This personal post is not part of the competition, as my intent is not to produce an original constructive response to cluelessness. However, Toby Tremlett encouraged me to share it insofar as it contributes to this week’s conversation.
TL; DR: I work in animal welfare and am clueless.[1] To avoid choosing actions at random, I decided to write a decision procedure for myself and to indicate the assumptions that accompany it. This does not constitute a new argument on the topic. Rather, this is a layman’s attempt to apply it. For those who want to skim, you can skip to the Decision procedure section, where I bolded the aspects of my decision procedure that seem most likely to elicit pushback.
Epistemic status and intent
Like many in animal welfare, I intuitively agree with cluelessness without understanding all of the philosophical steps behind it, and this will show in the post.[2] I focused on honestly representing my understanding of cluelessness, rather than high-quality philosophical reasoning.[3] I’m partly betting on Cunningham’s Law here.
This post does not aim to be a constructive general response like bracketing. It is also mostly derived from other writings on cluelessness. It assumes surface-level familiarity with bracketing.
However fragile and open to revision, this is a procedure I have already begun acting on. Since it has real-world consequences[4], it’s worth exposing it to criticism. And I’d love for the competition to give me ideas for a better procedure.
While the decision procedure and the list of assumptions already existed in an unpolished form, the post itself was time-capped at a day.
What Cluelessness Means To Me, a Layman
“Cluelessness”, to me, means that, when choosing how to act, my decision can’t actually track better expected effects on all of the beings which I care about.
I work in animal welfare, where cluelessness is both salient[5] and low-status. In EA animal welfare, action-bias rules[6]. The main ick around cluelessness in that field is not that it may justify repugnant conclusions[7], but that it licenses inaction.
But cluelessness doesn’t recommend inaction: it says that no action is impartially better-justified than another. This means that something else has to drive our decisions. But what is that? For the past year, I’d have answered something like:
“In animal welfare, I can sometimes know that I have a solid chance of reducing the suffering of some animals in the near-term, and that’s potentially worth doing as long as you’re, uh, careful about the risks of doing harm. And the bracketing people have said that this can be good, so there’s at least one element of justification.”
However, a month ago, I had to make some of the highest-stakes decisions of my life, with significant autonomy and no pre-existing framework to follow. This confronted me with the question of what procedure I should use.
I considered deference, some form of causal bracketing, metanormative bracketing, vibes, not supporting any intervention… And decided that I should at least write out how I would decide, if only to make it slightly consistent across the different options I had. I took the “timecapping” approach from the action-bias toolbox: setting myself a maximum amount of days of work to come up with a procedure that seemed “least bad” and which I’d have to stick to.[8]
Why not just use what exists?
All decision procedures seem flawed, in particular when applied by humans, and I guess this is why we’re hare, right? Moreover, if I copy others’ approaches to decision-making without understanding their flaws, there is a risk of deference cascades. This is particularly fragile for someone working to prioritize among interventions.
Why not maximize expected value?
There’s a little sequence on the topic that I heavily recommend reading.
What I came up with is probably worse than bracketing. However, my thoughts, aside from deference risks mentioned above, were:
It seems very sensitive to whether some edge case widens the possibilities we were considering. For example, say you want to work on a water quality intervention for shrimp, which seven papers of decent quality show positive evidence for. An eighth paper of similar quality shows that it may be somewhat harmful. Under bracketing, we are now more uncertain, and the intervention could be negative, so we rule it out. This seems to create a negative value-of-information problem, which proponents of bracketing have acknowledged.
Its main benefit of avoiding a certain kind of arbitrariness might still not be salient enough to me, especially insofar as its conclusions in animal welfare rarely seem very different from those of an EV-maxer that relies on evidence and acknowledges empirical uncertainty.[9] I think this means focusing on interventions with a track record and evidence of reducing suffering for the target populations, where we haven’t yet found evidence of robust negative 2nd-order effects that could strongly outweigh the beneficial effects. Maybe bracketing formally justifies deciding in that way, but I don’t think that it invented this approach to decision-making, and I mostly care about the practical part.
It changes very often, which makes it even more fragile if I can’t apply it myself.
Most ethical views are influenced by judgments on consequences. How I think about the effects of cage-free transitions on hens strongly influences whether I should support a cage-free campaign (and metanormative bracketing doesn’t tell me how to think about that).
It’s hard to know how to weigh up different virtues, though it does make some cases relatively obvious, like not hurting stray animals for fun. Like above, it feels insufficient to guide prioritization.
I could find myself convinced that there’s nothing preferable to this (especially metanormative bracketing focused on the recommendations of religious ethics), though, because my own decision procedure could lead to too many different conclusions.
Benefits of having my own decision procedure:
There are many downsides, such as falling into what DiGiovanni calls verdict-level curve-fitting, and of course, being more likely to make mistakes by thinking alone. But I’ll focus on the bright side.
