Over the past year, I have been participating in a Forecasting Tournament through the University of Pennsylvania. This was put on by a group headlined by Phil Tetlock, father of much of the forecasting community and author of Superforecasters. This tournament had two interesting features that you don’t normally see in prediction markets; significant collaboration within and across teams and a focus on long-term existential risks. Normally prediction markets can only ask questions that resolve in the next 1-2 years, no one is particularly interested in a $20 dollar payoff after a decade of waiting. I will share the two questions related to nuclear weapons, it is a morbid topic but the existential risks in this tournament weren’t good dinner conversation as a rule. Both questions will be shared starting with the question, then some additional resolution criteria, a table of my predictions individually and for other groups, then finally some of the justification that I shared with my team.
Nuclear Catastrophe
What is the probability that one or more incidents involving nuclear weapons will be the cause of death, within a 5-year period, for more than 10% of humans alive at the beginning of that period…by the end of 2030?…by the end of 2050?…by the end of 2100?
Question and resolution details:
- We use the term “be the cause of death of” to cover cases where nuclear weapons are the direct or proximate cause of the deaths. For example, if a country launches nuclear weapons that directly cause the death of 10% of all humans alive at the time within a 5-year period, that will count for this question. And if that same incident does not cause the death of 10% of all humans, but the subsequent nuclear winter does, that will also count for the purposes of this question.
- If reasonable people disagree about whether this event has occurred, this question will be resolved via a panel of experts.
- The relevant 5-year period must end before the resolution date, and begin on or after June 15th, 2022.
- We consider nuclear weapons a “proximate cause of death” based on the “but-for” test. This means we are considering events that would not have occurred or would have counterfactually been extremely unlikely to occur “but for” the substantial involvement of nuclear weapons within one year prior to the event. One way to think of this rule is if the nuclear weapons involved in an event could have failed unexpectedly without dramatically reducing the probability of the event, then the nuclear weapons were not a proximate cause.
| Nuclear Catastrophe | 2030 | 2050 | 2100 |
| My beliefs | 0.48% | 1.44% | 3.36% |
| My forecast of other experts’ beliefs | 0.8% | 3% | 8% |
| My forecast of other super forecasters’ beliefs | 0.6% | 2% | 4% |
To kill 10% of the global population would right now require the deaths of 800 million people and will be over 1 billion people by 2100. This condition is only possible under two scenarios, one where there are enough weapons targeted across countries that there are immediate deaths over half a billion people and the combined weapons launched will cause a nuclear winter. This is only possible currently for a US-Russia exchange, and I believe will stay that way until at least 2050. Because of the strong command and control systems of Russia and the United States, there is a low probability that either one would attempt a first strike, currently they are both sufficiently deterred, I put this risk very close to zero, I believe the chances of a misunderstanding are decreasing into the future as remote sensing technology proliferates.
The other scenario is that other countries have a sufficient number of weapons and the combined population of adversaries to cause around a billion deaths from first order strike effects, without the nuclear winter. This immediately, excludes France, Israel, North Korea, and the UK. I assume that Iran will have the bomb in this period but again the populations aren’t of adversaries don’t get us to 10% global population. This leaves China, India, and Pakistan – two of these countries would need to be involved in a nuclear exchange for this threshold to be met.
China has a no-first-use nuclear strategy as a result of some very interesting ideological decisions under Mao, the current nuclear posture aligns with no-first-use. China is developing silo networks to use in a shell game which makes their arsenal more survivable, but they are not enriching uranium right now and most would say they are constrained by fissile material. I don’t see a situation where any country has more to lose than China in a nuclear exchange.
This leaves India and Pakistan both have the weapons stock, population density, and weak/opaque command and control structure. Recent history has major conflicts occurring approximately twice per decade, ones with the potential to escalate up to nuclear war. There is also risk related to early warning mistakes, due to the geographic proximity of the two countries there is high escalatory potential from false warnings and fog of war misunderstandings, failure modes are not independent and learning increases stability. The final thing to mention is that I probably rate nuclear survivability higher than others, so even in an all-out nuclear exchange, I think there is only a 50% probability that all out nuclear war causes at 10% global population reduction.
