Hide table of contents

This is a writeup of the findings from the Causal Networks Model, created by CEA summer research fellows Alex Barry and Denise Melchin. Owen Cotton-Barratt provided the original idea, which was further developed by Max Dalton. Both, along with Stefan Schubert, provided comments and feedback throughout the process.

This is of a multipart series of posts explaining what the model is, how it works and our findings. We recommend you read the ‘Introduction & user guide’ post first before this post to give the correct background to our model. The structure of the series is as follows:

  1. Introduction & user guide (Recommended before reading this post)

  2. Technical guide (optional reading, a description of the technical details of how the model works)

  3. Findings (this post)

  4. Climate catastrophe (one particularly major finding)


The structure of this post is as follows:


  1. Summary of important findings

  2. Highlighted areas for further research

  3. Effects of specific inputs


Some of the results listed in the last section are fairly minor, so readers may wish to focus only on the first two sections.


The model is both very simplified, and many of the results depend to a large extent on particular variables with values about which we have very little information. Because of this, and because of the general limitations of the model, any findings should be taken as at most invitations to further research, rather than as concrete pronouncements of effectiveness. (In the past we have found a fair number of mistakes involving numbers being wrong by a few orders of magnitude!)


For the sake of space and readability we largely omit these qualifiers throughout the rest of the post, although will bring it up when results seem particularly uncertain.


We strongly recommend you read part I (particularly sections 2 and 5) to get appropriate background for the model before this post.


1. Summary of important findings

1.1 Steakless Salvation

Even with very small probabilities of success, research into developing cost-effective farmed meat alternatives (‘clean meat’) can be cost-competitive compared with other animal welfare alternatives.


This is due to the potential for ‘clean meat’ to gain a large proportion of the market share very quickly once it is lower in price but equal in quality to conventional meat. This means that it will scale far better than most animal interventions. Additionally this seems to be potentially a very effective way to reduce climate change, which has many other beneficial effects, as explained below.


This is true even for relatively limited forms of ‘clean meat’, as in our model we only consider the possibility of developing cost-competitive ‘clean’ ground meat, which seems much more attainable in the short term; this is still effective enough to seem plausibly better than conventional animal outreach.


This all assumes that the market is not actively hostile to clean meat, and all else equal will simply chose the cheaper option. This seems particularly applicable in the ground meat case.


1.2 Climate Catastrophe:

Climate change seems to be a much bigger problem than most people normally consider, both due to the potential damage caused by the Earth’s temperature rising 2-3 degrees, as well as the tail risk of runaway warming being a global catastrophic risk.


Unfortunately, many typical EA activities (e.g. giving more resources to the global poor, improving farmed animal conditions) probably cause increases in CO2 emissions, and so could potentially be negative overall due to the effects of climate change. This argument is particularly worrying if you think the potential of the far future morally dominates decision-making. For more discussion and elaboration on the climate x-risk connection see Part IV.


1.3 Cagefree Costs

One example of how considering climate change could cause an apparently positive intervention to be negative is corporate outreach focused on animal welfare. In particular this affects Mercy For Animals’s success in 2016 in getting large businesses to pledge to change from battery to cage-free eggs, affecting a total of 80 million laying hens a year. While this is a large win for animal welfare, cage-free hens are somewhat less efficient, causing more CO2e emissions per egg. This effect is large enough that for every year earlier MFA caused this to happen compared to when it would have happened otherwise, it will cost (very, very approximately) 500 QALYs due to death and disease from climate change before 2050.


This means that if you value a chicken-QALY at less than 1/20,000 of a human QALY, our model outputs this intervention as neutral, or even negative. [1]


1.4 Existential Effectiveness

Using our default estimates of the chance of existential risk and researchers’ ability to reduce it (or even estimates orders of magnitude lower), existential and global catastrophic risk research and policy work dominates other categories in terms of value. This is true even when comparing to other interventions in terms of QALYs saved before 2050. [2]

Therefore, you could justify giving to existential risk research charities even if you took a person-affecting view or strongly discounted future lives, as long as you put enough chance on research being able to reduce the risks. (That said, if you value guaranteed impact over high expected impact, then e.g. global poverty charities might still be more attractive).


1.5 Morality Matters

Many of the actions considered in the model end up being positive under some moral theories and negative under others. Examples include the farmed animal welfare case set out above, or actions that could negatively affect the far future while providing value today. This suggests that charity recommendations should be more dependant on the particulars of people's moral theories.


Despite seeming obvious when stated, this seems to be somewhat at odds with how the EA community actually operates, where charity evaluators etc. don’t really talk much about morality when giving recommendations.