It makes my decisions consistent
I can submit it to critique and improve it
I can more easily update over time as I get some information (one strong reason why I prioritize animal welfare over AI safety), compared to copying other procedures, which may involve more deference
I can more easily translate others’ methodology into my decision procedure,[11] and I also know where to cut corners under time constraints. Focusing on practical considerations like these in decision-making is controversial, but it can make a big difference in my context.
Keeps complex cluelessness alive: the sign of an intervention easily flips when we learn of a sufficiently well-evidenced 2nd order effect
Section 1: My decision procedure
Disclaimers
I’m optimizing for what I hope will help some animals, rather than for originality. This is heavily influenced by the writings of Michael St Jules, Anthony DiGiovanni, and Jesse Clifton. In particular, some of the main ideas here are covered in greater detail in What to do about near-term cluelessness in animal welfare. Some aspects are similar to forms of bracketing, and the more problematic ones might arise precisely where the procedure diverges from it.
Some of the language is seriously vague, as I may have an intention but haven’t always settled on the right way to formalize it. For consistency’s sake, I’ll make my commitments more precise over time, as long as this doesn’t paralyze me. Also, the points aren’t always clear; I only spent a few hours making them more legible.
This involves some bargaining with ethical views that are more common than mine, see below.
My decision procedure
I care about reducing aggregate suffering. I can’t claim that this very is well-specified (see #10).
I intend to roughly quantify the impact of my decisions, in particular for donations, prioritization research, and when considering a new role, considering some of the following:
Back of the envelope calculations (BOTECs), with 10th percentile, 90th percentile, and median estimates for each input concerning the effects I decide to evaluate (see #5). I see this as a way of representing my uncertainty, but this might diverge strongly from what people usually call “imprecise credences”. I’d particularly appreciate comments on the difference there.
My BOTECs’ inputs include qualitative judgments. For example, whether an organization pursuing an intervention seems to care about updating its intervention choice after M&E raises the 10th-percentile outcome for target individuals, which is usually negative. I also try to think of how my donation, beyond the marginal impact, may shape deserving institutions.[12] I agree that there may not be one uniquely justified way to quantify this, of course.
My BOTECs include “robust” 2nd-order effects (effects that seem to point in one specific direction—DiGiovanni defines the concept well in this post). I set the evidence bar fairly high for robust 2nd order effects, and this is the criteria on which I avoid getting systematically paralyzed by 2nd order effects.[13] However, if I am strict with this approach, it’s likely to face the same negative value of information problem as bracketing: the more I learn about 2nd order effects, the less likely I am to find interventions that look positive.[14] I’ll note that while this only evaluates a small subset of effects, that are unlikely to dominate the cosmos-wide impact of our actions, this is what many people in EA Animal Welfare would call “expected value”.
In part because I believe that AI could do the vast majority of the tasks that humans currently do within the next 5 years (controversial assumption), I apply an annual deweighting − 12% a year over the next 5 years, then slower for the years that follow, maybe 5% - to my confidence that the intervention will actually have the intended effects (this is after already accounting for the usual uncertainties in achieving outcomes—this is purely for “the world has changed too much for the reference class levers to achieve”).[15]
A superficial analysis under bracketing, though this currently amounts to deference to others who understand the procedure better than I do. This can be particularly helpful to map out the different effects of the intervention.
For deference on object-level interventions: I’d trust the literature over the animal welfare community in cases that aren’t movement-centric. Be explicit if my belief hinges on one specific deference point.[16] This is often the case in the animal movement—support for shrimp stunning seems to go back to SWP being confident that it was worthwhile, rather than people in the field. I’ll often have to go with very little evidence, which may just lead me to count the claim but give it a fairly wide range in a BOTEC, and not put the median too high above 0. In practice, when under time constraints (say, 3h), I often follow 3 steps.
Write what I think is true (it can be 2 lines, like “Salmon stunning is often botched but seems to be better evidenced than stunning for other fish species”)
Discuss with an LLM.
Ask a human expert whether what the LLM is saying seems accurate. This is insufficient, but a bit less of an echo chamber than the default.
I broadly apply a hits-based approach, with special caution[17] around certain types of interventions (ones that are particularly irreversible in the conventional sense, such as policy advocacy on highly contentious issues). Monte Carlo simulations enable me to see when intervention effectiveness is dominated by a few tail inputs that are likely to be overoptimistic (see also DiGiovanni on small probabilities here).
Some violations of norms can rule out supporting organizations or interventions altogether, in particular committing a large-scale harm that is not socially accepted (direct violence against humans, pyromania), or sharing deliberately untrue information as a core part of the intervention. However, I avoid being excessively strict in assessing whether an intervention follows norms, as this can rule out nearly everything. In contrast, the following harms are a reason to downweigh an intervention, but not to rule it out: animal experimentation in order to better help animals of the same species as the animals that are being experimented on; selective emphasis with no untrue information, where the main ethical motive is not disclosed.
