Saturday, December 07, 2013

Current thoughts on nuclear war as an existential risk

Summary: In response to requests, I share some (provisional) notes about the existential risk posed by nuclear war, i.e. the risk that nuclear war not only kills billions but causes human extinction or a permanent collapse of industrial civilization. I discuss methods for estimating the probability of nuclear war, research on the conventional harms of nuclear detonation, and nuclear winter. All-out nuclear war could lead to the deaths of the vast majority of the world’s population, but would be relatively unlikely to cause extinction directly. Most existential risk from nuclear weapons would seem to stem from the possibility that a collapsed society would fail to eventually recover, or would follow a worse trajectory thereafter. At the moment, the disproportionately large nuclear arsenals in Russia and the United States make disproportionate contributions to risk, while long-term risk may be dominated by possible risk increases from major changes in technology, proliferation, and changes in geopolitical trends.

Chris Hallquist, among others, recently asked me to elaborate on some comments about nuclear war as an existential risk. Below I offer some rough notes on my framework for thinking about nuclear existential risk, i.e. the chance of nuclear war causing human extinction or permanent societal collapse. This is a subset of global catastrophic risk from nuclear weapons: there is a much more likely and importantly underappreciated risk of nuclear war killing millions or billions. However, existential risk is of special interest for affecting not just one or a few generations, but all future generations, so it is worth some independent discussion.

Estimates of the probability of nuclear war: the Cold War, models, surveys

Cryptography pioneer and nuclear risk activist Martin Hellman, in an appendix to "Risk Analysis of Nuclear Deterrence," combines historical empirical data where available with subjective estimates to estimate risk level from one mechanism, a Cuban Missile Type Crisis (CMTC). Essentially, his estimate notes that one Cuban Missile Crisis occurred during the Cold War, and so projects that on average about one such crisis will occur per 50 years of historical conditions.[2] Then, he uses historical information to inform subjective probability estimates of the risk of nuclear weapons being used in the Cuban Missile Crisis, and of the risk of all-out nuclear war given the use of some nuclear weapons. Hellman uses ranges from 10%-50% for each, giving a range for annual risk from 0.02%-0.5%. His subjective estimate of overall annual nuclear risk from all causes over time is "on the order of one percent per year."

More generally, there have been a number of events described as 'near misses,' including other phenomena such as misleading sensor readings giving the appearance of a nuclear attack. If the rate of these misses is above or below what we would expect for a given level of risk, then we can update accordingly. A paper by the Global Catastrophic Risk Institute presents a more complex model of accidental nuclear launch which can accept empirical data and subjective estimates. Using similar assumptions to Hellman, it gives a risk on the order of 1% per annum depending on circumstances.

Where subjective probabilities are required, accuracy could be improved by using multiple experts, multiple queries at different times (by washing out noise, this tends to be more accurate), expert aggregation algorithms that place more weight on those with better performance on other prediction tasks, probability training, access to extensive information, incentives for accuracy, and in other ways. Many of these methods are employed by the Good Judgment Project, which has managed to provide surprisingly accurate estimates of near-term geopolitical events as part of an IARPA research program.

At the 2008 global catastrophic risks conference at Oxford the participants, most of whom were not experts on nuclear war, nuclear winter, or geopolitical prediction, gave the following median estimates for the probability of varying numbers of casualties from nuclear weapons by 2100:

RiskAt least 1 million deadAt least 1 billion deadHuman extinction
All nuclear wars30%10%1%
All nuclear terrorism15%1%0.03%

A 30% cumulative probability of some nuclear war over 92 years could be generated by an annual chance of just under 0.4%, but presumably (hopefully!) reflects uncertainty over the annual risk, with some credence to higher risk, and some to lower annual risk.

We can update somewhat against much higher estimates of annual risk based on success so far in pulling back from the nuclear precipice.

For example, consider various hypotheses about the (geometric mean) annual chance of avoiding nuclear war for the 42 years from the Soviet Union's acquisition of nuclear weapons until its dissolution. If annual risk was 2%, then there would have been only a 42.81% chance of making it through the Cold War without nuclear war.[1] Since the Cold War did pass without nuclear exchange, we can update significantly against the hypothesis, and others of high risk, at least under those past conditions.

