Thanks for your work here, and it’s always great to see people doing a deep dive on nuclear winter and abrupt sunlight reduction scenarios (ASRS). As Alliance to Feed the Earth in Disasters (ALLFED) we are highly concerned about these issues and certainly feel that they are neglected, and our analysis also suggests that the field is high impact and cost effective to mitigate.
However, there are a number of points we would like to raise, where we differ, at least in part, with your analysis:
We assign a higher probability that a nuclear conflict occurs compared to your estimates, and also assume that conditional on a nuclear conflict occurring that higher detonation totals are likely. This raises the likelihood and severity of nuclear winters versus your estimates.
The weightings of the commenters vs Metaculus is your choice, but we would suggest that a prediction market result should have a higher weighting, due to its aggregation of a large amount of expert opinions.
Your probability analysis excludes some high quality work (such as peer reviewed publications) which have a higher probability of nuclear conflict, potentially at 1% annually. This would primarily be via the risk of an inadvertent exchange, however the dynamics of an inadvertent exchange can be more damaging versus a deliberate conflict, as a response to an enemy launch is likely to focus more on critical infrastructure rather than weapon sites (which are assumed to have just fired at you).
The threshold for a catastrophic nuclear war in the XPT was very high—causing at least 10% of humanity to die over 5 years or less—and so should be considered as the probability of a nuclear conflict killing at least 800 million people, rather than a nuclear exchange.
However, your readers can also adjust this themselves in their heads reasonably easily if they wish, with a similar scale adjustment to the impact factor.
In terms of expected weapon detonations given at least 100, we also feel like your estimate of a uniform distribution is too low, and the logic of nuclear warfare suggests that “use it or lose it” would apply for at least the vulnerable land based weapon systems (including bombers). This pushes the distribution of detonations for a future NATO/Russia/China exchange towards the upper point of deployed weaponry, rather than a more uniform/skewed towards lower values distribution. This raises the expected severity of the following nuclear winter, if one occurs.
We estimate that the expected mortality from supervolcanic eruptions (VEI 8+) would be comparable to VEI 7 eruptions, so their inclusion could increase cost effectiveness significantly.
We feel that you are selling short the importance of research in building resilience to nuclear winters in particular and ASRSs in general (page 27, and page 41 onwards), possibly by conflating research with just one of its subsets (pilot schemes and field tests of resilient food technologies).
Research covers many activities, and a good amount of the analysis you link to in the report is based upon fundamental research of the likely dynamics of food consumption, production and trade, which did not exist before organizations like ALLFED started working on them.
These issues are highly complex and understudied, there is a significant risk of ineffective or even counterproductive actions if one rushes in without proper consideration, so new policy advocacy and engagement should result from careful consideration.
Research at ALLFED covers many different fields, for example analyzing nutrition and diets in these scenarios, the likely production/yields of these sources under extreme conditions, the cost of their production and the likely dynamics of trade, accessibility, pricing and storage. In addition, we are proposing gathering some experimental data or carrying out pilot studies in cases where it would generate useful insights or build capacity, but this is only part of the story.
The impact on the long-term future is likely to be relatively larger from the most extreme catastrophes than the mortality, which is a further reason that we focus on the larger scenarios. Of course some of this work could provide tangible benefits for tackling smaller scale events too.
For example, you highlight uncertainty about the impact of novel resilient foods in all but the largest scenarios, as they can only provide around 19% of global calories in a no-international food trade scenario. Research is a way of bridging this gap in understanding by getting to the core of where they could be useful. For example, where might prices go in a variety of scenarios? How resilient are the different food sources to the different shocks, and how much would they cost to produce? Can they integrate into the food system as feed or biofuels to free up human edible foods? A 1% shock to output leads to around a 7% rise in prices, being able to produce 19% more food at short notice is not a trivial factor in many crises of varying severity, and resilience like that could save millions of lives.
Some resilient foods are already cost effective for small quantities, such as seaweed and greenhouses, so they would be scaled up in lesser shocks. Also, we think of crop relocation to existing planted areas and crop area expansion as resilient foods, and these are likely to be a big part of the response in lesser catastrophes (and these are not included in the 19% figure).
Overall, we see research as the foundation on which you then build the policy work and other actions. Broadening and strengthening this foundation is therefore vital in allowing the work that finally effects change to occur—it isn’t an either/or. Now that there is a solid enough research base, it is possible to take some policy action, hence ALLFED’s expansion into this area of work, but more research will allow better and additional resilience-building in the future.