Two deep unknowns influence my BOTECs’ results more than anything else: pain-intensity trade-offs and differences in the intensity of experience across animal groups. Currently, I “handle” this by running four separate Monte Carlos across where inputs on the above are fixed by the following quadrant:
Excruciating pain dominating vs 10:1 ratio between AIM’s disabling-equivalent days and excruciating pain[18]
“All animals being capable of significant suffering” vs “Only vertebrates being capable of significant suffering”. In both cases, the Monte Carlo runs across various potential moral weights scenarios.[19]
The actual decision: After the quadrant described in #11 is built, I take a portfolio approach if possible. This means supporting something based on whether it looks very good in at least one quadrant, as long as it’s not likely to be very harmful in another quadrant. Looking good is still not well-defined, and usually means looking positive across most of the distribution, wth a likely upside that we have reason to believe is larger than the likely downside. This is probably where my decision procedure strays the furthest from EV-maxing. Where I differ from some forms of bracketing is that I might support an intervention if the harmful effects in one quadrant seem likely to be much smaller than the positive effects of the entire portfolio in that quadrant.[20] Note that it’s less clear how to apply this to cases where you can only support one thing, In that case, I may be sympathetic to either taking the thing that have the highest expected value on the effects I’m evaluating, or taking what seems least uncertain (since this is not very clearly discriminating, this may end up being case-by-case, and reliant on qualitative judgment).
I am allowed to redefine my approach, even radically: this is an iterative process as I learn more about decision science and ethics. I also aim, over time, to keep the number of inputs in my procedure low-ish, as it’s commonly considered to be healthy in decision-making to use simpler models. I am even allowed to revise it simply if the verdicts are particularly bad by ordinary norms (overlaps with #10) or lead to no action ever being privileged (the thing this procedure is meant to avoid).
Brief tentative conclusion
In practice, the recommendations from my procedure are not very sensitive to whether I go with my higher or lower inputs on elements of the BOTEC such as the chance of success leading to outcomes, or the number of animals affected.
The big, undetermined, crucial considerations are: as I learn more, new, previously unconsidered effects are likely to outweigh the positive impacts I was evaluating; what trade-offs to make between pain intensity and duration; and how much to weigh different animals. The first consideration, I try to “cover” by updating over time if a new effect becomes sufficiently[21] evidenced. For the other two considerations, I try to cover them with a portfolio, but when that is not possible, I consider multiple options worthwhile under my procedure and tiebreak on other factors.[22]
Early updates from applying the procedure
“Alignment” influences my conclusions a lot. If I take a subordinate role in an org whose director is thinking about impact in roughly the same way as I am, they’re less likely to pursue things which I perceive as harmful, as long as they have access to the same evidence as I do.
It’s often about finding the thing that we have the least reason to think is harmful, and that makes me less keen on applying many hard constraints to what actions can be taken. Of course, one could reply that this should push me toward not taking any large-scale action, rather than being more permissive.
Section 2: My judgment on underdetermined matters
A lot of my judgment relies on matters where that cannot yet be empirically resolved (underdetermined questions). Since I have to make such assumptions in order to make decisions, I’ll make a list of the positions I’ve settled on here.[23]
Three core elements influence this list: arguments I’ve been exposed to in EA, the moral intuitions that feel unbreakable to me, and input from other ethical views, such as the main monotheistic religions. This mix of inputs is probably where most of the objections will lie. Sometimes, reconciling the three is impossible, and I add a “my own view” sub-bullet where the assumptions that satisfice the 3 conflict with my deepest-seated intuitions (in those cases, I tend to go with my own view but pressure myself to integrate constraints from the more acceptable view).
Eight underdetermined assumptions
It is important to help others.
I can focus on reducing suffering. I prioritize it, and for reasons that are less salient to me, I will try to consider other goods. Thus, if a being can suffer, that being matters to me.
My own view: only reducing suffering matters to me in high-stakes decisions, though I think following norms is extremely helpful, and I would not be able to take very repugnant actions in order to reduce more suffering.
Doing good is primarily defined by its impact on others, not by avoiding involvement in harm. However, gratuitous harm remains unacceptable, and norm-following can build character in ways that could help me reduce suffering.
My own view: only impact on others counts; I’m suspicious of caring too much about the act-omission distinction, especially as I fear that it would be paralyzing for high-stakes decisions under uncertainty (since you always have a possibility of causing harm with your decision).
The scientific method and quantification can help us in doing good. (Controversial) Expected value (in animal welfare, “suffering reduced”) is a heuristic I can use, though I should apply caution around “speculative” probabilities driving a result (DiGiovanni justifies it well there, and Monte Carlo simulations help me catch it).
Even without knowing whether it’s truth-tracking, I can try to quantify individual suffering in order to guide decisions.
The same logic leads me to use the philosophically controversial notion of counterfactual impact.
Constraints in my decision procedure above allow me to restrict that to interventions that are hopefully more evidence-based, and where our estimates of impacts on some categories of animals are better-justified.
This entire point 4 is a controversial leap. While these seem like the most “practical” approaches to assess interventions based on how they impact others, I’m very open to alternatives.