1.6 Larder Logic

Whether one expects interventions that reduce the number of farmed animals to be positive or negative often depends on whether one thinks factory farmed cows have lives worth living. This is also true for other animals, but cows seem to the biggest case for disagreement.


2. Highlighted areas for future research

As discussed in the first post, many of the model’s results depend to a large extent on values we know very little about, and there are many important areas of the world we were not able to include in the model due to complexity or time constraints.


In particular, the model suggests that it would be very useful to learn more about the following areas:

  • How much can research actually reduce the chance of existential or global catastrophic risk?

  • What is the base rate of global catastrophic or existential risks?

  • What is the probability of clean meat becoming cheaper than normal meat in the near future (even just in limited forms, such as ground meat), and how much could additional funding increase this chance?

  • How well does veg*n outreach work?

  • Do factory farmed animals have net positive or negative lives? (This seems particularly unclear for cows.)

  • How negative are the effects of climate change likely to be, and what is the chance of extreme runaway global warming acting as an severe global catastrophic or existential risk?

  • How likely is it that EA will be able to influence government policy, and how strongly can policy affect the things we care about (e.g. x-risk, farmed animals)?


Some things that we were unable to capture in the model that also seem important to investigate include:

  • The interaction of outreach with (potentially) time-dependant areas, and accurately modelling movement growth more generally.

  • EA’s impact on government institutions and the effect this has or could have.

  • The value of cause prioritisation (in particular the likely distributions of cause effectiveness and how likely research is to find high-value causes).

  • Suffering and the environmental impact of fishing.

  • Wild animal suffering.

  • The long-term effects of economic growth on global institutions and technological progress.

  • The long-term effects of value changes.

  • Long-term effects more generally


3. Effects of specific inputs

In this section we go into detail about the effects of changing the different funding inputs on the user tool, how these effects depend on the ‘important variables’ and the impact of different moral interpretations of the results.


(In this section ‘good for climate change’ means reducing CO2 emissions etc.)


Veg*n outreach

  • Lowers farmed animal populations, which is in turn somewhat good for climate change and so has a minor positive impact on humans (our defaults give $375 per human QALY).

  • Can be negative overall if you think cows have positive lives and are significantly sentient.

  • Depends very strongly on the effectiveness of this veg*n outreach.


Government / corporate animal welfare reform

  • Exceedingly positive for animals in terms of increasing their welfare, regardless of how good you think their current lives are.

  • Strongly negative for humans (minus one QALY per $100 by default assumptions!) via significantly increased climate change, due to decreased efficiency of cage-free eggs etc. How bad this is strongly depends on how bad you think climate change will be (in particular its chance of being a global catastrophic or x-risk, and how much it will lower average quality of life).

  • Still ends up positive overall for views that value animals >~1/10,000 of a human.

  • Not very sensitive to any variables beyond value of animals vs humans, and those relating to climate change.

  • Additional uncertainty of how much these interventions actually help animals.


 Animal Product Alternatives

  • If you think funding animal product alternatives research has any real chance of speeding up or increasing the chance of cost-competitive cultured ground beef, then this ends up dominating veg*n outreach in terms of reducing the number of farmed animals.

  • As usual this means that if you think cows have net positive lives and are significantly sentient this may be net negative. (Assuming they are not replaced by something better!)

  • With default values this is also an exceedingly efficient way of reducing CO2 emissions, at a cost of just 2 cents per tonne of CO2, which also makes it a great way of saving QALYs, at $4 per QALY. If you are very sceptical about the value of x-risk research, this could also be the most cost-effective way of reducing x-risk (via climate change reduction).

  • Depends strongly on: the chance of cultured meat, chance of global catastrophic and x-risk, animal quality of life, importance of climate change.



  • Increased consumption slightly increases CO2, which then slightly reduces population and increases x-risk.

  • Still robustly positive on short-term total view due to increasing happiness, ending up with around $500 per QALY on default values.

  • To a small extent increases the number of farmed animals, which is bad if you think they have net-negative lives, but this is swamped by the positive effects on humans under all weightings that value animals less than or equal to humans (and no matter how bad you assume animal lives are on our scale).

  • Seems robustly good (ignoring far future effects) and not very sensitive to any variables.



  • Essentially just operates as a multiple of GiveDirectly, depending on the “effect of childhood deworming on future income” variable, so the comments on GiveDirectly apply here wholesale.

  • For the default values it ends up being around 30 times better than GiveDirectly, giving $16 per human QALY, due to the potential of deworming massively boosting future income.

  • Again this is entirely dependant on the “effect of deworming on lifetime earnings” variable.


Against Malaria Foundation

  • Effectively simply saves QALYs at AMF’s standard rate, and very slightly increases population and wellbeing.

  • Increases number of animals being farmed but by negligible amounts.

