This is a linkpost for Is Extinction Risk Mitigation Uniquely Cost-Effective? Not in Standard Population Models by Gustav Alexandrie and Maya Eden, which was published on 18 August 2025 as chapter 20 of the book Essays on Longtermism: Present Action for the Distant Future. I got Claude Opus 5 to write the summary and 2 objections below after some iteration. The objections came from me. Gustav said the summary "looks correct from very quickly skimming it, though I may well have missed if there are mistakes". I also think it is accurate based on my read of the article. Thanks to Gustav for feedback on the draft objections, and pushing me to shorten the summary, which now has 1/3 as many words as the main text of the chapter (excluding abstract, footnotes, references, and appendix).
TL;DR
Alexandrie and Eden take aim at the long-run argument for prioritising extinction risk mitigation: the Parfit-derived claim that the gap between a 99% catastrophe and a 100% catastrophe dwarfs the gap between no catastrophe and a 99% one, because extinction alone forecloses all future generations. They grant the argument's ethics entirely — total, time-neutral aggregation of welfare — and attack the demography it quietly presupposes.
The argument needs two empirical assumptions. Recovery: any catastrophe short of extinction leaves long-run population levels unchanged, because humanity bounces back. Priority of saving lives: the most cost-effective way to raise long-run population is a lifesaving intervention rather than something else, such as permanently expanding the supply of natural resources. Recovery fails in two of the three standard economic models of fertility; in the Malthusian model, the one where Recovery does hold, Priority of saving lives fails instead.
Their headline positive result needs no calibration. Because fertility decisions in standard models depend on the ratio of capital to labour rather than on the scale of the economy, any shock that reduces population and capital in the same proportion lowers long-run population permanently and proportionally. Mitigating such shocks is therefore exactly as valuable, in their framework, as mitigating extinction. Their challenge is one of non-uniqueness rather than of magnitude: extinction risk mitigation may be worth everything its advocates claim, and still not be the best buy.
1. The target argument
Parfit's comparison sets up the problem. Consider (i) no catastrophe, (ii) a catastrophe killing 99% of the world's population, and (iii) one killing 100%. Counting only the present generation, the (i)-to-(ii) gap dwarfs (ii)-to-(iii); counting all generations equally, Parfit argues, the reverse holds, because extinction forecloses every subsequent generation. Alexandrie and Eden call this the long-run argument; it has driven published cost-effectiveness analyses of asteroid risk (Matheny 2007), climate change (Ng 2016), and pandemics (Millett & Snyder-Beattie 2017).
They stylise it as:
- (P1) The social value of a population-affecting intervention is approximately proportional to how much it raises expected long-run population.
- (P2) Marginal philanthropic spending on extinction risk mitigation raises expected long-run population more than marginal spending on any other population-affecting intervention.
- (C) Therefore marginal spending on extinction risk mitigation is more socially valuable than marginal spending on any other population-affecting intervention.
"Population-affecting interventions" are those whose impact comes mainly through how many people or life-years exist rather than through quality of life — both lifesaving ones (bed nets, extinction risk mitigation) and non-lifesaving ones (shifting fertility norms, expanding natural resource supply).
(P1) rests on two claims they state, note the counterarguments to, and then grant: Generalized totalism (social value rises linearly in the number of good lives, regardless of when they exist) and Astronomical stakes (in expectation, the vast majority of lives lie in the far future, and are good). The entire chapter attacks (P2), which needs Recovery and Priority of saving lives.
2. The framework: the long-term value ratio
This is the chapter's most reusable contribution.
Two objectives. Write U for what real-world policymakers maximise in expectation: the expected number of current lives, normalised so the no-catastrophe baseline equals 1, so an event killing a quarter of the current population takes U to 0.75. Write W for what a longtermist maximises: the expected number of current and future lives, again normalised to 1, so 0.75 means a permanent quarter reduction in the sum of current and future population. Near-term extinction sets both to zero.
Policymakers. Each event's probability is endogenous, depending on money spent averting it through a "risk function" assumed decreasing (money lowers risk) and convex (diminishing returns). Policymakers maximise expected U subject to a fixed mitigation budget. Assuming an interior solution — something is spent on every risk — their optimum satisfies, for some positive constant λ,
C′i(m∗i)(Ui−U0)=λfor all events i(1)
where the starred term is optimal spending on event i and the derivative is the rate at which an extra dollar lowers its probability. In words: at the optimum, the marginal benefit of a dollar is equal across all risks, since otherwise money would flow to the better-returning risk until returns equalised.