Thanks again for your work, and the openness with which it was conducted. It’s important to talk about and dig deep into these issues, and we hope others will do the same.
Firstly, thanks to you and all of ALLFED for your willingness to let me prod and poke at your work in the past year.
You make some excellent points and I think they will help readers to decide where they stand on the important cruxes here.
We assign a higher probability that a nuclear conflict occurs compared to your estimates, and also assume that conditional on a nuclear conflict occurring that higher detonation totals are likely. This raises the likelihood and severity of nuclear winters versus your estimates.
For anyone wanting to get up to speed on my nuclear winter model, plus a quick intro to why nuclear cooling is so uncertain, see my just-released nuclear winter post.
Weightings of Metaculus, XPT vs individual estimates: we do place high weightings on individual’s estimates, which is not ideal. The main reason for this is that the Metaculus and XPT estimates are not calibrated to a 100+ detonation nuclear conflict involving US and/or Russia.
The Metaculus estimate still forms around 20% of the total weight. Overall, it seems credible that my final estimate (0.10%) is around a third of the Metaculus estimate (0.30%) for 100+ detonations, given that most experts I spoke to considered India/Pakistan to be one of the main threats of a 100+ detonation nuclear exchange.
The XPT estimate only forms around 2% of the total weight.
Distribution of the number of detonations in a 100+ detonation conflict. I have done my best to incorporate different sources of evidence. One expert I spoke to broadly agreed with you, Mike, and guessed that in a 100+ conflict involving US/Russia there would be 80-90% risk of over 1000 detonations. Another expert was less pessimistic and placed higher weight on the possibility of ‘moderate’ escalation involving hundreds of weapons.
I re-ran the numbers just looking at conflicts with over 1000 detonations and found that the cooling levels increase only modestly: the 75th-percentile cooling level increases from 0.66 degrees (100+ detonations) to 1.24 degrees (1000+ detonations). The initial assumption about soot volumes is far more important (see below)
Detonation-soot relationship: Yes, the results are extremely sensitive to the initial assumption about how much soot is produced in a small nuclear exchange targeting cities. When I re-ran the numbers under the more pessimistic assumption that soot levels were 10%-100% those predicted by Toon et al., 75th-percentile cooling increased from 0.66 degrees to 3 degrees.
We estimate that the expected mortality from supervolcanic eruptions (VEI 8+) would be comparable to VEI 7 eruptions, so their inclusion could increase cost effectiveness significantly.
I don’t exclude supereruptions; I estimate that the right tail of my volcanic cooling model already accounts for them.
We feel that you are selling short the importance of research in building resilience to nuclear winters in particular and ASRSs in general [...] Overall, we see research as the foundation on which you then build the policy work and other actions. Broadening and strengthening this foundation is therefore vital in allowing the work that finally effects change to occur—it isn’t an either/or.
I want to be clear that I recommend that funders prioritize policy advocacy over R&D on the margin at this point in time. I totally agree that advocacy on such an uncertain topic can only be effective if it is grounded in research, and that ALLFED’s research will very likely form the foundations of policy work in this area for years to come.
One key takeaway from my analysis is that mild and moderate scenarios form a larger proportion of the threat than the lore of nuclear winter might suggest. Resilient foods would likely have a role to play in these scenarios, but I think the calories at stake in distribution and adaptation are likely to be more pivotal.
One reason for my focus on resilient food pilot studies is that they are a possible next step for ALLFED if it were to receive a funding boost. ALLFED has been ticking along with modest but reliable core funding for some time now, and perhaps I am guilty of taking its theoretical research for granted.
Feel free to set the record straight and give some indication of the kinds of work ALLFED might be interested in accepting funding for.
Michael—I agree with your assessment here, both that the CEARCH report is very helpful and informative, but also that their estimated likelihood of nuclear (only 10% per century) seems much lower than seems reasonable, and much lower than other expert estimates that I’ve seen.
Just as a lot can happen in a century of AI development, a lot can happen over the next century that could increase the likelihood of nuclear war.
I just wanted to clarify your “likelihood of nuclear (only 10% per century [9.44 % = 1 - (1 − 9.91*10^-4)^100])” refers to a nuclear conflict with at least 100 nuclear detonations involving China, the US and Russia, not just to the chance of at least 1 nuclear detonation (which would be higher).
