A critique of EA’s focus on longtermism
Hello everyone,
A couple of days ago I had to present a critique of EA for an in-depth course and to try to steelman the argument as much as possible. I was encouraged to post it here as well, perhaps it can spawn some interesting discussion. It is not necessarily a terribly new critique, nor is it very original; nevertheless, I believe the arguments are pretty solid and the critique might convince some within EA to shift focus on some of its cause areas.
The summary of my critique of EA runs as follows: EA has placed too large of an emphasis on longtermism as a central (though broad) cause area; EA would benefit from a return to its roots in the form of less-uncertain global health/poverty, animal welfare and (arguably) climate change interventions.
For people who are not so familiar with longtermism, one common definition is that longtermism entails believing that positively influencing the long-term future is a (or the) key moral priority of our time.
Some of the theoretical/philosophical basis for longtermism can be found in books by influent EA figures. For example, in What We Owe the Future, Will MacAskill illustrates the moral equality of future and present people through a simple thought experiment. Suppose you leave a broken bottle in a forest. If someone cuts their hand on it tomorrow, you would rightly feel responsible for their injury. But if, instead, the bottle remains there for a hundred years and someone then gets hurt, the moral situation is unchanged: the harm is the same, and your causal role is identical. The time delay does not diminish the victim’s suffering or your moral responsibility. MacAskill’s point is that temporal distance alone cannot justify discounting the interests of future individuals. A person’s location in time, like their location in space, is morally irrelevant.
I do agree with this framing and with the proposition that, in principle, the lives of future, potential people count equally with those of present, actual people. The key element is the „in principle” part, which might become different in practice given the uncertainties involved. I am also open to theories whereby we apply an discount rate on the value of future people that grows with the time separating us from them, as long as the discount rate is not too steep. I am genuinly agnostic on which of these approaches is better or „more true”; however, I am quite confidently opposed to any moral theories that assigns zero moral worth to future, potential people. The arguments of my critique of EA makes no use of such assumptions.
To summarize, I would like to make it perfectly clear what I am not claiming here: I am not claiming that the lives of future people count for less than that of actual people. Instead, I am simply claiming that there is more uncertainty about both the probabilities of the individual causes affecting the survival of humanity, and uncertainty about whether we can be effective in doing something about it. Because longtermist interventions have extremely high variance and weak feedback loops, their expected-value estimates are often dominated by speculative assumptions rather than empirical validation. I am also not claiming that longtermism should stop being a focus inside EA altogether. I am simply arguing for a shift in focus in EA on other issues.
My argument can be broken down into three parts:
Epistemic uncertainty: We have low confidence in estimates of the likelihood of specific existential risks. Toby Ord’s The Precipice explicitly acknowledges that uncertainty in probability estimates of extinction risk is vast. At the same time, Ord admits some estimates (like AI extinction risk) are based more on expert judgment than hard empirical data. The epistemologist Christian Tarsney and the Global Priorities Institute have noted that longtermist decision-making involves deep uncertainty. All of these considerations should give us more epistemic humility regarding our understanding of the risks involved.
Tractability: Even if such risks are real, interventions may not be effective given their geopolitical or technical nature. I am going to focus on three popular causes within EA’s longtermist approach: extinction risk from nuclear war, man-made pandemics or artificial superintelligence.
Disclaimer: I am not an expert in any of the fields I am going to talk about, so I might be wrong about technical details. If you know more about these fields than I do and think my arguments break down somewhere, please let me know!
Nuclear war: it seems to me that any money donated to NGOs that aim to reduce nuclear proliferation can achieve little in a world where currently the vast majority(about 90%[1]) of nuclear weapons are in the hands of two arguably mentally unstable sociopaths: I am talking of course about Vladimir Putin and Donald Trump. Since they have some veto power over the use of these weapons, and anyway the circle of people making such decisions is small, the average EA can do very little to reduce probability of extinction via nuclear war. Historical research shows that most “near-miss” nuclear incidents in history were prevented by luck or individual restraint (e.g., Vasili Arkhipov, Stanislav Petrov), not by NGO intervention.
