Summary and purpose

This post was originally published here. It is written to be accessible and persuasive to people without an EA background, and aims at providing a light, readable introduction to the nuclear risk problem.

This is a review of Daniel Ellsberg's The Doomsday Machine: Confessions of a Nuclear War Planner, recounting some of the key historical anecdotes, providing a brief summary of the research on how bad nuclear war would be (which, as far as I can tell, does not support Ellsberg's claim that modern nuclear arsenals amount to a "doomsday machine" that might cause the end of human civilisation through nuclear winter), mentioning Ellsberg's main recommendations (US/Russia arsenal reductions, taking ICBMs off high-alert, and more treaties), and connecting the problem to the more general one of designing robust institutions.

The most interesting parts for those already familiar with the topic are likely the specific examples of bureaucratic insanity that Ellsberg witnessed firsthand (described in the first section), revolving around the lack of separate war plans for the USSR and China, and US generals' attitudes towards this.

Review

Here’s what former RAND Corporation nuclear strategy analyst (and later, Pentagon Papers leaker) Daniel Ellsberg and his colleague thought about the movie Doctor Strangelove – a dark and brilliant comedy about accidental nuclear war – after watching it in 1964:

We came out into the afternoon sunlight, dazed by the light and the film, both agreeing that what we had just seen was, essentially, a documentary.

Doctor Ellsberg, or: How I Learned to Start Worrying and Hate the Bomb

The age of mass slaughter of civilians as war strategy did not start with Hiroshima, but rather years before with British and, later, American bombing campaigns. No new moral or strategic choice was made in the decision to drop the atomic bombs on Japan; it was the natural outgrowth of the policies that had already incinerated Dresden and Tokyo.

Of course, nuclear technology meant an escalation of its scale. A single plane carrying an atomic bomb is more efficient at delivering mass death than a bomber fleet. Hydrogen weapons, in which the atomic bomb is a mere detonating cap for a fusion reaction, scale up the destructive power a thousandfold. Thanks to missiles that can strike anywhere on Earth within an hour and the insistence of many nuclear countries in keeping weapons on high alert, each nuclear power has a loaded gun trained at the civilian population of the others.

The perverse logic of this hostage situation leads to the sorts of insanities that make Ellsberg call Doctor Strangelove a documentary.

For example, the lack of any way to recall bombers was a true and deliberate part of the US nuclear response mechanism. The fictional horror scenario of unintentionally launched bombers continuing towards their targets while the rest of the world spends its final hours waiting powerlessly was at most fifteen minutes from becoming reality throughout the early Cold War.

(Thankfully the switch from bombers to faster missiles later removed this anxiety-inducing pre-Armageddon wait.)

Why? Presumably because a recall code could be stolen by the enemy and used to misdirect an attack. The logic of mutually assured destruction demands certain response without delay.

When the US Air Force was told to place electronic locks on Minuteman missiles to prevent unauthorised launch, they decided that the unlock code would always be set at 0000 0000, so that a launch would never be blocked because the code was missing (or because a nervous launch officer couldn’t punch in anything more complicated).

Delegating the authority to launch weapons is another way of ensuring launch readiness. If the president, vice president, and everyone else in the line of succession right down to the White House chef are nuked into oblivion by a surprise strike, it can’t interfere with the ability to retaliate, or else that’s exactly what the Soviets would immediately start planning to do. And so (as Ellsberg carefully investigated) Eisenhower discreetly gave the admiral of the US Indo-Pacific Command the right to start the nuclear war plan by his own initiative. Communication links across the Pacific can be unreliable, so the admiral further delegated launch authority down the chain. To Ellsberg it is unclear if even the president was aware of the further delegation, but perfectly clear that this is insanity: all it would take is a geopolitical crisis, some bad weather over Hawaii, and suddenly some over-eager general on a distant Pacific island thinks that nuclear war has broken out and it is their duty to join the fun.