Vasco—understood. The estimate still seems much lower than most other credible estimates I’ve seen. And much lower than it felt when we were living through the 70s and 80s, and the Cold War was still very much a thing.
Your probability analysis excludes some high quality work (such as peer reviewed publications) which have a higher probability of nuclear conflict, potentially at 1% annually.
To clarify, CEARCH’s estimated a probability of a conflict involving at least 100 nuclear detonations in the US, Russia or China of 0.0991 % per year, which is 9.91 % of the 1 % you mention. However, this refers to the probability of nuclear launch:
The US or Russia launching a nuclear weapon does not have to lead to 100 nuclear detonations in the US, China and Russia. The US or Russia could be attacking other countries, or the nuclear conflict fall short of escalating to at least 100 nuclear detonations. So I do not think the above estimates are obviously at odds with CEARCH’s.
The threshold for a catastrophic nuclear war in the XPT was very high—causing at least 10% of humanity to die over 5 years or less—and so should be considered as the probability of a nuclear conflict killing at least 800 million people, rather than a nuclear exchange.
I think the median particpant of The Existential Risk Persuasion Tournament (XPT) is very pessimistic about nuclear risk. The annual nuclear extinction risk from 2023 to 2050 respecting the median superforecaster and expert is 602 k and 7.23 M times mine of 5.93*10^-12.
It is also worth noting XPT’s forecasters varied a lot in their predictions. 6.21 % (10/161) forecasted a nuclear extinction risk from 2023 to 2100 of exactly 0 (which is obviously too low and wrong, but still illustrates my point).
ALLFED’s estimate suggest dividing by 46.9 (= 1⁄0.0213)?
David Denkenberger (ALLFED’s co-founder and research director)modelled the “percent of combustible material that burns that turns into soot” as a lognormal distribution with 2.5th and 97.5th percentiles equal to 1 % and 4 % (see Table 2), whose mean is 2.13 %.
Reisner 2018 consider the above fraction to be 1 (emphasis mine):
“Further, because the current version of FIRETEC assumes BC production to be inversely proportional to oxygen depletion (no soot model was employed), that is, all the carbon in the fuel participated in the reaction and was turned into BC, the estimates, which represent upper bounds for the given fuel loadings, are higher (worst case) than they would be if a detailed chemical combustion model was used for soot production”.
For readers’ reference, I have explained why I think Mike’s nuclear winter analysis is too pessimistic.
Dear Stan,
Thanks for your work here, and it’s always great to see people doing a deep dive on nuclear winter and abrupt sunlight reduction scenarios (ASRS). As Alliance to Feed the Earth in Disasters (ALLFED) we are highly concerned about these issues and certainly feel that they are neglected, and our analysis also suggests that the field is high impact and cost effective to mitigate.
However, there are a number of points we would like to raise, where we differ, at least in part, with your analysis:
We assign a higher probability that a nuclear conflict occurs compared to your estimates, and also assume that conditional on a nuclear conflict occurring that higher detonation totals are likely. This raises the likelihood and severity of nuclear winters versus your estimates.
The weightings of the commenters vs Metaculus is your choice, but we would suggest that a prediction market result should have a higher weighting, due to its aggregation of a large amount of expert opinions.
Your probability analysis excludes some high quality work (such as peer reviewed publications) which have a higher probability of nuclear conflict, potentially at 1% annually. This would primarily be via the risk of an inadvertent exchange, however the dynamics of an inadvertent exchange can be more damaging versus a deliberate conflict, as a response to an enemy launch is likely to focus more on critical infrastructure rather than weapon sites (which are assumed to have just fired at you).
The threshold for a catastrophic nuclear war in the XPT was very high—causing at least 10% of humanity to die over 5 years or less—and so should be considered as the probability of a nuclear conflict killing at least 800 million people, rather than a nuclear exchange.
However, your readers can also adjust this themselves in their heads reasonably easily if they wish, with a similar scale adjustment to the impact factor.
In terms of expected weapon detonations given at least 100, we also feel like your estimate of a uniform distribution is too low, and the logic of nuclear warfare suggests that “use it or lose it” would apply for at least the vulnerable land based weapon systems (including bombers). This pushes the distribution of detonations for a future NATO/Russia/China exchange towards the upper point of deployed weaponry, rather than a more uniform/skewed towards lower values distribution. This raises the expected severity of the following nuclear winter, if one occurs.