Man-made pandemic risk: this is also an area where the danger can arguably come from a small team of individuals that are hard to be controlled or regulated by the average EA donor. That is because risks arise from small, uncontrolled actors (e.g., rogue labs, state bioweapons programs). I actually happen to think it is probably a very good idea to regulate gain-of-function research. Gain-of-function research involves deliberately enhancing the transmissibility or virulence of pathogens in laboratory settings to study how dangerous variants might evolve. While scientifically informative, it carries the risk that an engineered strain could escape containment, potentially causing a pandemic. Laboratory leaks have occurred multiple times in the past: for example, smallpox escaped from a UK laboratory in 1978, killing a medical photographer, and SARS-CoV leaked from research facilities in Beijing in 2004, infecting several laboratory workers.
It probably is a good idea to regulate gain-of-function research; this however necessitates broad political action and international cooperation, both of which are hard to achieve using standard EA tools like donating to broad existential risk funds. The main levers are policy regulation and international treaties, both notoriously slow and resistant to small-scale donor influence. As you can notice, my arguments focus mostly on donations and earning to give; they might also apply to some degree to those EA members who try to help the world through their career.
AI risk: some of our current most advanced models are LLMs like ChatGPT, which are based on deep learning neural networks. From a scientific, epistemic perspective, the problem with neural networks is that computer scientists themselves don’t even understand how the black-box neural network arrives at its results. The decision processes within large neural networks are highly opaque and poorly understood in practice, making mechanistic interpretability a major open research challenge. This to me makes AI alignment research a very difficult endavour indeed. Moreover, there is as of now very little evidence of agentic AI having goals of its own. Current AI systems are non-agentic pattern predictors; no empirical evidence yet of goal-oriented autonomy. Even alignment researchers admit they are still defining what “misalignment” concretely means. Finally, the salaries of AI safety researchers tend to be very large, especially in labs such as Anthropic, OpenAI, or DeepMind, while the payoff is arguably very small: based on these reasons alone, giving to AI safety seems to me hardly cost effective. AI safety research often requires highly specialized technical expertise, making it relatively expensive compared with direct global health interventions. I am aware that longtermists attempt to make such donations cost-effective by using expected value calculations; any small and uncertain probability of averting extinction can yield great value if multiplied by a large enough benefit in terms of the future worth of humanity. Nevertheless, my understanding is that such calculations are highly speculative and can be made to show whatever one wants by simply manipulating the values inserted into the calculation: since such values are expert opinions at best (meaning an argument from authority) and subjective evaluations at worst, I do not think very highly of them. Such expected value calculations are also vulnerable to Pascal’s mugging, a deeply technical problem within decision theory that has, as far as I know, no non-controversial solution.
3. Neglectedness reconsidered: EA has overcorrected toward longtermism, leaving its proven “roots” (global health, poverty reduction, animal welfare) underfunded. Funding data from Open Philanthropy show that global health and development programs receive a smaller fraction of EA-aligned grants today than existential risk mitigation, despite the former having far stronger empirical track records.[2]
Finally, my pragmatic approach is to say that average small-time EA donors should direct their money towards proven, cheap interventions into global health and poverty, animal welfare etc, rather than into speculative existential risk funds which are most likely not cost-effective.
My conclusion is a portfolio diversification argument: Given extreme uncertainty and low tractability of existential risk mitigation, small donors should allocate primarily to global health, poverty, and animal welfare, while large donors/foundations can diversify into speculative, high-variance longtermist causes.
As you can see, I am mainly focusing on the day-to-day donation decisions of average EA small-time donors such as myself. I am someone who is not going to earn huge sums of money throughout my life, and therefore I am quite risk averse when it comes to where I want my charitable giving to go. If I were to win the lottery tomorrow, or be in a position to give advice to a billionaire or even millionaire, I would definitely change some of the focus of my critique and encourage them to speculatively donate to more uncertain existential risks areas. Nevertheless I think that the epistemic focus on longtermism from the EA community in the past couple of years, which perhaps stems from the fact that important figures like Will MacAskill, Toby Ord and Nick Bostrom have books on such topics, is not ideal.