It’s contested whether we can compare suffering between animals of different species, as well as comparing different intensities of suffering. While many views agree that preventing the suffering of any sentient animal is good all else equal, and that preventing some suffering above a threshold is good all else equal, all else is generally not equal. Yet I have to make a judgment call.
This nudges me toward hedging by supporting different interventions that end up positive on at least one worldview, rather than maximizing across all of them.
However, since small animals are a smaller part of the total animal welfare portfolio, maybe preventing long-lasting pain in vertebrates shouldn’t be part of my portfolio, depending on how I define it.
Worse: while this works well for donations, this is less clear in cases where you have to work for several years on the same intervention (here, I’d most likely have to tiebreak with other factors).
Many animals, if not ~all[24], have a possibility of sentience that is not trivial enough to exclude them from my decision-making.
Wild animals undergo significant suffering, and since their lives are comparatively short, this often a non-negligible fraction of their existence.
My own view: I think that for virtually all farmed and wild animals, it is better not to exist (because Iof my suffering-focused views). I aim to bargain with more common-sense views on that question, since most humans see this as a serious moral mistake.
My own view: Artificial Intelligence has a 50%+ chance of doing the vast majority of tasks a human can do in the next 5 years, and this could influence the levers I use to do good.
Next steps
This is an accurate reflection of how I have made what I perceive to be extremely important ethical decisions. While I’m usually aware of how fragile my work is, having to write it out makes it as salient as it ought to be. The suffering of beings in the world horrifies me, and I want to do “the best I can” in that context. I expect that there are many holes in my decision procedure, and through reading other posts and hopefully receiving feedback on this one, I hope to patch them or entirely revise my approach.
Immensely grateful to Michael St Jules, Anthony DiGiovanni, and Jesse Clifton for their writings, and the influence they’ve had on my ethics and prioritization. No one gave feedback on this post, as I wrote it just a few hours before applications closed.
The decision to work in animal welfare was influenced by the fact that it seems particularly tricky to justify any course of action in AI safety. Though I can’t say I’ve ascertained that animal welfare is “safe”. I will not discussed why I focus on animal welfare in this post, as I didn’t have time for a section on this. If I realize that my decision procedure and assumptions can license AI safety (or other more “speculative” causes), I’ll consider it. I’m also open to realizing that my decision procedure, applied more broadly, may be more positive on even less speculative causes. Unfortunately, at the moment, I don’t know of many smaller-scale, “safer” ways to prevent suffering.
My experience is that people will often agree with the claim “We have no idea what the consequences on all sentient beings are”. This is lower among people who are longtermist-leaning, in my experience, but the sample size is extremely low.
This is a different criterion than causal bracketing, which rules out effects that make the intervention potentially positive or negative, though it often leads to considering similar effects.
I’ll add that this has a similar issues to various forms of bracketing, though: if we get passable evidence that a broad lever like economic growth has a very large 2nd-order effect on at least one group of beings which we care about, such as increasing the amount of pasture in the world and decreasing the number of r-selected animals, we may have to evaluate ~all interventions through their contribution to economic growth over anything else. I’m not sure this is a problematic feature: in practice, if I got such evidence, I would want to act on it. In practice, I think we have very little evidence on such broad levels.
Since that is hard to model under time pressure, I’ll probably just deweight the bulk of the impact by 12% for every year it takes between starting the intervention and achieving the expected impact if the intervention is successful.
They don’t make a massive difference, since the cruxy thing is generally that there is a threshold beyond which the best invertebrate interventions look significantly better than the best vertebrate interventions. Similarly for the pain tradeoffs ratio: we could have many inputs, but it boils down to “duration could be more important”, and
I admit this is morally risky, especially given our epistemic limitations. An example would be: doing an animal welfare reform that has a possibility of slightly increasing excruciating pain, like moving hens from furnished cages to cage-free housing, while also making a large grant to advance less painful chicken slaughter. This is harder to apply with your career choices or when your own organization is facing a potential.has strategy pivot.
Since the takeaway here is that there’s no absolutely obvious way to decide, I figure that I don’t have to formalize how I’d tiebreak. Also, it seems unlikely that I could strictly precommit to tiebreak methods for, say, career decisions.
I’ve removed assumptions that are already explicit in the decision procedure, along with what is generally always implicit in decision-making (the outside world exists, there’s some fact of the matter about better and worse), since they don’t seem to affect my decisions.
How One Clueless Layman Makes Decisions in Animal Welfare
This personal post is not part of the competition, as my intent is not to produce an original constructive response to cluelessness. However, Toby Tremlett encouraged me to share it insofar as it contributes to this week’s conversation.
TL; DR: I work in animal welfare and am clueless.[1] To avoid choosing actions at random, I decided to write a decision procedure for myself and to indicate the assumptions that accompany it. This does not constitute a new argument on the topic. Rather, this is a layman’s attempt to apply it. For those who want to skim, you can skip to the Decision procedure section, where I bolded the aspects of my decision procedure that seem most likely to elicit pushback.
