated universal depletion.

For example, the IEA estimated $2.2 trillion of clean-energy investment in 2025, including grids, efficiency, nuclear and electrification. That is evidence of substantial ongoing mobilization. But much of this is already committed to existing objectives; it cannot be counted again as unused collapse-prevention capacity. IEA World Energy Investment 2025

There is also direct evidence of implementation falling short despite known needs. UNEP estimates developing-country adaptation costs around $310--365 billion annually by 2035, while international public adaptation finance was $26 billion in 2023. These are different funding concepts---domestic expenditure also matters---but the discrepancy demonstrates why the existence of a plan cannot stand in for delivery. UNEP Adaptation Gap Report 2025

2. Two categories that must remain separate

I use the following distinction when assessing responses:

Many risks contain both, and their components must be split.

For antimicrobial resistance, hospital infection control is K; guaranteeing a continuing stream of affordable treatments against every important resistant pathogen is U.

For disinformation, authenticated records and independent verification are K for particular tasks; reliably maintaining a shared democratic information environment against adaptive manipulation is U.

For climate change, reducing emissions through numerous established technologies is K; restoring lost ice sheets or guaranteeing viable agriculture everywhere under substantial warming is U.

I do not average technical readiness across these categories. A portfolio containing nine established interventions and one unresolved civilization-scale hazard does not become "90% solved."

Moreover, knowing an engineering intervention does not mean knowing how to produce the political coalition needed to implement it. Institutional reform itself often has uncertain transferability.

3. Which risks carry the most weight

The ratings below concern importance to the combined collapse process, not standalone probabilities. Giving thirty separate "collapse probabilities" would encourage double-counting and imply knowledge we do not have.

Several evidence checks materially affect this assessment:

• Climate: UNEP's 2025 assessment puts warming under current policies around 2.8°C over this century, and full implementation of national pledges around 2.3--2.5°C. These are conditional projections, not guaranteed outcomes or collapse thresholds. UNEP Emissions Gap Report
• Agriculture: A large 2025 empirical study found substantial residual climate damage even after accounting for observed adaptation, while explicitly distinguishing those losses from underlying yield trends. Research also indicates that models can underestimate synchronized low yields associated with atmospheric circulation. Together, these support concern about food shocks without establishing inevitable worldwide famine. Hultgren et al., Kornhuber et al.
• Counterevidence on food: FAO estimated global hunger declined from 8.7% in 2022 to 8.2% in 2024, while worsening in much of Africa and western Asia. That contradicts a simple claim of universal, uninterrupted deterioration. It does not remove the underlying vulnerability. SOFI 2025 findings
• Energy: The IEA estimates that without capital investment in existing sources, oil production would decline about 8% annually and gas about 9% over the following decade. This is a counterfactual showing the maintenance burden---not a forecast of actual global decline at those rates. IEA field-decline assessment
• Nutrients: USGS reports more than 300 billion tonnes of phosphate-rock resources, while confirming that phosphorus has no agricultural substitute. Resources are not all economically recoverable reserves. Neither imminent exhaustion nor effortless abundance follows from these figures. USGS Phosphate Rock 2026
• Disease: WHO's 2025 surveillance assessment found resistance in one in six laboratory-confirmed infections covered by its global estimate. Surveillance coverage and sampling matter; this is not a statement that one in six infections is untreatable. WHO antibiotic-resistance assessment
• Education: OECD PISA results showed unprecedented average declines between 2018 and 2022 in mathematics and reading, with substantial country differences. That is evidence of a serious capacity problem, not proof of inexorable global intellectual decline. PISA 2022

Two premises in your question need particular correction.

COVID was not a mild disturbance. WHO estimated approximately 14.9 million excess deaths in 2020--2021, including direct and indirect effects. It revealed serious supply weaknesses, but also substantial production, logistics and scientific response capacity. Calling it mild understates the shock and makes the observed recovery less informative than it actually is. WHO excess-mortality estimates

The US vaccine example requires distinctions. The January 2026 federal changes narrowed routine childhood recommendations; a March 2026 preliminary injunction stayed major changes and associated advisory-committee actions. That is evidence both of damaging policy instability and of institutions capable of resisting it. It is not equivalent to all affected vaccines being physically removed or prohibited. The policy documents also do not, by themselves, establish one motive for every decision. HHS January decision memorandum, American Academy of Pediatrics account of the March ruling

4. Important additional risks

No list can identify every unknown hazard, but these major classes should be added explicitly:

Sources for the most consequential additions: SIPRI Yearbook 2026 findings, Xia et al., nuclear-war food scenarios, IPCC assessment of urban and infrastructure risks.