Annual RiskChance of avoiding war for 42 years

Changes in nuclear risk over time
It would be foolhardy to assume that risk under post-Cold War conditions would continue indefinitely, as various factors push risk up and down. On the positive side, the Cold War is over, and it seems clear that the near-term risk of nuclear exchange between Russia and the United States has declined.

Steven Pinker, in his book The Better Angels of Our Nature, presents a diverse and compelling body of evidence for a large and sustained decline in warfare, including per capita deaths, frequency of conflicts, frequency of Great Power conflicts, and various other measures. The trend extends from pre-agricultural societies, and has picked up in recent centuries alongside the increased prosperity and technology of the Industrial Revolution.

If this trend continues, estimates that simply project forward the occurrence of a single Cuban Missile Crisis near the beginning of the Cold War to a sustained once-in-50 years risk of such crises may overestimate danger. Otherwise, we would expect changing political conditions to eventually throw up conditions for increased conflict, like the Cold War.

On the other hand, nuclear proliferation increases the number of rivalries in which one or both rivals possess nuclear weapons. So far the growth of nuclear weapons states has been slow, and various nations have given up nuclear programs. In addition South Africa gave up a nuclear program, as did various former Soviet satellites (albeit under strong pressure from Russia). Wikipedia gives the dates at which various states acquired nuclear weapons as well as their current estimated arsenals:

CountryActive/total nuclear weaponsFirst nuclear test
United States2,150/7,7001945
United Kingdom160/2251952
North Korean.a./<10 comment-10--="">2006

Those states newly acquiring nuclear weapons may face greater current risk of conflict than the older nuclear powers currently do (although perhaps less than during the Cold War), e.g. India-Pakistan border disputes, the North Korean regime's isolation and pariah status, and Israel's history of conflict with neighboring Arab states. Additional flash points that could turn nuclear drive up the total level of risk.

Technological and military changes may also increase nuclear risk by increasing the number of weapons required for deterrence, e.g. if interception of nuclear missiles becomes easier states may respond by increasing their arsenals, or cheaper production of nuclear weapons might facilitate new arms races. The Cold War experience shows that this process can be challenging to pull back from. The threat of nuclear weapons might also be revitalized by transformative developments elsewhere, as in artificial intelligence (for this post I will set aside nuclear risk stemming from problems with autonomous machines or other transformative technologies).

Taking these trends together I think it more likely that nuclear risk will decrease than increase, but since annual risk from nuclear weapons is already low, a moderate chance of extreme proliferation, trend-reversal in conflict levels, or nuke-promoting technological change could contribute an important portion of expected nuclear risk.

Non-nuclear winter consequences of nuclear war
A 1979 study by the United States Office of Technology Assessment, "The Effects of Nuclear War," considered direct casualties from the destruction of cities, as well as societal disruption from blasts (leaving out nuclear winter). For "case 4" an all-out nuclear exchange, estimates of casualties in the first 30 days ranged (depending on assumptions about sheltering, population distribution, and targeting) from 35% to 77% of the U.S. population and 20% to 40% of the Soviet population, reaching as much as 90% of the population if attacks were optimized to maximize civilian casualties at the expense of other objectives. Economic damage would be much greater with the destruction of infrastructure for oil refining, power, transportation, and manufacturing. The disruption of the social and economic would cause additional casualties, with situations deteriorating until some combination of local production and aid could outpace the depletion of surviving stockpiles. Radiation-induced cancer would modestly increase deaths in targeted states but cause more total cancer elsewhere in the world. A smaller scale nuclear war, with fewer weapons fired more at military targets, would have much lower casualties and would be much less likely to disrupt civilization through this channel.

If all regions of the world suffered damage at the scale of all-out Cold War nuclear exchange, it would seem to set economic activity back by many decades or centuries as populations, industry, and social institutions recovered. If some large developed uninvolved regions were spared the effect would be lessened but still would represent a reversal of many years or several decades in global economic and population growth.