Epistemic status and intent
Like many in animal welfare, I intuitively agree with cluelessness without understanding all of the philosophical steps behind it, and this will show in the post.[2] I focused on honestly representing my understanding of cluelessness, rather than high-quality philosophical reasoning.[3] I’m partly betting on Cunningham’s Law here.
This post does not aim to be a constructive general response like bracketing. It is also mostly derived from other writings on cluelessness. It assumes surface-level familiarity with bracketing.
However fragile and open to revision, this is a procedure I have already begun acting on. Since it has real-world consequences[4], it’s worth exposing it to criticism. And I’d love for the competition to give me ideas for a better procedure.
While the decision procedure and the list of assumptions already existed in an unpolished form, the post itself was time-capped at a day.
What Cluelessness Means To Me, a Layman
“Cluelessness”, to me, means that, when choosing how to act, my decision can’t actually track better expected effects on all of the beings which I care about.
I work in animal welfare, where cluelessness is both salient[5] and low-status. In EA animal welfare, action-bias rules[6]. The main ick around cluelessness in that field is not that it may justify repugnant conclusions[7], but that it licenses inaction.
But cluelessness doesn’t recommend inaction: it says that no action is impartially better-justified than another. This means that something else has to drive our decisions. But what is that? For the past year, I’d have answered something like:
“In animal welfare, I can sometimes know that I have a solid chance of reducing the suffering of some animals in the near-term, and that’s potentially worth doing as long as you’re, uh, careful about the risks of doing harm. And the bracketing people have said that this can be good, so there’s at least one element of justification.”
However, a month ago, I had to make some of the highest-stakes decisions of my life, with significant autonomy and no pre-existing framework to follow. This confronted me with the question of what procedure I should use.
I considered deference, some form of causal bracketing, metanormative bracketing, vibes, not supporting any intervention… And decided that I should at least write out how I would decide, if only to make it slightly consistent across the different options I had. I took the “timecapping” approach from the action-bias toolbox: setting myself a maximum amount of days of work to come up with a procedure that seemed “least bad” and which I’d have to stick to.[8]
Why not just use what exists?
All decision procedures seem flawed, in particular when applied by humans, and I guess this is why we’re hare, right? Moreover, if I copy others’ approaches to decision-making without understanding their flaws, there is a risk of deference cascades. This is particularly fragile for someone working to prioritize among interventions.
Why not maximize expected value?
There’s a little sequence on the topic that I heavily recommend reading.
Why not apply bracketing?
What I came up with is probably worse than bracketing. However, my thoughts, aside from deference risks mentioned above, were:
It seems very sensitive to whether some edge case widens the possibilities we were considering. For example, say you want to work on a water quality intervention for shrimp, which seven papers of decent quality show positive evidence for. An eighth paper of similar quality shows that it may be somewhat harmful. Under bracketing, we are now more uncertain, and the intervention could be negative, so we rule it out. This seems to create a negative value-of-information problem, which proponents of bracketing have acknowledged.
Its main benefit of avoiding a certain kind of arbitrariness might still not be salient enough to me, especially insofar as its conclusions in animal welfare rarely seem very different from those of an EV-maxer that relies on evidence and acknowledges empirical uncertainty.[9] I think this means focusing on interventions with a track record and evidence of reducing suffering for the target populations, where we haven’t yet found evidence of robust negative 2nd-order effects that could strongly outweigh the beneficial effects. Maybe bracketing formally justifies deciding in that way, but I don’t think that it invented this approach to decision-making, and I mostly care about the practical part.
It changes very often, which makes it even more fragile if I can’t apply it myself.
Why not apply metanormative bracketing?
Most ethical views are influenced by judgments on consequences. How I think about the effects of cage-free transitions on hens strongly influences whether I should support a cage-free campaign (and metanormative bracketing doesn’t tell me how to think about that).
It’s hard to know how to weigh up different virtues, though it does make some cases relatively obvious, like not hurting stray animals for fun. Like above, it feels insufficient to guide prioritization.
I could find myself convinced that there’s nothing preferable to this (especially metanormative bracketing focused on the recommendations of religious ethics), though, because my own decision procedure could lead to too many different conclusions.
Benefits of having my own decision procedure:
There are many downsides, such as falling into what DiGiovanni calls verdict-level curve-fitting, and of course, being more likely to make mistakes by thinking alone. But I’ll focus on the bright side.
It makes my decisions consistent
I can submit it to critique and improve it
I can more easily update over time as I get some information (one strong reason why I prioritize animal welfare over AI safety), compared to copying other procedures, which may involve more deference
It can easily integrate some of my intuitions[10]
I can more easily translate others’ methodology into my decision procedure,[11] and I also know where to cut corners under time constraints. Focusing on practical considerations like these in decision-making is controversial, but it can make a big difference in my context.