  • QALY effects depend strongly on number of QALYs given to preventing the death of a child under 5, but otherwise robust.


EA Outreach

  • Not currently well-implemented, basically just acts as multiplier funding other areas.

  • Almost all value with our default settings comes from x-risk, as this generally dominates.

  • Sensitive to all variables to different degrees, with x-risk variables having the biggest effect.


Cause prioritisation

  • Not currently integrated at all: ended up seeming like a separate project to model well. While we came up with a model for cause prioritisation, we needed better data on things like the underlying distribution of cause effectiveness and researchers’ ability to discover cause effectiveness.

  • This does however seems potentially very important, and we would like to see it studied further.


Global catastrophic and x-risk policy and strategy

  • Very strong x-risk reduction effects under default values (0.13% points reduction per million dollars) which leads to increased expected population. This can be negative in the total view if you think most humans have neutral or negative lives.

  • Depends strongly on global catastrophic and x-risk base chance, and how much government policy can influence x-risk.

  • Superiority over other x-risk reduction paths is entirely dependant on very uncertain (and quite arbitrary) values for how much policy effects x-risk etc.


 Global catastrophic and x-risk research

  • Medium x-risk reduction under default values (0.05% points reduction per million dollars), with same implications as above.

  • Depends strongly on base chance of global catastrophic and x-risk, and how much research reduces x-risk.


 Far future (global catastrophic and x-risk) outreach

  • Small x-risk reduction under default values (0.03% points reduction per million dollars), with same implications as above. Under default assumptions almost all impact comes from the effects of government policy, not research, but this is very uncertain (as above).

  • Depends strongly on the base chance of global catastrophic and x-risk, and how much how much government policy can influence x-risk.


General animal outreach / far future / global poverty

  • These nodes function simply by adding funding to the specific areas in their cause, e.g. funding Global Poverty just funds AMF, Givedirectly and Deworming.

  • Their effects are just combinations of the effects of the nodes discussed earlier, often dominated by whichever area happens to be largest. Discussion of the impacts of funding these ‘cause level’ nodes will be excluded as we can look at the specific areas within causes for a more complete picture.


Our next post will be a writeup of one particular finding of the model, the climate change catastrophic risk connection findings of the model which constitutes Part IV of the series.


Feel free to ask questions in the comment section, or email us (denisemelchin@gmail.com or alexbarry40@gmail.com).


[1]: Non-caged hens seem to emit about 15-25% more greenhouses gases per egg (due to increased food, heating and land requirements and a higher proportion of eggs being lost), which results in around 500,000 tonnes more CO2 emitted per year. Our very rough estimates lead to the conclusion that approximately each additional 1,000 Tonnes of CO2e emitted will cause the loss of one QALY before 2050.


[2]: Default assumptions used in our model were 7% x-risk chance by 2050, and 10,000 researches working for a decade could half x-risk to 3.5%.Hence 1 researcher year reduces risk by 0.000035% (percentage points) and say one researcher year costs $50,000, so 7*10^-10 percentage points reduction in x-risk per $. If extinction costs 25*7 billion = 175 billion QALYS, multiplying out gives just over 1 QALY saved per $.


Sorted by Click to highlight new comments since:

Thanks for the write-up!

I found the figures for existential-risk-reduced-per-$ with your default values a bit suspiciously high. I wonder if the reason for this is in endnote [2], where you say:

say one researcher year costs $50,000

I think this is too low as the figure to use in this calculation, perhaps by around an order of magnitude.

Firstly, that is a very cheap researcher-year even just paying costs. Many researcher salaries are straight-up higher, and costs should include overheads.

A second factor is that having twice as much money doesn't come close to buying you twice as much (quality-adjusted) research. In general it is hard to simply pay money to produce more of some of these specialised forms of labour. For instance see the recent 80k survey of willingness to pay of EA orgs to bring forward recent hires, where the average willingness to forgo donations to move a senior hire forward by three years was around $4 million.

Ah good point on the researcher salary, it was definitely just eyeballed and should be higher.

I think a reason I was happy to leave it low was as a fudge to take into account that the marginal impact of a researcher now is likely to be far greater than the average impact if there were 10,000 working on x-risk, but I should have clarified that as a separate factor.

In any case, even adjusting the cost of a researcher up to $500,000 a year and leaving the rest unchanged does not significantly change the conclusion, with the very rough calculation still giving ~$10 per QALY (but obviously leaves less wiggle room for skepticism about the efficacy of research etc.)

Indeed, the Oxford Prioritisation Project found cost-effectiveness about an order of magnitude lower than yours for AI. But still it was more cost-effective than global poverty interventions even in the present generation. And alternate foods for agricultural catastrophes are even more cost effective for the present generation.

Curated and popular this week
Relevant opportunities