Notice that I have left out climate change from my criticisms of longtermist causes. There is one reason for that: the three cause areas I singled out (nuclear, biorisk, and AI) can be characterized as involving the potential for existential risk coming from a very small team of people. Such localized, agent-dependent threats are very difficult to prevent by the average EA donor without broad political intervention and international cooperation between countries. Climate change is an issue that also requires broad political intervention and international cooperation, but the threat is no longer coming from a small number of hard-to-control people, i.e. it is not agent-dependent. The solutions are also much clearer and more tractable in the case of climate change and other environmental concerns.[3] These are some of the reasons why I think tackling climate change should remain a core EA cause.
To close, in decision theory, when uncertainty overwhelms our ability to compare options, a cautious strategy favoring robustly good actions over speculative gambles becomes rational; especially for small donors.
Finally, I would argue that my criticisms places me in rather good company: in a Q&A, Peter Singer also expressed concern with some of EA’s focus on longtermism to the expense of present cause areas.
[1] Federation of American Scientists (FAS), Nuclear Notebook 2024
[2] Open Philanthropy 2023 Annual Report
[3] Drawdown.org and 80,000 Hours both list climate change as moderately tractable and more tractable than AI or nuclear risk.
The recent anthology Essays on Longtermism, which is open access and free to read here, has several essays with good criticisms of longtermism. You might find some of those essays interesting. The authors included in that anthology are a mix of proponents of longtermism and critics of longtermism.
This is not necessarily to disagree with any of your specific arguments or your conclusion, but I think for people who have not been extremely immersed in effective altruist discourse for years, what has been happening with effective altruism over the last 5-10 years can easily be mis-diagnosed.
In the last 5-10 years, has EA shifted significantly toward prioritizing very long-term outcomes (i.e. outcomes more than 1,000 years in the future) over relatively near-term outcomes (i.e. outcomes within the next 100 years)? My impression is no, not really.
Instead, what has happened is that a large number of people in EA have come to believe that there’s more than a 50% chance of artificial general intelligence being created within the next 20 years, with many thinking there’s more than a 50% chance of it being created within 10 years. If AGI is created, many people in EA believe there is a significant risk of human extinction (or another really, really bad outcome). “Significant risk” could mean anywhere from 10% to over 50%. People vary on that.
This is not really about the very long-term future. It’s actually about the near-term future: what happens within the next 10-20 years. It’s not a pivot from the near-term to the very long-term, it’s a pivot from global poverty and factory farming to near-term AGI. So, it’s not really about longtermism at all.
The people who are concerned about existential risk from near-term AGI don’t think it’s only a justified worry if you account for lives in the distant future. They think it’s a justified worry if you only account for people who already alive right now. The shift in opinion is not anything to do with arguments about longtermism, but about people thinking AGI is much more likely much sooner than they previously did, and also them accepting arguments that AGI would be incredibly dangerous if created.
The pivot in EA over the last 5-10 years has also not, in my observation, been a pivot from global poverty and factory farming to existential risk in general, but a pivot to only specifically existential risk from near-term AGI.
To put my cards on the table, my own personal view is:
Longtermism is, in principle, correct (i.e. all else being equal, future lives matter as much as present lives or past lives and, if we can, we should try to ensure there are a lot of good future lives), but even after years of discussion, it seems really hard for anyone to give a good example of what actions longtermism justifies that we wouldn’t have already been doing anyway — other than actions related to existential risk, the discussion of which predates the discussion of longtermism by many years.
Whereas many people in EA seem to think the probability of AGI being created within the next 7 years is 50% or more, I think that probability is significantly less than 0.1%.