Keeps complex cluelessness alive: the sign of an intervention easily flips when we learn of a sufficiently well-evidenced 2nd order effect
Section 1: My decision procedure
Disclaimers
I’m optimizing for what I hope will help some animals, rather than for originality. This is heavily influenced by the writings of Michael St Jules, Anthony DiGiovanni, and Jesse Clifton. In particular, some of the main ideas here are covered in greater detail in What to do about near-term cluelessness in animal welfare. Some aspects are similar to forms of bracketing, and the more problematic ones might arise precisely where the procedure diverges from it.
Some of the language is seriously vague, as I may have an intention but haven’t always settled on the right way to formalize it. For consistency’s sake, I’ll make my commitments more precise over time, as long as this doesn’t paralyze me. Also, the points aren’t always clear; I only spent a few hours making them more legible.
This involves some bargaining with ethical views that are more common than mine, see below.
My decision procedure
I care about reducing aggregate suffering. I can’t claim that this very is well-specified (see #10).
I intend to roughly quantify the impact of my decisions, in particular for donations, prioritization research, and when considering a new role, considering some of the following:
Back of the envelope calculations (BOTECs), with 10th percentile, 90th percentile, and median estimates for each input concerning the effects I decide to evaluate (see #5). I see this as a way of representing my uncertainty, but this might diverge strongly from what people usually call “imprecise credences”. I’d particularly appreciate comments on the difference there.
My BOTECs’ inputs include qualitative judgments. For example, whether an organization pursuing an intervention seems to care about updating its intervention choice after M&E raises the 10th-percentile outcome for target individuals, which is usually negative. I also try to think of how my donation, beyond the marginal impact, may shape deserving institutions.[12] I agree that there may not be one uniquely justified way to quantify this, of course.
My BOTECs include “robust” 2nd-order effects (effects that seem to point in one specific direction—DiGiovanni defines the concept well in this post). I set the evidence bar fairly high for robust 2nd order effects, and this is the criteria on which I avoid getting systematically paralyzed by 2nd order effects.[13] However, if I am strict with this approach, it’s likely to face the same negative value of information problem as bracketing: the more I learn about 2nd order effects, the less likely I am to find interventions that look positive.[14]
I’ll note that while this only evaluates a small subset of effects, that are unlikely to dominate the cosmos-wide impact of our actions, this is what many people in EA Animal Welfare would call “expected value”.
In part because I believe that AI could do the vast majority of the tasks that humans currently do within the next 5 years (controversial assumption), I apply an annual deweighting − 12% a year over the next 5 years, then slower for the years that follow, maybe 5% - to my confidence that the intervention will actually have the intended effects (this is after already accounting for the usual uncertainties in achieving outcomes—this is purely for “the world has changed too much for the reference class levers to achieve”).[15]
A superficial analysis under bracketing, though this currently amounts to deference to others who understand the procedure better than I do. This can be particularly helpful to map out the different effects of the intervention.
For deference on object-level interventions: I’d trust the literature over the animal welfare community in cases that aren’t movement-centric. Be explicit if my belief hinges on one specific deference point.[16] This is often the case in the animal movement—support for shrimp stunning seems to go back to SWP being confident that it was worthwhile, rather than people in the field. I’ll often have to go with very little evidence, which may just lead me to count the claim but give it a fairly wide range in a BOTEC, and not put the median too high above 0. In practice, when under time constraints (say, 3h), I often follow 3 steps.
Write what I think is true (it can be 2 lines, like “Salmon stunning is often botched but seems to be better evidenced than stunning for other fish species”)
Discuss with an LLM.
Ask a human expert whether what the LLM is saying seems accurate. This is insufficient, but a bit less of an echo chamber than the default.
I broadly apply a hits-based approach, with special caution[17] around certain types of interventions (ones that are particularly irreversible in the conventional sense, such as policy advocacy on highly contentious issues). Monte Carlo simulations enable me to see when intervention effectiveness is dominated by a few tail inputs that are likely to be overoptimistic (see also DiGiovanni on small probabilities here).
Some violations of norms can rule out supporting organizations or interventions altogether, in particular committing a large-scale harm that is not socially accepted (direct violence against humans, pyromania), or sharing deliberately untrue information as a core part of the intervention. However, I avoid being excessively strict in assessing whether an intervention follows norms, as this can rule out nearly everything. In contrast, the following harms are a reason to downweigh an intervention, but not to rule it out: animal experimentation in order to better help animals of the same species as the animals that are being experimented on; selective emphasis with no untrue information, where the main ethical motive is not disclosed.
Two deep unknowns influence my BOTECs’ results more than anything else: pain-intensity trade-offs and differences in the intensity of experience across animal groups. Currently, I “handle” this by running four separate Monte Carlos across where inputs on the above are fixed by the following quadrant:
Excruciating pain dominating vs 10:1 ratio between AIM’s disabling-equivalent days and excruciating pain[18]
“All animals being capable of significant suffering” vs “Only vertebrates being capable of significant suffering”. In both cases, the Monte Carlo runs across various potential moral weights scenarios.[19]
The actual decision: After the quadrant described in #11 is built, I take a portfolio approach if possible. This means supporting something based on whether it looks very good in at least one quadrant, as long as it’s not likely to be very harmful in another quadrant. Looking good is still not well-defined, and usually means looking positive across most of the distribution, wth a likely upside that we have reason to believe is larger than the likely downside. This is probably where my decision procedure strays the furthest from EV-maxing.