This is not the end of bureaucratic madness. Ellsberg recounts his surprise after learning that the US had no war plan involving just the Soviet Union. Any nuclear attack would hit China as well. The admirals Ellsberg asked about this were incensed, leading Ellsberg to conclude:

Thus, if the president gave an order to attack only Soviet targets, CINCPAC [US Indo-Pacific Command, now called USINDOPACOM] forces, having destroyed Vladivostok and a few other minor targets in eastern Russia, would essentially have to sit out the war as observers—“on the sidelines,” as they thought of it—during the big game.

This was something that the admirals thought intolerable.

It gets worse:

[I]t had long been clear to me that if the highest authorities did give [an order that excluded China] it would be virtually impossible to implement that order quickly in the Pacific. That was true for technical as well as bureaucratic reasons. CINCPAC planners were working extremely hard, around the clock each year, just to produce one single plan for nuclear war against the Sino-Soviet bloc, and they simply didn’t have the ability to produce a second plan for war with the Soviet Union alone.

Why was it so difficult to create a nuclear war plan? A major reason was the enormous number of calculations needed to schedule the bombers so that they wouldn’t be swatted out of the sky by nukes dropped by other bombers:

Plans specified that a particular explosion would go off at time-over-target, or TOT (for example, 117 minutes and 32 seconds after the Execute order), and then a nearby explosion would go off 2 minutes and 12 seconds later, and so forth. If everything went according to plan, no plane would be struck down by the explosion from a bomb dropped by another plane; no “fratricide” would occur.

Practical inconveniences, like the fact that not all planes would manage to get themselves in the air equally fast, or the existence of weather, were ignored.

I pointed these two problems out to a planner once. “Yes, I’ve thought of these problems before,” he said. “Well, doesn’t that make you question the value of making all these calculations and plans?” “These men are risking their lives flying out there. We’ve got to do what we can to save their lives.” “But it doesn’t seem that this plan has any chance to save any lives at all. It would save lives only if the execution followed the plan down to the second, and there’s not even the remotest possibility of that happening.” “Well, we’re ordered to make these calculations, so that’s what we do.”

No sane person can think to themselves “let’s plan to kill three hundred million Chinese peasants in order to protect the egos of a few admirals, and because someone wants to make really detailed spreadsheets”. A badly designed bureaucracy won’t even blink.

How to describe that, other than insanity? Should the Pentagon officials and their subordinates have been institutionalized? But that was precisely the problem: they already were. Their institutions not only promoted this insanity, they demanded it. And still do. As do comparable institutions in Russia.

Cuban roulette

Ellsberg’s account of the Cuban missile crisis is particularly haunting. Both Kennedy and Khrushchev were eager to avoid war, more cautious than many of their advisors, and willing to step down the bravado even at steep political cost.

Yet at the peak of the crisis on October 27th, 1962, there were two occasions when nuclear war was averted by chance. The first occasion was when the captain of a Soviet submarine being hounded by American destroyers decided to launch a nuclear torpedo at the destroyers. On most submarines the agreement of the captain and political officer would have sufficed, but flotilla commander Vasili Arkhipov happened to be onboard this particular submarine, and had the authority to overrule the captain and the political officer.

Had Arkhipov been stationed on a different submarine:

The source of [the explosion caused by the nuclear torpedo] would have been mysterious to other commanders in the [US] Navy and officials on the ExComm, since no submarines known to be in the region were believed to carry nuclear warheads. The clear implication on the cause of the nuclear destruction of this antisubmarine hunter-killer group would have been a medium-range missile from Cuba whose launch had not been detected. That is the event that President Kennedy had announced on October 22nd would lead to a full-scale nuclear attack on the Soviet Union.

Perhaps Kennedy would have decided to take back his redline, and maybe the conflict might have deescalated even then. But the odds would have been long.