The longtermist. A philanthropist with a small budget who cares about W has a marginal benefit equal to that same derivative times the change in W. Rearranging (1) to express the derivative in terms of λ and the change in U, then substituting, gives a marginal benefit proportional to what the authors call the long-term value ratio (LVR):
LVR=Wi−W0Ui−U0,
heuristically, proportional lives lost in the long run divided by proportional lives lost in the short run.
Why this is elegant. The policymakers' optimality condition prices out tractability: if myopic actors have already equalised marginal returns measured in short-run lives, differences in short-run cost-effectiveness have been arbitraged away and only the persistence ratio remains. Everything collapses to one number per risk, and no absolute risk estimates are needed anywhere.
Because extinction zeroes both objectives, extinction risk mitigation has an LVR of exactly 1 by construction. Recovery says the proportional long-run loss from any non-extinction shock is smaller than its proportional short-run loss, which rearranges to an LVR below 1. So under Recovery, nothing beats extinction risk mitigation. The rest of the chapter asks whether anything can reach or exceed 1.
3. Pure population shocks: three models
First, shocks that kill people but leave other factors of production intact. If 50% of people vanish, each survivor works with twice as much capital, so standard theory predicts higher wages and higher average wealth — matching the evidence that the Black Death raised living standards in late medieval Europe (Jedwab, Johnson & Koyama 2022). How does that feed back into fertility?
(a) The social determinants model. Fertility is set by cultural norms about ideal family size and is insensitive to wages and wealth. A 50% shock therefore leaves population permanently 50% lower than it would otherwise have been — the population can keep growing, but never closes the gap. Recovery fails. The LVR equals 1 for catastrophes of any size, so extinction mitigation is neither better nor worse than mitigating smaller catastrophes.
(b) The Malthusian model. Population is capped by a binding natural resource constraint. A shock leaves more resources per head, raising income; the income effect pushes fertility above replacement; population climbs back to its ceiling. Recovery holds — the only one of the three models that unequivocally delivers it, and hence the long-run argument's best hope. But then Priority of saving lives fails: permanently expanding resource supply permanently raises the steady-state population, so lifesaving is not the uniquely best lever.
(c) The Barro-Becker model (Becker & Barro 1988; Barro & Becker 1989), the workhorse for modern fertility, in which all capital is reproducible so there is no fixed resource ceiling. Parents value both consumption and children, so the model contains the income effect and a competing substitution effect: labour scarcity raises wages, raising the opportunity cost of time spent on children and pushing fertility down. If substitution dominates, a shock is amplified and the LVR exceeds 1. Their calibration finds substitution strengthens with shock size: below ~13% the steady-state drop is proportionally smaller than the shock (LVR under 1); above ~13% the shock is amplified (LVR above 1); the LVR peaks at just under 2.4 at a shock of roughly 35%. They read this only as an existence claim, and caution explicitly against drawing anything stronger from the calibration.
How plausible is the Malthusian model? Since it is the long-run argument's only support, they assess it directly. Against: it is broadly considered irrelevant to modern fertility, since industrialisation lifted resource constraints and fertility has fallen as income has risen. For: Malthusian dynamics might re-emerge, through evolutionary selection for higher fertility, or through AI-driven growth in machine labour and capital eventually hitting binding energy or land constraints. Against that in turn: fertility is now below replacement across most high- and middle-income countries.
4. Shocks to all factors of production: the robust result
This is the strongest part of the chapter: it needs no calibration and holds across models.
The economic determinants of fertility are invariant to the scale of the economy. Parents' choices depend on their own wealth and income, not on how many other people exist. Under constant returns to scale — doubling both workforce and capital exactly doubles output — per-capita wealth and income are set by the capital-to-labour ratio, not by the size of the economy. So shrinking people and capital proportionally leaves wages, rental rates, and therefore fertility decisions unchanged. Population stays permanently lower by exactly the shock size, so proportional long-run loss equals proportional short-run loss and the LVR is exactly 1.
Their example: a nuclear war kills 50% of the population and 50% of the capital stock. Capital-labour ratio unchanged → wages and rental rates unchanged → people have the same number of children they would have had → population permanently 50% lower than it would otherwise have been → LVR of 1, on par with extinction mitigation. And this isn't a contrivance: wars level factories and cities, and climate change destroys natural resources as well as lives.
5. The back-of-the-envelope calculation
To pre-empt the objection that no such intervention is purchasable, they construct one: save lives while proportionally increasing the capital stock, holding the capital-labour ratio constant.