Where I differ from some forms of bracketing is that I might support an intervention if the harmful effects in one quadrant seem likely to be much smaller than the positive effects of the entire portfolio in that quadrant.[20] Note that it’s less clear how to apply this to cases where you can only support one thing, In that case, I may be sympathetic to either taking the thing that have the highest expected value on the effects I’m evaluating, or taking what seems least uncertain (since this is not very clearly discriminating, this may end up being case-by-case, and reliant on qualitative judgment).
I am allowed to redefine my approach, even radically: this is an iterative process as I learn more about decision science and ethics. I also aim, over time, to keep the number of inputs in my procedure low-ish, as it’s commonly considered to be healthy in decision-making to use simpler models. I am even allowed to revise it simply if the verdicts are particularly bad by ordinary norms (overlaps with #10) or lead to no action ever being privileged (the thing this procedure is meant to avoid).
Brief tentative conclusion
In practice, the recommendations from my procedure are not very sensitive to whether I go with my higher or lower inputs on elements of the BOTEC such as the chance of success leading to outcomes, or the number of animals affected.
The big, undetermined, crucial considerations are: as I learn more, new, previously unconsidered effects are likely to outweigh the positive impacts I was evaluating; what trade-offs to make between pain intensity and duration; and how much to weigh different animals. The first consideration, I try to “cover” by updating over time if a new effect becomes sufficiently[21] evidenced. For the other two considerations, I try to cover them with a portfolio, but when that is not possible, I consider multiple options worthwhile under my procedure and tiebreak on other factors.[22]
Early updates from applying the procedure
“Alignment” influences my conclusions a lot. If I take a subordinate role in an org whose director is thinking about impact in roughly the same way as I am, they’re less likely to pursue things which I perceive as harmful, as long as they have access to the same evidence as I do.
It’s often about finding the thing that we have the least reason to think is harmful, and that makes me less keen on applying many hard constraints to what actions can be taken. Of course, one could reply that this should push me toward not taking any large-scale action, rather than being more permissive.
Section 2: My judgment on underdetermined matters
A lot of my judgment relies on matters where that cannot yet be empirically resolved (underdetermined questions). Since I have to make such assumptions in order to make decisions, I’ll make a list of the positions I’ve settled on here.[23]
Three core elements influence this list: arguments I’ve been exposed to in EA, the moral intuitions that feel unbreakable to me, and input from other ethical views, such as the main monotheistic religions. This mix of inputs is probably where most of the objections will lie. Sometimes, reconciling the three is impossible, and I add a “my own view” sub-bullet where the assumptions that satisfice the 3 conflict with my deepest-seated intuitions (in those cases, I tend to go with my own view but pressure myself to integrate constraints from the more acceptable view).
Eight underdetermined assumptions
It is important to help others.
I can focus on reducing suffering. I prioritize it, and for reasons that are less salient to me, I will try to consider other goods. Thus, if a being can suffer, that being matters to me.
My own view: only reducing suffering matters to me in high-stakes decisions, though I think following norms is extremely helpful, and I would not be able to take very repugnant actions in order to reduce more suffering.
Doing good is primarily defined by its impact on others, not by avoiding involvement in harm. However, gratuitous harm remains unacceptable, and norm-following can build character in ways that could help me reduce suffering.
My own view: only impact on others counts; I’m suspicious of caring too much about the act-omission distinction, especially as I fear that it would be paralyzing for high-stakes decisions under uncertainty (since you always have a possibility of causing harm with your decision).
The scientific method and quantification can help us in doing good. (Controversial) Expected value (in animal welfare, “suffering reduced”) is a heuristic I can use, though I should apply caution around “speculative” probabilities driving a result (DiGiovanni justifies it well there, and Monte Carlo simulations help me catch it).
Even without knowing whether it’s truth-tracking, I can try to quantify individual suffering in order to guide decisions.
The same logic leads me to use the philosophically controversial notion of counterfactual impact.
Constraints in my decision procedure above allow me to restrict that to interventions that are hopefully more evidence-based, and where our estimates of impacts on some categories of animals are better-justified.
This entire point 4 is a controversial leap. While these seem like the most “practical” approaches to assess interventions based on how they impact others, I’m very open to alternatives.
It’s contested whether we can compare suffering between animals of different species, as well as comparing different intensities of suffering. While many views agree that preventing the suffering of any sentient animal is good all else equal, and that preventing some suffering above a threshold is good all else equal, all else is generally not equal. Yet I have to make a judgment call.
This nudges me toward hedging by supporting different interventions that end up positive on at least one worldview, rather than maximizing across all of them.