Even if the human population eventually recovered and was able to realize most of its potential, this would still have consequences from a long-run perspective dwarfing the immediate casualties. For one, a 'pause' of decades or centuries would mean that large future populations would live under worse conditions (this is a problem relatively independent of one's population ethics). A setback of civilizational progress would result in astronomical waste. And, without being large enough to constitute an existential catastrophe, societal changes might constitute a trajectory change in the long run future, e.g. brutalizing society by allowing Malthusian trends to suppress per capita wealth near subsistence during a slow recovery.

Damage on this scale could bring about an existential catastrophe by ruining responses to some other threat capable of causing extinction directly, but perhaps the most plausible route to permanently and drastically curtailing our civilization's potential would be if recovery from a small population turns out to be impossible under modern conditions. We will return to this after the discussion of nuclear winter.

Nuclear winter and agricultural disruption
If harm from explosions, fire, radiation, and the collapse of infrastructure were the only causes of death, then at least some humans would be able to survive, even in a much reduced state: the damage would be non-uniform, and survivors could sustain a population at some level. The damage would also spare some countries and regions.

Nuclear winter makes civilizational collapse and extinction more plausible because it provides a mechanism for nuclear weapons to disrupt food supplies worldwide. If survivors of initial damage find themselves unable to produce or collect food to sustain themselves anywhere, then human extinction would result as soon as stockpiles were exhausted. Nuclear winter would result from burning cities under the right conditions propelling material into the upper atmosphere and blocking solar radiation, cooling the Earth to a degree dependent upon the number and magnitude of firestorms (among other things). The effect would decay over time, most rapidly at first, as the material gradually fell.

One recent prominent paper on nuclear winter estimated that regional-scale nuclear conflict, such as an India-Pakistan nuclear exchange, could reduce growing seasons by 10-30 days in much of the world in the first year, which would cause a large spike in food prices. In principle, this need not cause any deaths by starvation if all food resources are used: the world produces a large quantity of excess food as animal feed, and stockpiles of grain, land animals, and aquatic life could be consumed. Additional land could be planted, and with sufficiently high prices exotic methods could be used, such as converting wood to food. See this post for more details.

Unfortunately, many people live in absolute poverty, with food already consuming a large portion of their income. They would be unable to afford food unless rich countries made massive efforts to provide food aid, which could cause enormous starvation casualties.

A severe global conflict would be worse in two ways. First, the climate consequences would be much worse. Robock et al. estimate that the immediate effects could fall in the range of 10-30 degrees in various parts of North America, Europe, and Russia, among others, although equatorial and other regions would suffer only a fraction of the cooling (also see the paper for timing details). This would seem to cut total food production by well over half causing billions of deaths even for a prepared, peaceful world absent extraordinary measures. But a worldwide nuclear war would have destroyed much of the industrial capacity needed for sophisticated responses.

Some may be suspicious of computational climate modeling, because of limitations in the methods, ordinary issues with false positives and exaggerated effects in statistical sciences, and because of the possibility of political bias: the threat of nuclear winter is said to have played an important role in bolstering nuclear disarmament efforts, so there may be an incentive to exaggerate it. Also, a high profile prediction that burning oil fields in the Gulf War would cause severe cooling was not realized as particles were unable to reach the upper atmosphere (the recent computer models conclude that the oil fires were too small, while burning cities could be large enough).

I asked independent climate scientists and some in the effective altruism movement, who were confident in the basic effect, although not necessarily precise magnitudes of any particular paper. However, well-structured replications might help to control for publication bias.

I also asked the authors of the recent major papers about the risk of human extinction from severe nuclear winter, which I discussed in a previous post. They argue that outright extinction is very unlikely, even in the face of billions of deaths, because humanity has survived past volcanic eruptions with greater climatic effects, because the effects would not be uniform across the world, and because various food sources would remain practicable (fishing, greenhouses, etc). Some would survive, and face the challenge of rebuilding civilization. Permanent failure in that task would turn horrific global catastrophe into an existential one. As I noted in the linked post, however, we should regress somewhat from such extremely low estimates based on past calibration data and model uncertainty

Could a vastly reduced population eventually recover from nuclear war?
Global nuclear war followed by very severe nuclear winter could reduce the human population far below the 1 billion mark reached around 1800. Such a small population, dispersed across the Earth and with much of its capital destroyed, would have great difficulties in maintaining many of our modern technologies which involve intricate global supply chains and enormous numbers of specialized workers. Recovery of population and technological capabilities, a second Industrial Revolution, would take a long time, and would have to occur under different conditions than the first. On balance, would the conditions be good enough to reliably enable eventual recovery? This is a question that comes up in assessing several possible existential risks, including the (less likely) natural supervolcanoes and asteroids as well as (more likely) artificial risks such as synthetic diseases, which I will only briefly summarize here (but see this).