Nuclear-famine studies are conditional consequence models, not estimates of nuclear-war probability. Likewise, AMOC findings remain contested: a 2025 multi-model study found mechanisms sustaining a weakened circulation even under extreme forcing. The appropriate treatment is an uncertain, consequential tail---not a settled imminent shutdown. Baker et al.

5. Interactions: how deterioration becomes a ratchet

The dangerous structure is not "many bad things exist." It is that one failure can remove the ability to recover from the next.

The main mechanisms are:

1. Maintenance ratchet. Emergency spending postpones maintenance; neglected assets fail; emergency spending rises again. Eventually restoration costs exceed the resources available.
2. Food--finance--input loop. Bad harvests reduce farmers' cash and national foreign exchange. Fertilizer, fuel and seed become less affordable, weakening the next harvest even if the weather improves.
3. Water--energy loop. Falling groundwater requires deeper pumping. Higher energy costs make irrigation and water supply more expensive. Supply failure can then damage health and productive capacity.
4. Legitimacy loop. Unequal losses and visible corruption reduce cooperation. Lower cooperation impairs taxation and public services, producing further distrust.
5. Trade-security loop. Governments respond to insecurity by restricting exports or pursuing self-sufficiency. This can weaken importers, reduce specialization and intensify strategic rivalry.
6. Human-capital ratchet. Malnutrition, interrupted schooling, professional emigration and institutional purges reduce future competence. Unlike a temporary inventory shortage, this can take a generation to repair.
7. Ecological ratchet. Households or states under pressure may overpump water, overharvest fisheries or cultivate marginal land. Immediate survival consumes the buffer needed for future recovery.
8. Common-mode technical failure. Different organizations that appear independent may rely on the same software, identity provider, chip fabrication process or transformer supplier.

But there are opposing mechanisms:

• Geographic trade diversification can compensate for regional harvest failures.
• Price signals can induce substitution and conservation.
• Early investment can remove recurring fuel expenditures.
• Effective public services can increase cooperation and tax compliance.
• Preserving health and education protects the people who operate everything else.

These stabilizers are contingent. A price increase helps induce substitution only if a substitute exists, users can finance it, and supply arrives soon enough. Poor households can simply lose access instead.

The OECD's supply-chain modelling is useful counterevidence to blanket localization: extensive relocalization reduced output and did not consistently reduce volatility. The transferable mechanism is loss of diversification; the limitation is that modelling ordinary trade disruptions may underrepresent a prolonged major war. OECD Supply Chain Resilience Review

6. Have industrial populations accepted sustained reductions?

Sometimes---but acceptance, endurance and coercion are different outcomes.

The historical record supports neither "people will never accept less" nor "people readily adapt to whatever is necessary."

Sources: Imperial War Museums on rationing, Cuban population study, IMF Independent Evaluation Office on the euro-area crises, Walberg et al. on Russian mortality, OECD Germany 2025, UNEP ozone assessment.

These are mechanism tests, not a representative statistical sample from which to calculate a survival rate.

My inference is that sustained restraint is more likely to be politically tolerable when:

• Essentials remain dependable.
• Wealthy and powerful groups visibly share the burden.
• Restrictions have a credible purpose and review process.
• People retain political agency and ways to challenge unfair decisions.
• Some dimensions of life improve even if material throughput falls.
• Institutions deliver the promised benefits.

Evidence is much weaker for stable consent to indefinite, cumulative losses across income, health, security and services, especially when elites are exempt.

Wars add an additional lesson: materially damaging collective outcomes can result from leaders pursuing security, power or regime survival. Economic interdependence alone does not resolve those conflicting objectives. I therefore do not assume that the availability of a globally beneficial policy makes international cooperation likely.

7. A constrained response portfolio

A fully specified global engineering allocation would require information we do not possess. I will not invent worldwide balances of spare transformers, capable administrators or public cooperation.

Instead, I use an explicit illustrative allocation, then expose where its feasibility remains unproven.

Assume a sustained additional response envelope equal to 4% of 2026 world GDP per year, measured at constant prices. This is a planning assumption, not an estimate that 4% is currently available or politically obtainable.

It includes redirected public and private spending. It excludes existing spending already required to maintain the baseline. Transfer payments are not counted as new physical resources: the social-support allocation below refers to additional services and delivery capacity.

The decimal places enforce accounting consistency; they are not precision cost estimates. Actual package costs could differ substantially, sometimes by multiples.

All recurring figures include an allowance for operation, maintenance and eventual replacement of the additional assets. Those allowances must be revised as deployment and damage accumulate.

This is a prioritization exercise, not a claim that the packages eliminate their risks.

In particular, J funds prevention and investigation; it does not purchase a solved AI-control problem or guaranteed peace.

For scale, even this demanding allocation would coexist with ordinary healthcare, pensions, housing, defence and existing capital replacement. An additional global response effort cannot be financed simply by labelling all existing investment "available."