The second time was above Siberia and the Bering Sea. An American U-2 spy plane had wandered off-course into Soviet airspace. MiGs were scrambled to intercept it (perhaps believing it to be a reconnaissance plane for a larger attack), and American F-102As scrambled in turn to intercept the MiGs before they could get to the U-2. The F-102As were armed only with nuclear air-to-air weapons, since they were meant to be used against Russian nuclear bomber formations.

Secretary of Defense Robert McNamara reportedly ran out of a Pentagon meeting hysterically yelling “this means war with the Soviet Union” upon hearing the news. Kennedy, however, was calmer:

In a panic, [the chief of the Bureau of Intelligence and Research] rushed in to tell the president there was a U-2 over Russia being pursued by MiGs. Kennedy, very cool, responded from his rocking chair (as Hilsman reported) with an old Navy joke: “There’s always some son-of-a-bitch who didn’t get the word.”

Even assuming leaders who would rather lose face than commit genocide, their control over events is not perfect. Government bureaucracies, trigger-happy generals, and your generic sons-of-bitches who don’t get the message have a lot of inertia, which any central organising force will struggle to halt. Combined with the hair-trigger launch capability demanded by deterrence through mutually assured destruction, this means nuclear war cannot be removed from the realm of the possible.

’Tis but a scratch

How bad is nuclear war, really?

Ellsberg comes with a prepackaged answer: the nuclear weapons currently deployed by the US and Russia (all other arsenals combined are less than 10% of the total) are equivalent to a doomsday machine which, if activated, would result in a nuclear winter that ends human civilisation.

As far as I can tell, the research is not nearly as clearcut as Ellsberg would like to tell. Ellsberg writes of the “recent scientific confirmation of the thirty-year old nuclear winter “hypothesis””, but I’m not sure what this is meant to refer to.

The Doomsday Machine is about nuclear history and policy, not the effects of nuclear war, so it makes sense for Ellsberg to omit a detailed analysis of what exactly we think might happen to the atmosphere. However, as best as I can tell given other sources, Ellsberg’s claims of a scientific consensus for civilisation-ending nuclear winter following a war waged with post-Cold War nuclear stockpiles is simply too strong given the current evidence. This is a shame. Nuclear winter is a serious threat, and serious threats do not need exaggeration.

So what is our current understanding of nuclear winter? In a word: complicated.

Some older models of nuclear winter were challenged when burning oil wells in Kuwait during the 1991 Gulf War failed to cause global or even continental cooling. Later papers suggest that sufficiently large burning areas, such as entire cities, might lift smoke much higher than isolated burning oil wells and hence cause greater effects. Others argue that modern cities are not very likely to become firestorms. A recent study estimated 3-17% losses in various crops from a limited Indo-Pakistani war alone. Still others claim they’re being stigmatised as “closet Doctor Strangeloves” for their criticism of the nuclear winter hypothesis.

If we assume that the more aggressive nuclear winter models are not totally off the mark, this analysis estimates that billions of people might plausibly starve to death following a modern US-Russia nuclear exchange. However, to arrive at such estimates involves a long chain of assumptions.

I think all we can say for sure are two things: first, that this is not an experiment we ever want to try, and second, that there exists at least one foolproof solution to global warming.

Regardless of a hypothetical nuclear winter, any nuclear war is bad.

Consider the greatest disasters in human history. Events like World War II, the Black Death, the Great Chinese Famine, and the Spanish flu all had a death toll between 10 and 100 million people (though the high-end estimates for the Black Death go to twice that number).

In the early 1960s, the US Joint Chiefs of Staff estimated that a US first strike on the USSR and China would kill 275 million people immediately, and another 50 million within those countries over the next six months due to injuries and fallout. Attacks on Warsaw Pact countries would kill another 100 million. Collateral damage on neutral countries from fallout would depend on which way the wind blows, but likely add at least another 100 million across nearby countries like Finland, Japan, Sweden, and Afghanistan.

Ellsberg recounts his reaction:

I remember what I thought when I first held the single sheet with the graph on it. I thought, This piece of paper should not exist. It should never have existed. Not in America. Not anywhere, ever. It depicted evil beyond any human project ever. There should be nothing on earth, nothing real, that it referred to.