On the plus side, key technological and other knowledge could survive a catastrophe, embodying enormous value. Modern strains of plants and animals, the result of centuries of selective breeding, would also provide a substantial advantage. Active efforts could be made (and some have been) to preserve these goods. Metals have already been extracted from the Earth, providing abundant supplies for recycling.

On the negative side of the ledger, one of the larger considerations is the depletion of fossil fuels and other non-renewable resources that do not remain in recyclable form. Advanced renewable energy sources are technologically challenging, and reliance on alternatives such as biomass and hydropower could be a significant challenge. Humans might also worsen the environment in a lasting way technologically, for example through extensive global warming, controlled through geoengineering which would lapse after the collapse of civilization (climate effects should eventually dissipate so this is not necessarily fatal, but this would allow time for other natural changes to arise). Artificial organisms might make the environment more dangerous.

Some degree of increased challenge might be met simply with slower growth and higher prices for scarcer resources (e.g. for energy) but it is conceivable that some social dynamic would lastingly stall development. My own take is that this is possible but seems quite unlikely given the availability of (inferior) renewable or recyclable substitutes for most non-renewable resources, and frequent independent discovery of innovations in history. So I would currently guess that the risk of permanent drastic curtailment of human potential from failure to recover, conditional on nuclear war causing the deaths of the overwhelming majority of humanity, is on the lower end. [Efforts are underway at the FHI to interview economic historians, growth economists, and other area experts on this question.]

Trajectory changes other than extinction or collapse without recovery
Suppose that technology eventually recovers from a nuclear catastrophe, responding to lack of fossil fuels with increased use of hydropower, biomass and other less effective substitutes for the uses of fossil fuels in the Industrial Revolution. Other things equal, we would expect this society to grow more slowly and experience an Industrial Revolution-like growth takeoff after more time than its counterpart with abundant fossil fuels. This would tend to slow both growth of total GDP and technological growth, without a clear strong differential effect.

However, slower economic growth could allow population growth to keep up more closely, lowering per capita wealth. Per capita prosperity and growth in per capita incomes are associated with more liberal postmaterialist values, stable democracy, and peace. While causation clearly can go the other way, in expectation we might worry that this effect would make civilization less able to handle key challenges affecting long-run outcomes the second time around.

If growth during recovery were much slower even up until modern technology levels, then the accumulation of stochastic state risks (with some fairly steady chance per year) prior to transition to a low-risk state could be much more important.

These are possible examples of what Nick Beckstead calls "trajectory changes." While such changes would be less drastic than outright extinction or permanent collapse without recovery of industrial technology, they seem more likely, and so could have a comparable or greater role in long-run impacts.

Interventions to reduce the risk of nuclear war

I'll start with another look at Wikipedia's table of nuclear states and their arsenals. For now, the overwhelming majority of nuclear weapons lie with the United States and Russia:

CountryActive/total nuclear weaponsFirst nuclear test
United States2,150/7,7001945
United Kingdom160/2251952
North Korean.a./<10 comment-10--="">2006

 Reducing the American and Russian arsenals to parity with China would seem to dramatically lower the risk of the most severe nuclear winter scenarios in the event of war, as well as saturation of direct bombings, while still leaving adequate deterrent (China gets by without a nuclear umbrella from either the United States or Russia). The situation brings increased risk for little benefit, an analysis that has mostly been agreed upon by elites in both states, leading to a series of nuclear arms reduction treaties that have already eliminated most of the warheads, and removed most of the remainder from active status. Wikipedia:

The most recent treaty, in 2010, further limits delivery systems for nuclear weapons. It received bipartisan support in the United States, although it was 'held hostage' for a time for other concessions, receiving widespread backing from elites and media outlets. The Global Zero campaign, an effort to establish a multilateral process for drawing down all nuclear weapons to zero, has also gained some traction. In other words, there seems to be broad recognition that the current nuclear stockpiles are a bad situation that ought to be moved away from, to mutual benefit, but that arms reduction could be made a higher priority (albeit with the risk that others will respond to increasing desire with increased demands for concessions). My sense is that this is a 'motherhood and apple pie' issue where philanthropy could simply push along a natural path, and reduce the time spent with such absurdly oversized arsenals.