Soot lofting is very complicated and has serious uncertainties, but our estimates are that far higher levels are possible versus your estimates. Dividing Reisner 2018 (which cannot replicate real world firestorms) by 30 may be driving this, as well as your lower assumptions for detonations in a nuclear conflict.
We estimate that the expected mortality from supervolcanic eruptions (VEI 8+) would be comparable to VEI 7 eruptions, so their inclusion could increase cost effectiveness significantly.
We feel that you are selling short the importance of research in building resilience to nuclear winters in particular and ASRSs in general (page 27, and page 41 onwards), possibly by conflating research with just one of its subsets (pilot schemes and field tests of resilient food technologies).
Research covers many activities, and a good amount of the analysis you link to in the report is based upon fundamental research of the likely dynamics of food consumption, production and trade, which did not exist before organizations like ALLFED started working on them.
These issues are highly complex and understudied, there is a significant risk of ineffective or even counterproductive actions if one rushes in without proper consideration, so new policy advocacy and engagement should result from careful consideration.
Research at ALLFED covers many different fields, for example analyzing nutrition and diets in these scenarios, the likely production/yields of these sources under extreme conditions, the cost of their production and the likely dynamics of trade, accessibility, pricing and storage. In addition, we are proposing gathering some experimental data or carrying out pilot studies in cases where it would generate useful insights or build capacity, but this is only part of the story.
The impact on the long-term future is likely to be relatively larger from the most extreme catastrophes than the mortality, which is a further reason that we focus on the larger scenarios. Of course some of this work could provide tangible benefits for tackling smaller scale events too.
For example, you highlight uncertainty about the impact of novel resilient foods in all but the largest scenarios, as they can only provide around 19% of global calories in a no-international food trade scenario. Research is a way of bridging this gap in understanding by getting to the core of where they could be useful. For example, where might prices go in a variety of scenarios? How resilient are the different food sources to the different shocks, and how much would they cost to produce? Can they integrate into the food system as feed or biofuels to free up human edible foods? A 1% shock to output leads to around a 7% rise in prices, being able to produce 19% more food at short notice is not a trivial factor in many crises of varying severity, and resilience like that could save millions of lives.
Some resilient foods are already cost effective for small quantities, such as seaweed and greenhouses, so they would be scaled up in lesser shocks. Also, we think of crop relocation to existing planted areas and crop area expansion as resilient foods, and these are likely to be a big part of the response in lesser catastrophes (and these are not included in the 19% figure).
Overall, we see research as the foundation on which you then build the policy work and other actions. Broadening and strengthening this foundation is therefore vital in allowing the work that finally effects change to occur—it isn’t an either/or. Now that there is a solid enough research base, it is possible to take some policy action, hence ALLFED’s expansion into this area of work, but more research will allow better and additional resilience-building in the future.
Thanks again for your work, and the openness with which it was conducted. It’s important to talk about and dig deep into these issues, and we hope others will do the same.
Hi Mike,
Firstly, thanks to you and all of ALLFED for your willingness to let me prod and poke at your work in the past year.
You make some excellent points and I think they will help readers to decide where they stand on the important cruxes here.
For anyone wanting to get up to speed on my nuclear winter model, plus a quick intro to why nuclear cooling is so uncertain, see my just-released nuclear winter post.
Weightings of Metaculus, XPT vs individual estimates: we do place high weightings on individual’s estimates, which is not ideal. The main reason for this is that the Metaculus and XPT estimates are not calibrated to a 100+ detonation nuclear conflict involving US and/or Russia.
The Metaculus estimate still forms around 20% of the total weight. Overall, it seems credible that my final estimate (0.10%) is around a third of the Metaculus estimate (0.30%) for 100+ detonations, given that most experts I spoke to considered India/Pakistan to be one of the main threats of a 100+ detonation nuclear exchange.
The XPT estimate only forms around 2% of the total weight.
Distribution of the number of detonations in a 100+ detonation conflict. I have done my best to incorporate different sources of evidence. One expert I spoke to broadly agreed with you, Mike, and guessed that in a 100+ conflict involving US/Russia there would be 80-90% risk of over 1000 detonations. Another expert was less pessimistic and placed higher weight on the possibility of ‘moderate’ escalation involving hundreds of weapons.