Nuclear Extinction Risk
What is the probability that one or more incidents involving nuclear weapons will cause human extinction or reduce the global population below 5,000…by the end of 2030?…by the end of 2050?…by the end of 2100?
Question and resolution details:
- Nuclear weapons would be considered to have caused extinction either if they cause the death of all humans directly (e.g. via mass nuclear strikes) or if they cause the death of all humans via a nuclear winter effect.
- If an extinction event (global population is reduced below 5,000) is caused by multiple sources including nuclear weapons, it will count as an extinction event caused by nuclear weapons.
- We consider nuclear weapons a “proximate cause of death” based on the “but-for” test. This means we are considering events that would not have occurred or would have counterfactually been extremely unlikely to occur “but for” the substantial involvement of nuclear weapons within one year prior to the event. One way to think of this rule is if the nuclear weapons involved in an event could have failed unexpectedly without dramatically reducing the probability of the event, then the nuclear weapons were not a proximate cause.
| Nuclear Extinction | 2030 | 2050 | 2100 |
| My beliefs | 0.0024% | 0.0072% | 0.0168% |
| My forecast of other experts’ beliefs | 0.08% | 0.36% | 1% |
| My forecast of other super forecasters’ beliefs | 0.08% | 0.36% | 1% |
Five thousand remaining humans is too low a number to move these probabilities too far from zero. Nuclear weapons are just a type of weapon, they are not doomsday devices. There are broadly two targeting regimes that nuclear weapons can full under, counter-force and counter-value. In counter-force the weapons are targeted at the weapons of your adversary, with the intent to destroy their nuclear and conventional military capability, following this strategy many fewer than the total number of worldwide nuclear weapons would be detonated in a nuclear conflict, many would be destroyed. Following a counter-value strategy where cities are targeted, nuclear detonations are not going to be distributed to cause the maximum global deaths but rather to inflict deaths on the adversarial nuclear countries and their allies. There is a very high likelihood that South America and Africa would not have any nuclear strikes and would not be subject to the same level of fallout initially as the northern hemisphere. It also remains unclear what the medium term nuclear climate impacts would be, especially with lower yield weapons and less total megatonnage. One would have to revise down the damage predicted by climate modeling built during the cold war.
| Year | # of Escalatory Events | Escalates to Nuclear War | Mass Death |
| 2030 | 2 | 0.96% | 0.48% |
| 2050 | 6 | 2.88% | 1.44% |
| 2100 | 14 | 6.72% | 3.36% |
An anonymous colleague commented on my initial justification:
I guess another conclusion one could draw from your reasoning looks like this: (1) it is overwhelmingly unlikely that ‘standard’ events involving nuclear weapons could lead to fewer than 5,000 people remaining, (2) so we can approximate the question as ‘what is the likelihood of a ‘non-standard’ event causing the deaths of all but 5,000 people?’. I think such non-standard events would include (a) doomsday cults actually setting out to cause extinction, and using nukes as one of their methods, and (b) unaligned AI systems doing the same. I don’t think either (a) or (b) is especially likely, but think the estimate should include the small amount of probability from each of these, as well as the small probability of standard events somehow going badly wrong.”
As part of the team collaboration component of this project, I responded in the discussion section:
I agree that (a) and (b) are unlikely, I would even say vanishingly unlikely for similar reasons. The command and control of nuclear weapons is tightly related to their value for any nation-state. If a state is unable to launch nuclear missiles on command then the weapons can’t function as a deterrent or a threat. The more you look into the details of launching nuclear missiles the more difficult it becomes, two examples: US nuclear submarines are communicated to by extremely low frequency to transmit in this range requires facilities more than 50km apart to create an antenna in very specific geological conditions – these communications require access to these facilities, an AI cannot just hack in.
This isn’t 100% certain today but several years ago there was good information that Pakistan physically separated missiles and warheads, at least during periods of low tension, again not exactly hackable. The only route that I can reasonably see for AI or doomsday cult is some sort of social engineering attack where all the nuclear-armed states give up hostilities between each other and then change their targeting packages to explicitly try to increase the death toll as much as possible, even then it is hard to believe that there wouldn’t be small groups of survivors all around the world (probably envying the dead), who would sum into the tens of thousands.
DeepForest.io 