The scale is an order of magnitude above any other disaster. In terms of human life lost, it is as if all of World War II (from the Holocaust to Hiroshima to Dresden to Leningrad), the Black Death, the Great Chinese Famine, and the Mongol conquests all happened in a day, followed by World War I, the Spanish flu, and every famine the British ever caused in India over the next few months. Finally, add in some risk of a nuclear winter that slowly kills a significant chunk of the rest through starvation. And this is what happens if we assume that the Soviets don’t hit back.

An argument in favour of nuclear weapons is that they help maintain peace between great powers. This is true, but inadequate.

When asked to put a number on the probability of the Cuban missile crisis escalating into a total nuclear war, Kennedy said “between one in three and even”. Assume this is right, and that the alternative to a nuclear standoff was another worldwide military conflict on the scale of World War II (fought with conventional weapons only). The harsh logic of expected value tells us that the crisis was still a bad deal: we shouldn’t gamble 500-1000 million lives on a coin flip to avoid a 50-100 million death conflict.

Thankfully, a nuclear war today might be less damaging.

First, the number of nuclear weapons has gone down. The US arsenal peaked at 30 000 weapons in the 1960s and Soviet/Russian one at 40 000 in the 1980s; both have since fallen to 6000 - 7000. At the same time, the accuracy of missiles has improved, and smaller, more accurate weapons have replaced huge multi-megaton bombs that can wipe out a city even if they miss by a few kilometres.

Second, increased accuracy allows for strategies to change, at least among the most advanced nuclear powers. Countervalue targeting, where the aim is to inflict maximum damage to an enemy by hitting cities, can be swapped for counterforce targeting, in which the enemy’s military is targeted. The US might plausibly carry out a counterforce attack (see here for thoughts on whether they would), but the same is not currently true of China, let alone Pakistan or India. Of course, with nuclear weapons it is impossible to avoid collateral damage, and the extent to which countervalue targeting has been swapped out is hard to tell given the secrecy of current nuclear war plans.

By one estimate, even a limited counterforce scenario for a US-Russia nuclear war would lead to 10 million immediate deaths each in the US, Russia, and (if it’s involved) western Europe. Add in countervalue targeting, and that’s another 100 million across the US and Russia (an estimate for western Europe is not given). By weighing the probabilities of each level of countervalue targeting, the author of this estimate came up with a mean death toll of 50 million direct deaths.

So what can we expect for a modern nuclear war? Any nuclear exchange will likely earn a place near the top of Wikipedia’s “list of wars and anthropogenic disasters by death toll” page. A total one between large nuclear powers will instantly shoot to first place from the number of direct deaths alone. It would, entirely literally, be the worst thing ever.

Then there’s the possibility of nuclear winter. It might be uncertain, but its potential scale means its contribution to the expected number of deaths is considerable. Every 1% increase in the chance of half the world’s population starving is, in expectation, another Canada gone.

Understanding exactly how bad nuclear war would be is important, both to guide policy and to judge its importance relative to other causes. Right now there does not yet seem to be a consensus about nuclear winter risks; if we draw a graph of number of deaths versus probability of it happening, the distribution would be very wide, with most of the expected harm coming from the tail end: scenarios of low probability, but involving billions of deaths. Hopefully the immense efforts rightly spent on modelling climate change will have spillover benefits for nuclear winter research, and allow us to be more certain in our predictions.

Institutional insanity, then and now

Ellsberg does not discuss much about US war plans after his time working with them, likely because after he leaked the Pentagon Papers there was no going back to his job at RAND.

(In fact, the Pentagon Papers were just half of the secret material Ellsberg had copied. Ellsberg decided to release the Vietnam papers first, fearing that if he also released the nuclear planning papers, the Vietnam stuff would be forgotten. His plan to release the nuclear papers later was derailed due to a complex chain of events including letting his friend hide them in a dumpster and flooding from a near-hurricane. Much of the material has since been declassified, however.)