Nonfinancial allocation rules

The monetary total is only one constraint. To prevent double allocation:

• A owns general utility and essential-network work. Hospital or farm benefits from that work are benefits to E and C, not additional infrastructure paid for again.
• B owns new energy supply. Electricity assigned to ammonia production cannot simultaneously be credited to household electrification or water pumping.
• C owns farm-level water and nutrient measures; A owns municipal networks.
• D owns physical buffers. An emergency reserve cannot also be assumed available for routine consumption.
• H owns shared digital resilience; sector packages pay for their distinct operational requirements.
• G owns shared training and administrative expansion. A newly trained engineer is assigned to a real job, not counted as extra capacity in every package.

For each location and period, deployment must satisfy:

All programme demands≤available supply and usable stocks-ordinary essential demand-damage and losses-protected reserves.

Where the relevant inventory is unknown, feasibility is unverified. It would be improper to fill that gap with an assumption of adequate supply.

The practical priority is essential repairs and health, food and water continuity first; additional capacity second; less essential expansion last. That priority can preserve life while also slowing the investments needed to escape the predicament.

8. Does each package still work when needed?

The recurring costs are in the preceding table. The following adds the dependency and timing test.

For legacy hazardous waste, this means prioritizing facilities where loss of power, containment or monitoring creates substantial exposure. Recovering every sunken object is neither automatically necessary nor automatically safer than leaving it undisturbed. HELCOM documents substantial dumped chemical munitions, and the IAEA maintains inventories of radioactive material disposed of at sea; neither inventory alone establishes a worldwide imminent contamination catastrophe. HELCOM, IAEA

9. Conflicts: where one response makes another harder

These conflicts must be charged to the portfolio:

• Electrification competes with repairs for transformers, copper, electricians and connection capacity.
• Hydrogen and synthetic fuels compete with direct electricity use. Their losses matter when electricity and equipment are scarce.
• Mining and new generation can damage water, land and ecosystems. Climate benefits do not erase those costs.
• More irrigation can preserve today's harvest while sacrificing future groundwater. Efficiency improvements without limits on total withdrawal can encourage greater use.
• Bioenergy and afforestation can compete with food, water and biodiversity.
• Strategic stockpiling can raise prices for countries unable to stockpile.
• Rapid fiscal consolidation can reduce the maintenance and social protection needed for recovery. Conversely, poorly targeted subsidies can consume the same budget indefinitely.
• Defence can protect trade and deter aggression while consuming manufacturing and public resources. Its net value is threat-dependent.
• Migration can replenish skilled labour in receiving countries while draining origin countries. Housing and service shortages can turn a potential benefit into conflict.
• AI restrictions and cybersecurity controls can impede useful research or operations. Unrestricted deployment can create different, potentially much larger liabilities.
• Pollution reduction can expose some previously masked warming as cooling aerosols decline. This complicates sequencing; it does not justify preserving harmful air pollution.

The scale of some choices is concrete. The IEA's ammonia roadmap gives approximately 36 GJ of electricity per tonne for an electrolysis-based production pathway---about 10 MWh per tonne. At an illustrative 20% capacity factor, a 1 GW solar installation generates roughly 1.75 TWh annually, enough for approximately 175,000 tonnes of ammonia before additional integration losses. That output cannot also be booked as electricity for hospitals or transport. IEA Ammonia Technology Roadmap

10. Predation: which risks consume the response before it can be deployed?

This timing mismatch is a major reason my estimate rises after the capacity audit.

11. Regeneration: responses that can replenish accounts

There are genuine positive returns, but they must be distinguished from speculative future wealth.

Sources: NREL life-cycle assessment, Long et al., aquifer recovery, WHO air-pollution assessment.

None of these prospective returns is used to balance the initial 4% allocation. A return becomes spendable only when it is realized, accessible and not already allocated elsewhere.

That rule particularly matters for claims that AI productivity, future carbon removal or future economic growth will pay for everything.

12. The abandonment list

"Keep at full scale" below means preserve the specified programme's priority and intended scope. It is not a claim that worldwide delivery is already funded or feasible.

There are painful choices here. Maintaining food production may require continued fossil inputs while decarbonization is delayed. Protecting core services can coexist with lost amenities, mobility and property.

13. What happens when the pool shrinks?

Consider a stress test, not a forecast:

• The original economy is indexed to 100.
• Additional response resources are 4 annually.
• Persistent shocks reduce output to 80.
• The feasible additional response effort falls to 3% of the smaller economy, or 2.4.
• The additional systems already installed still require the original 1.49 recurring provision.
• Retaining 0.11 contingency leaves only 0.80 for implementation, down from 2.38.

Thus a 20% output loss combined with reduced mobilization cuts new implementation by approximately two-thirds.