Bed nets cost about $5,000 per life saved (GiveWell 2022); holding the ratio constant means also adding one person's worth of capital. Global GDP per capita is around $12,000 (World Bank, 2021), about two-thirds of it labour income, so capital income per capita is around $4,000. Since capital must earn enough to cover both the return investors require and annual depreciation — taking both at 5%, a marginal product of 10% — the implied per-capita capital stock is $4,000/0.1 = $40,000. Total: around $45,000 per scale-preserving life saved, of which ~89% is the capital transfer. The returns must accrue to the people whose lives were saved, since their fertility depends on wealth as well as wages, so implementation means bed nets plus a wealth transfer to that region — one consisting of resources that would otherwise have been consumed, not invested.
Spending $100 therefore buys $100/$45,000 ≈ 0.0022 of a scale-preserving life, a permanent proportional population increase of 0.0022/(8 × 10⁹) against a world population of 8 billion. Because the increase is permanent and proportional it scales the entire future, so using Greaves & MacAskill's (2021) estimate of 10²⁴ expected future lives, the lives gained in the long run per $100 spent are
0.0022×10248×109=2.75×1011.\frac{0.0022 \times 10^{24}}{8 \times 10^{9}} = 2.75 \times 10^{11}.8×1090.0022×1024=2.75×1011.
For comparison, Greaves & MacAskill estimate 2 × 10⁸ lives per $100 on biosecurity and 3 × 10⁵ per $100 on asteroid detection — about 1,000 times the biosecurity figure.
They disclaim this energetically, citing Karnofsky (2011) on not taking expected-value estimates literally: it requires Barro-Becker to hold indefinitely, and the intervention is in their own words unconventional and perhaps politically impractical. Their aim is to establish that at least one intervention exists whose LVR matches or exceeds that of extinction risk mitigation — not that this particular package is among the best uses of marginal philanthropic money.
6. Increasing the stock of natural resources
The calculation above assumes reproducible capital can substitute for natural resources. In the Malthusian model, where it can't, the analogue is direct: permanently increasing resource supply by some proportion permanently raises population by that proportion. So preventing permanent destruction of arable land raises short-run and long-run social value alike — an LVR of 1, again matching extinction mitigation.
Then the speculative extension. If humanity becomes intergalactic, the binding constraint is the number of reachable galaxies, and each year of delay permanently loses (2 × 10⁻⁸)% of the reachable universe to accelerating cosmic expansion (Ord n.d.b). An intervention accelerating expansion with no short-run benefits at all could therefore have an LVR greater than 1 — though only if policymakers already spend something on space expansion, or the derivation in §2 doesn't apply. They contrast this with prior literature, noting that it "generally finds that extinction risk mitigation is much more cost-effective than speeding up space expansion (Bostrom 2003; Ord n.d.b)." That is an acknowledgement of prior work rather than a load-bearing step, but more contestable than it looks; see §8.2 below.
7. Conclusions
- Shocks proportionally reducing all factors of production have proportional, permanent effects on long-run population, so mitigating them is comparably cost-effective to mitigating extinction — and the back-of-the-envelope calculation suggests the bed-nets-plus-wealth-transfer package could even exceed it. The challenge is thus one of non-uniqueness, and it bites hardest if fertility stays low or the exogenous extinction hazard is high, since Malthusian dynamics are then unlikely to govern.
- A corollary for catastrophe taxonomy: how much capital a catastrophe destroys matters as much as how many people it kills. Capital-destroying events (asteroids, wars) plausibly produce proportional long-run losses under any constant-returns model, independently of which fertility model is correct. Capital-sparing events (pandemics) are the model-dependent case: Malthusian predicts a disproportionately small long-run effect, social determinants an exactly proportional one, and calibrated Barro-Becker a small effect below ~13% but a disproportionately large one above it.
8. Two things I'd push on
8.1 The model's separability across events is doing real work
The framework treats each bad event as having its own mitigation budget and its own independent risk function. Real mitigation isn't like that. A dollar of pandemic preparedness lowers the probability of a mild epidemic, of a catastrophic pandemic, and of an extinction-level pandemic all at once. Two consequences follow.
Spending per event isn't well-defined. Condition (1) then can't be applied event-by-event, and the substitution that yields the LVR loses its footing: the policymakers' problem is no longer additively separable across events in the way the derivation needs. This is the kind of objection the authors would presumably grant without much resistance — the model is deliberately stylised, and a richer one could shift the numbers.