Some things have gotten better. The insanity of a single war plan hitting both the USSR and China must have ended as the Sino-Soviet split progressed (or so I assume). Permissive Action Links (PALs) are now often, but not always, used to make unauthorised nuclear weapon use harder. Nuclear brinksmanship is (for now) less common than during the Cold War.

However, as Ellsberg cautions, to think that the modern nuclear situation is much saner than the one that Ellsberg witnessed in the 1950s and 60s would be a mistake.

A key point to understand about nuclear war is that, if it happens, the reason for it will be stupid.

Nuclear weapons are meant to be used. Their intended use is not as explosives, though. As Ellsberg points out:

[…] they have been used in the precise way that a gun is used when you point it at someone’s head in a direct confrontation, whether or not the trigger is pulled. For a certain type of gun owner, getting their way in such situations without having to pull the trigger is the best use of the gun. It is why they have it, why they keep it loaded and ready to hand.

The world has fallen into a tragedy-of-the-commons -type situation. The commons, in this case, is the absence of nuclear weapons. Such a world is ideal, but the equilibrium is unstable: the first country to get them can threaten others, and so over time things degenerate until most (big) countries develop them. The commons has become exhausted, no nuclear power is better off relative to the others, and they are all paying the cost: upkeep of weapons, delivery systems, and infrastructure, as well as a small but ever-present risk of accidental mass murder.

(This is a simplification, of course. Nuclear weapons allow some countries to better their position relative even to other nuclear powers; for example, both Pakistan and India have nukes, but Pakistan comes off better in the deal since its weapons help offset its disadvantages in population, resources, and territory (as does the asymmetry of their nuclear policies - India has adopted a no-first-use policy, but Pakistan refuses to).)

No sane person starts a nuclear war. If a nuclear weapon detonates, it has failed its purpose. The most important risk both during and after the Cold War is that of accidental nuclear war – technical glitches, inadvertent escalation, Kennedy’s “sons-of-bitches who don’t get the word”.

It is possible to imagine a world where the delicate balance of nuclear deterrents can be maintained with the millimetre precision required to ensure that the expected value of harm remains low. In this world, nuclear weapons may even be a net positive, paying back the costs of their upkeep and probability of accident by reducing the likelihood of non-nuclear conflict.

Is this our world?

If all curtains of secrecy were stripped from US nuclear planning, would we see a rational government carefully shouldering its Atlas-like burdens? What guarantee is there that China and Russia, both of which either already have or are soon likely to have dictators for life, will give appropriate weight to impartial concern for human welfare in their nuclear strategies? Didn’t Narendra Modi order air strikes on another nuclear power to increase his reelection odds just last year?

Building institutions that carry out complex tasks reliably is a very difficult problem.

Consider some of the greatest institutions humanity has come up with. Democracy promises that if you let people vote for their leaders, there’s some chance your country won’t slide into authoritarianism or dysfunction. Free markets boil down to the realisation that you can get away with making surprisingly few decisions about the economy. Scientific publishing allows for the mound of human knowledge to continuously expand, except occasionally there’s a replication crisis and the floor falls in for half a field.

Such institutions are among the greatest achievements of human organisation and intelligence. Yet I still wouldn’t bet my life on a breakthrough study replicating, the stock market updating on a building but predictable global pandemic in its early stages, or a European democracy never sliding into dictatorship. Ask me to bet tens of millions of lives on US, Russian, Chinese, French, British, Pakistani, Indian, and Israeli secret military institutions all working reliably over a time horizon of decades, and I start to wonder when the first ship leaves for Mars.

I am not generally fond of arguments about human hubris (too often they boil down to vague complaints that using our ingenuity to improve life would somehow be bad). This time, however, there is truth to them. For the governments of the United States and Russia to believe that they can wield a thousand-weapon arsenal with sufficiently low mistake probability to justify it is probably pretence (at least if we assume that large-scale conventional wars are held back by non-nuclear factors). Our current institution-building abilities are not up to the task.