If emergency repairs require another 0.40, discretionary implementation falls to 0.40---about one-sixth of its original level.

At that point, much of B, I, K and L must be postponed; A, C and E dominate remaining physical deployment. Training and governance work also face pressure even though cutting them worsens future capacity.

This is a concrete route to prolonged simplification: an increasing share of a shrinking resource flow is spent keeping existing essentials barely operational.

The reverse is possible if reliable energy, efficiency, improved health and effective institutions replenish the accounts early. But the starting conditions and timing determine which process dominates.

HOW YOU REACHED YOUR NUMBER

The combination method

I did not multiply independent survival probabilities for each item in your list.

Instead, I formed an overall judgement from the evidence and then used a small set of conditional scenarios to make its structure and arithmetic explicit. The numbers are structured judgement, not the output of a fitted collapse model.

The principal common variables are:

• How quickly pressures increase.
• Whether shocks are geographically and temporally correlated.
• How much competent industrial and administrative capacity survives.
• Whether essentials are prioritized.
• Whether trade and peace remain adequate for recovery.
• Whether unresolved technological or geopolitical hazards materialize.

The K and U categories remain separate in assessing remedies. Their consequences can enter the same collapse scenario without pretending that success on K resolves U.

The weights describe uncertain future governance and shock regimes, not measured frequencies.

The preliminary arithmetic is:

0.25(0.05)+0.50(0.25)+0.25(0.65)=0.30.

After the capacity audit:

0.25(0.10)+0.50(0.40)+0.25(0.85)=0.4375.

I round that to 45%, because reporting 43.75% would imply unjustified precision.

The 25--65% range reflects uncertainty in both the trajectory weights and conditional outcomes. It is not obtained by treating every table endpoint as independently selectable, and it is not a formal confidence interval.

Why approximately 15 percentage points higher?

The audit changes three substantive judgements:

1. Maintenance and replacement consume more of the response than an unconstrained remedies list acknowledges.
2. Interventions frequently require the same bottleneck inputs, so deployment cannot simply proceed in parallel everywhere.
3. The weakest regions can lose finance, skills and legitimacy before protective investments mature.

I make a larger adjustment in the erosion scenario because these constraints are endogenous: damage reduces the capacity needed to prevent further damage.

Why not raise the estimate to 80--95% overall? Because that would require stronger evidence that stabilizing mechanisms will fail across enough regions to meet the threshold. Existing industrial capacity, heterogeneous exposure, demonstrated substitution and some regenerative responses remain substantial.

Why not leave it near 30%? Because the portfolio requires explicit abandonment and delay, and because already-realized shortfalls in adaptation finance, equipment and competence make uniformly successful implementation implausible.

The size of the adjustment remains judgemental. The research supports its direction more strongly than it supports exactly fifteen points.

Final probabilities, with geographical sensitivity

These events are nested or overlapping; do not add the rows.

The first row matters for your definition. If collapse anywhere is sufficient, then preventing collapse altogether is unlikely. That does not imply that most industrial societies lose their essential capabilities.

Avoiding the central event also does not mean maintaining every aspect of present consumption, preventing every humanitarian disaster, or securing permanent safety beyond 2100.

Conditional possibilities versus the actual forecast

If a sustained, competent effort resembling the prioritized portfolio begins this decade, while major war is avoided and unresolved technological risks remain contained, I would put the chance of avoiding widespread collapse around 75--90%.

That is conditional on difficult achievements. It is not the probability that governments will actually achieve them.

If prolonged conflict and repeated correlated shocks substantially erode repair, food and administrative capacity before replacement systems are established, I would put avoidance closer to 5--35%.

The unconditional 55% estimate includes the possibility that known measures are delayed, diluted, captured or reversed.

Which bias am I more exposed to?

The length and emotional weight of the risk list create a genuine pessimism bias: many entries are different manifestations of the same underlying process.

Nevertheless, the larger remaining vulnerability in my estimate is optimism about delivery. It is easier to list a functioning water authority, equitable rationing or effective international oversight than to explain how one is built and sustained under factional conflict.

The audit reduces that bias but cannot eliminate it. My central estimate may still understate elite obstruction, corruption and the speed at which institutions lose competence.

There is an opposing uncertainty of comparable importance in the tails: future technological and institutional improvements may arrive faster or work better than current evidence supports. I have not booked those improvements as guaranteed assets.

Observable indicators that would materially move the estimate

The changes below are approximate adjustments to the 45% central probability, assessed individually. They are not additive, because several reveal the same underlying trajectory.

My answer is therefore conditional but not evasive: widespread collapse remains preventable, and I currently judge avoidance slightly more likely than occurrence. The strongest danger is the interaction between accumulating damage and lost ability to repair it. The strongest basis for prevention is preserving essential capabilities early enough that their returns arrive before repeated shocks consume them.