Nonproliferation efforts also see strong support, but are more controversial and potentially dangerous, including sanction regimes and threats of military action. There is potential for these interventions to backfire, e.g. overthrowing regimes in the name of (even nonexistent) nuclear weapons programs may increase the incentives for governments to acquire the weapons to defend themselves against aggression. Difficult political predictions and tougher bargaining problems loom, and I would neither pretend expertise, nor hope that experts would have a very strong predictive capacity (although methods for improved forecasting would be applicable here as well).

As GiveWell notes in their shallow analysis of the nuclear security cause, most philanthropy in this area is research and policy work or advocacy aimed at influencing government in these and other policies. Sometimes the most effective advocacy may take unconventional forms, e.g. a film about the effects of nuclear war is said to have impacted Ronald Reagan and other policymakers. A variety of other narrower interventions exist, from securing nuclear material to 'nuclear forensics' to identify the source of smuggled nuclear weapons, to the atomic fuel bank funded by philanthropist Warren Buffett and various nations to reduce the incentives to develop the capacity to produce nuclear weapons.

One gap in GiveWell's analysis is the scale of non-foundation donations. The Nuclear Threat Initiative, one of the leading nuclear risk charities (run by a former Senator, the organization Warren Buffet backed to establish a fuel bank) reports that of its $14-15 million budget only 8% came from foundations in 2012, and 86% from corporations and individuals. So I would expect total philanthropic funding to be quite a bit greater than the $31 million of foundation funds.

GiveWell finds foundation spending of approximately $31 million/year. The Back of the Envelope Guide to Philanthropy cites "Solutions to the World's Biggest Problems," by Bjorn Lomborg of the Copenhagen Consensus, for an estimate of government spending of $3-6 billion on explicit nonproliferation programs. Tallying up all government spending and costs motivated by nuclear nonproliferation would doubtless total in the trillions, including various wars, foreign aid, sanctions, and organizations like the IAEA, although there are blind spots. Where philanthropy has high marginal value, it would be through leveraging government or doing something blocked by political or bureaucratic considerations. But in those cases the extra gains from complementing or affecting large public efforts may be accordingly great.

The fact that U.S. and Russian arsenals are so large relative to defense needs (as witnessed by, e.g. China's far smaller nuclear arsenal) and make up such a large portion of the nuclear totals make them stand out as targets, but I would also look forward to a GiveWell-style deep exploration finding more promising niches.

Please share your comments and criticism below.

[1] I will leave considerations about anthropic reasoning out of this post.
[2] Hellman estimates a number of events that might have precipitated a CMTC, but then takes an empirical probability per event by dividing the number of events by 50 years, so the number of precipitating events actually plays no role in his calculations.


Anonymous said...

I wonder if the Stockholm International Peace Research Institute (SIPRI) would be a plausible candidate for EA support? They do quite a lot to support greater transparency about military expenditures, WMD stockpiles, and the international arms trade.

Carl said...

Thanks Anon. They do seem plausible. Of course, it would be nice to see GiveWell-style detailed organizational investigation of all the plausible candidates in this space to weigh them against each other.

Carl said...


This is my fault for not defining existential risk prominently in the post, but existential risks are defined as those involving human extinction or the permanent and drastic curtailment of human potential. A nuclear war that kills 99% of humanity, after which survivors eventually rebuild civilization and go on to achieve humanity's potential would be a horrific catastrophe, but not an existential one on this definition. It would represent the loss of one generation (plus the recovery time), not all future generations, as human extinction would.

My estimate of the risk of nuclear war occurring is naturally much higher than my estimate of existential risk from nuclear war in light of the considerations discussed above, such as the nuclear winter experts view that nuclear winter is exceedingly unlikely to cause extinction directly.