The arguments that AGI would be extremely dangerous rely on assumptions about the underlying technologies used to create AGI. I don’t think we know yet what those technologies will be, so I don’t accept these arguments. I’m not necessarily saying I know for sure AGI won’t be dangerous, either, I’m just saying knowing for sure would require information we don’t have.
I miss when EA was much more focused on global poverty and factory farming. (So, I agree with you there.)
Some existential risks are probably neglected, but not existential risk from near-term AGI. I think the world most likely still underinvests in global catastrophic risks of the humans vs. nature variety: asteroids, large volcanos, and natural pandemics. These risks are easier to calculate rigorous probabilities for.
The x-risks you discussed in your post are humans vs. humans risks: nuclear war, bioweapons, and the humans creating AGI. These are far more complex. Asteroids don’t respond to our space telescopes by attempting to disguise themselves to evade detection. But with anything to do with humans, humans will always respond to what we do, and that response is always at least somewhat unpredictable.
I still think we should do things to reduce the risk from nuclear war and bioweapons. I’m just saying that these risks are more complex and uncertain than risks from nature. So, it’s more harder to do the cost-effectiveness math that shows spending to reduce these risks is justified. However, so much in the world can’t be rigorously analyzed with that kind of math, so that’s not necessarily an argument against it!
As for climate change, I agree it’s important, and maybe some people in EA have done some good work in this area — I don’t really know — but it seems like there’s already so much focus on it from so many people, many of whom are extremely competent, it’s hard to see what EA would contribute by focusing on it. By contrast, global poverty charity effectiveness wasn’t a topic many people outside of international development thought about — or at least felt they could do anything about — before GiveWell and effective altruism. Moreover, there wasn’t any social movement advocating for people to donate 10% of their income to help the global poor.
Are you willing to bet on this?
In principle, yes, but in a typical bet structure, there is no upside for the person taking the other side of that bet, so what would be the point of it for them? I would gladly accept a bet where someone has to pay me an amount of money on January 1, 2033 if AGI isn’t created by then (and vice versa), but why would they accept that bet? There’s only downside for them.
Sometimes these bets are structured as loans. As in, I would loan someone money and they would promise to pay me that money back plus a premium after 7 years. But I don’t want to give a stranger from another country a 7-year loan that I wouldn’t be able to compel them to repay once the time is up. From my point of view, that would just be me giving a cash gift to a stranger for no particularly good reason.
There is Long Bets, which is a nice site, but since everything goes to charity, it’s largely symbolic. (Also, the money is paid up by both sides in advance, and the Long Now Foundation just holds onto it until the bet is resolved. So, it’s a little bit wasteful in that respect. The money is tied up for the duration of the bet and there is a time value of money.)
Right, I’d forgotten that betting on this is hard. I was thinking if one could do a sort of cross-over between an end-of-the-world bet crossed with betting a specific on a proportion of one’s net worth. This is the most fleshed-out proposal I’ve seen so far.
I wonder if this could be solved via a trusted third person who knows both bettors. (I think there are possible solutions here via blockchains, e.g. the ability to unilaterally destroy an escrow, but I guess that’s going to become quite complicated, not worth the setup, and using a technology I guess you’re skeptical of anyway)
You could bet on shorter-term indicators e.g. whether the METR trend will stop or accelerate.
Are you referring to the length of tasks that LLMs are able to complete with a 50% success rate? I don’t see that as a meaningful indicator of AGI. Indeed, I would say it’s practically meaningless. It truly just doesn’t make sense an indicator of progress toward AGI. I find it strange that anyone thinks otherwise. Why should we see that metric as indicating AGI progress anymore than, say, the length of LLMs’ context windows?
I think a much more meaningful indicator from METR would be the rate at which AI coding assistants speeds up coding tasks for human coders. Currently, METR’s finding is that it slows them down by 19%. But this is asymmetric. Failing to clear a low bar like being an unambiguously useful coding assistant in such tests is strong evidence against models nearing human-level capabilities, but clearing a low bar is not strong evidence for models nearing human-level capabilities. By analogy, we might take an AI system being bad at chess as evidence that it has much less than human-level general intelligence. But we shouldn’t take an AI system (such as Deep Blue or AlphaZero) being really good at chess as evidence that it has human-level or greater general intelligence.