What should we do?

What are the most effective things we can do to lower the expected harm of nuclear war – that is, reduce both its probability and the damage it would cause if it happened?

Ellsberg rightly recommends downsizing the US and Russian arsenals (together over 90% of the world total) as a first step, since this would reduce any risk of a catastrophic nuclear winter. For example, the US could start by unilaterally getting rid of its land-based missiles, which would reduce the time pressures on making a launch decision (land-based missiles will likely be the first targets of any attack and hence will be lost unless launched soon after a warning), and deprive Russian missiles of their first targets, making it more justifiable for Russia to cut back on its own arsenal.

Another step is for countries to take weapons off hair-trigger launch alert. Constantly being two mistakes and ten minutes away from nuclear launch is not sustainable in the long run. Deterrence could be maintained through a focus on hardier nuclear forces like submarines, and less reliance on sitting ducks like land-based missile silos.

To make both of these steps more likely, diplomatic efforts into nuclear arms control treaties should be increased. This has not been happening. The US suspended the Intermediate-Range Nuclear Forces Treaty in 2019 due to perceived Russian violations and because it didn’t cover China. The remaining major US-Russia treaty, New START (STrategic Arms Reduction Treaty), got close to expiring before both Putin and Biden made it a priority in the first days of the Biden administration. Treaties for old weapons aren’t enough, either; new technologies, like hypersonic gliders that can keep lower than ballistic missiles and perform evasive manoeuvres, might destabilise nuclear deterrence.

In the long run, the aim should be to either abolish weaponised nukes entirely, or, failing that, at least reach a state where only a few accountable states wield small numbers of weapons. Biological and chemical weapons of mass destruction have been reined in by treaties and mostly abolished. Nuclear weapons should be next.

Surviving by design

During the Cold War, it was easy to construct a compelling narrative around nuclear war: the climactic showdown between the forces of capitalism and communism and between democracy and totalitarianism, to be waged for infinite stakes with the ultimate fruits of modern science. With the end of the Cold War, the narrative was lost. Nuclear war was still possible – the only time a “nuclear briefcase” was ever activated was in 1995 – but it was largely relegated to the realm of technical glitches and accidents; not things that make for a good story.

As Ellsberg points out, the biggest nuclear threat never was, and still isn’t, an intentional conflict (or rogue states or terrorists). It is the potential for failure in the institutions, people, and machines that control the biggest nuclear arsenals.

If nuclear war starts, it won’t be grave geopolitical considerations that trigger it. It will be a country dropping a nuclear weapon on itself, someone accidentally inserting a nuclear war training tape into an operational computer, radar equipment getting confused by the moon, or Russian bureaucracy being Russian bureaucracy. (Each of these happened; see links.)

Stories are nice. It’s tempting to demand a certain narrative coherence from the world; to think that sufficiently bad things, for sufficiently dumb reasons, aren’t allowed to happen.

But we do not live in the world of narrative coherence. We live in the world where civilisation was put indefinitely on hold because of bat soup in China. The greatest risks we face aren’t wrapped up in compelling narratives, and they do not come from commensurate causes.

We did not survive the Cold War by design. We survived it by accident: because none of the close calls quite managed to escalate to full-blown war, and because of heroes like Vasili Arkhipov and Stanislav Petrov who decided not to press the button.

We should survive the 21st century by design, not accident. This is not a given; with ever greater technology comes an ever greater number of efficient ways to kill a lot of people.

Making nuclear war less likely and less disastrous is a useful step towards achieving this. It is not the the only part, nor necessarily the most urgent; while death tolls are high and there is some possibility of nuclear winter verging on X-risk territory, biological warfare and unfriendly AI both seem at least comparable in terms of expected harm. However, I will say this: if civilisation has to be severely damaged by some apocalyptic scenario, we might as well make sure that it’s something fancy and trendy like unfriendly AI, not something straight from a 1960s comedy film.