What you can actually buy is a shift in an entire severity distribution. So "mitigating a 35%-mortality catastrophe" is not an intervention distinct from "mitigating extinction"; both are severity-bundles purchased through spending on the same hazard class. The decision-relevant quantity is therefore not the LVR at any single severity, but something like a severity-weighted average over the density shift a real intervention produces.
But this shows less than it first appears to. The tempting next step is to say that the bundled LVR of non-extinction mitigation falls below 1 — since the calibrated LVR is under 1 for shocks below 13%, and most plausible hazard distributions put the bulk of their probability mass at low severities — and that this restores the uniqueness of extinction risk mitigation (recall from §2 that an LVR below 1 is exactly the condition under which nothing beats it). That inference doesn't go through, because the same bundling applies on the other side. Money labelled "extinction risk mitigation" also buys reductions in sub-extinction catastrophes whose LVRs differ from 1, so its effective LVR is a weighted average too, pulled away from 1 by the same low-severity mass.
So the honest conclusion is that bundling compresses differences between interventions rather than reordering them. It costs the LVR framework discriminating power in both directions at once: it makes it harder to establish that some non-extinction intervention beats extinction mitigation, and equally harder to establish that extinction mitigation is uniquely best. Showing that the ranking actually reverses would require an argument that the bundling effect is systematically larger on one side than the other — that hazard classes where you are paying to remove an extinction-level tail have differently shaped severity distributions from those where you are paying to remove a moderate tail. I don't have such an argument, and I'm not aware of one. Anyone wanting to press this objection has to supply it.
Two caveats on scope. First, this doesn't bear on the chapter's stated conclusion, which is more modest than the 35% peak in Figure 20.1B might suggest to a casual reader: Alexandrie and Eden present the calibration as showing only that there could exist mitigation efforts whose LVR exceeds 1, and explicitly caution against drawing anything stronger from it. The objection above is a warning about over-reading the figure, not a challenge to what they assert. Second, it leaves §4 untouched, since scale-invariance identifies a distinct type of intervention — one that changes the capital-labour ratio rather than the severity of a hazard — and needs no calibration at all.
8.2 A minor note on the space-expansion contrast
This is a small point about a single sentence, and nothing in the chapter's argument rests on it. Having noted that accelerating intergalactic expansion could in principle have an LVR above 1 — meaning it would be a better buy than extinction risk mitigation — the authors add that previous literature generally finds extinction risk mitigation much more cost-effective than speeding up space expansion (Bostrom 2003; Ord n.d.b). That is an acknowledgement of prior work, not a step in the argument. But it is worth registering that the acknowledgement is more contestable than it reads.
Bostrom (2003) is a genuine cost-effectiveness argument, not merely a claim about magnitudes: it is explicitly about what utilitarians ought to do, which necessarily weighs what a given effort can buy. What it isn't is an empirically grounded quantitative comparison of the kind Alexandrie and Eden themselves perform. The cosmological input — how much of the reachable universe is forfeited per year of delay — is well founded. The quantities that actually decide the comparison, namely how much extinction risk reduction or how much acceleration a given expenditure purchases, are stipulated rather than estimated.
More substantively, the finding isn't settled, and Ord supplies the reason himself. His contribution to this same volume ("Shaping Humanity's Longterm Trajectory", ch. 13) distinguishes four idealised ways short-term actions can alter the long-term trajectory, and the decisive property is whether a change scales with both the duration of humanity's future and its average value. Speed-ups and enhancements do, as does reducing extinction risk — putting them in one class. Pure advancements, reaching each value level some years earlier against a fixed end time, do not scale with duration, which is why they struggle to compete; Ord finds them competitive only where value rises exponentially right up to the end time, which he judges unlikely over the very long run. And his own footnote identifies Bostrom (2003) as comparing advancements to extinction risk reduction.
That distinction matters for the case at hand. A year of delay in intergalactic expansion permanently forfeits a fixed fraction of the reachable universe, which under the Malthusian model is a permanent proportional reduction in the resource base and hence in population — scaling with the entire future. In Ord's taxonomy that puts it on the competitive side of the line, rather than in the advancement category Bostrom analysed. So the prior literature sits less squarely against the chapter's own suggestion than the sentence implies.
References
Alexandrie, G. & Eden, M. (2025), 'Is Extinction Risk Mitigation Uniquely Cost-Effective? Not in Standard Population Models', in H. Greaves, J. Barrett & D. Thorstad (eds.), Essays on Longtermism (Oxford University Press), 343–361.