I re-ran the numbers just looking at conflicts with over 1000 detonations and found that the cooling levels increase only modestly: the 75th-percentile cooling level increases from 0.66 degrees (100+ detonations) to 1.24 degrees (1000+ detonations). The initial assumption about soot volumes is far more important (see below)
Detonation-soot relationship: Yes, the results are extremely sensitive to the initial assumption about how much soot is produced in a small nuclear exchange targeting cities. When I re-ran the numbers under the more pessimistic assumption that soot levels were 10%-100% those predicted by Toon et al., 75th-percentile cooling increased from 0.66 degrees to 3 degrees.
I don’t exclude supereruptions; I estimate that the right tail of my volcanic cooling model already accounts for them.
I want to be clear that I recommend that funders prioritize policy advocacy over R&D on the margin at this point in time. I totally agree that advocacy on such an uncertain topic can only be effective if it is grounded in research, and that ALLFED’s research will very likely form the foundations of policy work in this area for years to come.
One key takeaway from my analysis is that mild and moderate scenarios form a larger proportion of the threat than the lore of nuclear winter might suggest. Resilient foods would likely have a role to play in these scenarios, but I think the calories at stake in distribution and adaptation are likely to be more pivotal.
One reason for my focus on resilient food pilot studies is that they are a possible next step for ALLFED if it were to receive a funding boost. ALLFED has been ticking along with modest but reliable core funding for some time now, and perhaps I am guilty of taking its theoretical research for granted.
Feel free to set the record straight and give some indication of the kinds of work ALLFED might be interested in accepting funding for.
Michael—I agree with your assessment here, both that the CEARCH report is very helpful and informative, but also that their estimated likelihood of nuclear (only 10% per century) seems much lower than seems reasonable, and much lower than other expert estimates that I’ve seen.
Just as a lot can happen in a century of AI development, a lot can happen over the next century that could increase the likelihood of nuclear war.
Hi Geoffrey,
I just wanted to clarify your “likelihood of nuclear (only 10% per century [9.44 % = 1 - (1 − 9.91*10^-4)^100])” refers to a nuclear conflict with at least 100 nuclear detonations involving China, the US and Russia, not just to the chance of at least 1 nuclear detonation (which would be higher).
Vasco—understood. The estimate still seems much lower than most other credible estimates I’ve seen. And much lower than it felt when we were living through the 70s and 80s, and the Cold War was still very much a thing.
Hi Mike,
To clarify, CEARCH’s estimated a probability of a conflict involving at least 100 nuclear detonations in the US, Russia or China of 0.0991 % per year, which is 9.91 % of the 1 % you mention. However, this refers to the probability of nuclear launch:
The US or Russia launching a nuclear weapon does not have to lead to 100 nuclear detonations in the US, China and Russia. The US or Russia could be attacking other countries, or the nuclear conflict fall short of escalating to at least 100 nuclear detonations. So I do not think the above estimates are obviously at odds with CEARCH’s.
I think the median particpant of The Existential Risk Persuasion Tournament (XPT) is very pessimistic about nuclear risk. The annual nuclear extinction risk from 2023 to 2050 respecting the median superforecaster and expert is 602 k and 7.23 M times mine of 5.93*10^-12.
It is also worth noting XPT’s forecasters varied a lot in their predictions. 6.21 % (10/161) forecasted a nuclear extinction risk from 2023 to 2100 of exactly 0 (which is obviously too low and wrong, but still illustrates my point).
ALLFED’s estimate suggest dividing by 46.9 (= 1⁄0.0213)?
David Denkenberger (ALLFED’s co-founder and research director) modelled the “percent of combustible material that burns that turns into soot” as a lognormal distribution with 2.5th and 97.5th percentiles equal to 1 % and 4 % (see Table 2), whose mean is 2.13 %.
Reisner 2018 consider the above fraction to be 1 (emphasis mine):
“Further, because the current version of FIRETEC assumes BC production to be inversely proportional to oxygen depletion (no soot model was employed), that is, all the carbon in the fuel participated in the reaction and was turned into BC, the estimates, which represent upper bounds for the given fuel loadings, are higher (worst case) than they would be if a detailed chemical combustion model was used for soot production”.
For readers’ reference, I have explained why I think Mike’s nuclear winter analysis is too pessimistic.