If I wanted to settle for an indirect proxy for progress toward AGI, I could short companies like Nvidia, Microsoft, Google, or Meta (e.g. see my recent question about this), but, of course, those companies stock prices’ don’t directly measure AGI progress. Conversely, someone who wanted to take the other side of the bet could take a long position in those stocks. But then this isn’t much of an improvement on the above. If LLMs became much more useful coding assistants, then this could help justify these companies’ stock prices, but it wouldn’t say much about progress toward AGI. Likewise for other repetitive, text-heavy tasks, like customer support via web chat.
It seems like the flip side should be different: if you do think AGI is very likely to be created within 7 years, shouldn’t that imply a long position in stocks like Nvidia, Microsoft, Google, or Meta would be lucrative? In principle, you could believe that LLMs are some number of years away from being able to make a lot of money and at most 7 years away from progressing to AGI, and that the market will give up on LLMs making a lot of money just a few years too soon. But I would find this to be a strange and implausible view.
So, to be clear, you think that if LLMs continue to complete software engineering tasks of exponentially increasing lengths at exponentially decreasing risk of failure, then that tell us nothing about whether LLMs will reach AGI?
I expect most EAs who have enough money to consider investing them to already be investing them in index funds, which, by design, long the Magnificent Seven already.
I’m not sure if you’re asking about the METR graph on task length or about the practical use of AI coding assistants, which the METR study found is currently negative.
If I understand it correctly, the METR graph doesn’t measure an exponentially decreasing failure rate, just a 50% failure rate. (There’s also a version of the graph with a 20% failure rate, but that’s not the one people typically cite.)
I also think automatically graded tasks used in benchmarks don’t usually deserve to be called “software engineering” or anything that implies that the actual tasks the LLM is doing are practically useful, economically valuable, or could actually substitute for tasks that humans get paid to do.
I think many of these LLM benchmarks are trying to measure such narrow things and such toy problems, which seem to be largely selected so as to make the benchmarks easier for LLMs, that they aren’t particularly meaningful.
In terms of studies of real world performance like METR’s study on human coders using an AI coding assistant, that’s much more interesting and important. Although I find most LLM benchmarks practically meaningless for measuring AGI progress, I think practical performance in economically valuable contexts is much more meaningful.
My point in the above comment was just that an unambiguously useful AI coding assistant would not by itself be strong evidence for near-term AGI. AI systems mastering games like chess and go is impressive and interesting and probably tells us some information about AGI progress, but if someone pointed to AlphaGo beating Lee Seedol as strong evidence that AGI would have been created within 7 years of that point, they would have been wrong.
In other words, progress in AI probably tells us something about AGI progress, but just taking impressive results in AI and saying that implies AGI within 7 years isn’t correct, or at least it’s unsupported. Why 7 years and not 17 years or 77 years or 177 years?
If you assume whatever rate of progress you like, that will support any timeline you like based on any evidence you like, but, in my opinion, that’s no way to make an argument.
On the topic of betting and investing, it’s true that index funds have exposure to AI, and indeed personally I worry about how much exposure the S&P 500 has (global index funds that include small-cap stocks have less, but I don’t know how much less). My argument in the comment above is simply that if someone thought it was rational to bet some amount of money on AGI arriving within 7 years, then surely it would be rational to invest that same amount of money in a 100% concentrated investment in AI and not, say, the S&P 500.
The argument AI safety work is more cost-effective than AMF when considering only the next few generations is pretty weak.
Doesn’t that still depend on how much risk you think there is, and how tractable you think interventions are?
I think it’s still accurate to say that those concerned with near term AI risk think it is likely more cost effective than AMF.
This is, of course, sensitive to your assumptions.