Armstrong, S. & Sandberg, A. (2013), 'Eternity in Six Hours: Intergalactic Spreading of Intelligent Life and Sharpening the Fermi Paradox', Acta Astronautica 89: 1–13.
Barro, R. J. & Becker, G. S. (1989), 'Fertility Choice in a Model of Economic Growth', Econometrica 57/2: 481–501.
Basten, S., Lutz, W. & Scherbov, S. (2013), 'Very Long Range Global Population Scenarios to 2300 and the Implications of Sustained Low Fertility', Demographic Research 28: 1145–1166.
Becker, G. S. (1988), 'Family Economics and Macro Behavior', American Economic Review 78/1: 1–13.
Becker, G. S. & Barro, R. J. (1988), 'A Reformulation of the Economic Theory of Fertility', Quarterly Journal of Economics 103/1: 1–25.
Blackorby, C., Bossert, W. & Donaldson, D. (1995), 'Intertemporal Population Ethics: Critical-Level Utilitarian Principles', Econometrica 63/6: 1303–1320.
Bostrom, N. (2003), 'Astronomical Waste: The Opportunity Cost of Delayed Technological Development', Utilitas 15/3: 308–314.
Bostrom, N. (2004), 'The Future of Human Evolution', in N. Bostrom, R. C. W. Ettinger & C. Tandy (eds.), Death and Anti-death (Ria University Press), 339–371.
Collins, J. & Page, L. (2019), 'The Heritability of Fertility Makes World Population Stabilization Unlikely in the Foreseeable Future', Evolution and Human Behavior 40/1: 105–111.
Cordoba, J. C. (2015), 'Children, Dynastic Altruism and the Wealth of Nations', Review of Economic Dynamics 18/4: 774–791.
Geruso, M. & Spears, D. (2025), 'Depopulation and Longtermism', in Greaves, Barrett & Thorstad (eds.), Essays on Longtermism.
GiveWell (2022), Our Top Charities.
Greaves, H. & MacAskill, W. (2021), 'The Case for Strong Longtermism', GPI Working Paper No. 5-2021.
Hanson, R. (2016), The Age of Em: Work, Love, and Life When Robots Rule the Earth (Oxford University Press).
Jedwab, R., Johnson, N. D. & Koyama, M. (2022), 'The Economic Impact of the Black Death', Journal of Economic Literature 60/1: 132–178.
Jones, C. I. (2022), 'The End of Economic Growth? Unintended Consequences of a Declining Population', American Economic Review 112/11: 3489–3527.
Karnofsky, H. (2011), 'Why We Can't Take Expected Value Estimates Literally (Even When They're Unbiased)', The GiveWell Blog.
Kolk, M. & Jebari, K. (2022), 'Sex Selection for Daughters: Demographic Consequences of Female-Biased Sex Ratios', Population Research and Policy Review 41: 1–21.
Korinek, A. & Stiglitz, J. E. (2018), 'Artificial Intelligence and its Implications for Income Distribution and Unemployment', in A. Agrawal, J. Gans & A. Goldfarb (eds.), The Economics of Artificial Intelligence: An Agenda (University of Chicago Press), 349–390.
Malthus, T. (1798), An Essay on the Principle of Population.
Matheny, J. G. (2007), 'Reducing the Risk of Human Extinction', Risk Analysis 27/5: 1335–1344.
Millett, P. & Snyder-Beattie, A. (2017), 'Existential Risk and Cost-Effective Biosecurity', Health Security 15/4: 373–383.
Ng, Y. K. (2016), 'The Importance of Global Extinction in Climate Change Policy', Global Policy 7/3: 315–322.
Ord, T. (2025), 'Shaping Humanity's Longterm Trajectory', in Greaves, Barrett & Thorstad (eds.), Essays on Longtermism.
Ord, T. (n.d.b), 'The Edges of Our Universe', arXiv:2104.01191.
Parfit, D. (1984), Reasons and Persons (Oxford University Press).
Saint-Paul, G. (2003), 'Economic Aspects of Human Cloning and Reprogenetics', Economic Policy 18/36: 73–122.
Spears, D. & Zuber, S. (2021), 'Foundations of Utilitarianism Under Risk and Variable Population', IZA Discussion Paper No. 14515.
Tarsney, C. & Thomas, T. (2020), 'Non-additive Axiologies in Large Worlds', GPI Working Paper No. 9-2020.
Thorstad, D. (2023), 'High Risk, Low Reward: A Challenge to the Astronomical Value of Existential Risk Mitigation', Philosophy & Public Affairs 51/4: 373–412.