However, since small animals are a smaller part of the total animal welfare portfolio, maybe preventing long-lasting pain in vertebrates shouldn’t be part of my portfolio, depending on how I define it.
Worse: while this works well for donations, this is less clear in cases where you have to work for several years on the same intervention (here, I’d most likely have to tiebreak with other factors).
Many animals, if not ~all[24], have a possibility of sentience that is not trivial enough to exclude them from my decision-making.
Wild animals undergo significant suffering, and since their lives are comparatively short, this often a non-negligible fraction of their existence.
My own view: I think that for virtually all farmed and wild animals, it is better not to exist (because Iof my suffering-focused views). I aim to bargain with more common-sense views on that question, since most humans see this as a serious moral mistake.
My own view: Artificial Intelligence has a 50%+ chance of doing the vast majority of tasks a human can do in the next 5 years, and this could influence the levers I use to do good.
Next steps
This is an accurate reflection of how I have made what I perceive to be extremely important ethical decisions. While I’m usually aware of how fragile my work is, having to write it out makes it as salient as it ought to be. The suffering of beings in the world horrifies me, and I want to do “the best I can” in that context. I expect that there are many holes in my decision procedure, and through reading other posts and hopefully receiving feedback on this one, I hope to patch them or entirely revise my approach.
Immensely grateful to Michael St Jules, Anthony DiGiovanni, and Jesse Clifton for their writings, and the influence they’ve had on my ethics and prioritization. No one gave feedback on this post, as I wrote it just a few hours before applications closed.
The decision to work in animal welfare was influenced by the fact that it seems particularly tricky to justify any course of action in AI safety. Though I can’t say I’ve ascertained that animal welfare is “safe”. I will not discussed why I focus on animal welfare in this post, as I didn’t have time for a section on this. If I realize that my decision procedure and assumptions can license AI safety (or other more “speculative” causes), I’ll consider it. I’m also open to realizing that my decision procedure, applied more broadly, may be more positive on even less speculative causes. Unfortunately, at the moment, I don’t know of many smaller-scale, “safer” ways to prevent suffering.
You just have to say “insects” and everyone understands why cluelessness is intuitive.
My mistakes may reveal that some points are harder for non-philosophers to grasp, which could help future writings on cluelessness.
About which, on an impartial altruistic level, I am still clueless…
My experience is that people will often agree with the claim “We have no idea what the consequences on all sentient beings are”. This is lower among people who are longtermist-leaning, in my experience, but the sample size is extremely low.
Caveat: I’ve heard action-biased claim the exact opposite, so don’t take me seriously
People are very much used to that.
I’d say this ended up being about 30 hours, though many hours looking into object-level interventions also influenced it.
This is a particularly crucial and uncertain claim, and I’d love to hear pushback.
This might actually constitute a downside.
I’m pointing at something like “Sorry, I can’t read your BOTEC, I use metanormative bracketing and this does not bear on my decision”.
Though I’ve put very little thought into it.
This is a different criterion than causal bracketing, which rules out effects that make the intervention potentially positive or negative, though it often leads to considering similar effects.
I’ll add that this has a similar issues to various forms of bracketing, though: if we get passable evidence that a broad lever like economic growth has a very large 2nd-order effect on at least one group of beings which we care about, such as increasing the amount of pasture in the world and decreasing the number of r-selected animals, we may have to evaluate ~all interventions through their contribution to economic growth over anything else. I’m not sure this is a problematic feature: in practice, if I got such evidence, I would want to act on it. In practice, I think we have very little evidence on such broad levels.
Since that is hard to model under time pressure, I’ll probably just deweight the bulk of the impact by 12% for every year it takes between starting the intervention and achieving the expected impact if the intervention is successful.
Though I don’t weigh this quantitatively. It could be used as a tiebreaker though, but I haven’t put thought into it.
Not yet sure whether that’s an aspiration to discount irreversible interventions, or to fully rule them out. I feel unlikely to do the latter.
This basically always leads to prioritizing longer-lasting pain
They don’t make a massive difference, since the cruxy thing is generally that there is a threshold beyond which the best invertebrate interventions look significantly better than the best vertebrate interventions. Similarly for the pain tradeoffs ratio: we could have many inputs, but it boils down to “duration could be more important”, and
I admit this is morally risky, especially given our epistemic limitations. An example would be: doing an animal welfare reform that has a possibility of slightly increasing excruciating pain, like moving hens from furnished cages to cage-free housing, while also making a large grant to advance less painful chicken slaughter. This is harder to apply with your career choices or when your own organization is facing a potential.has strategy pivot.
No clear bar, but I aim for some sense of consistency with effects I’ve previously considered
Since the takeaway here is that there’s no absolutely obvious way to decide, I figure that I don’t have to formalize how I’d tiebreak. Also, it seems unlikely that I could strictly precommit to tiebreak methods for, say, career decisions.
I’ve removed assumptions that are already explicit in the decision procedure, along with what is generally always implicit in decision-making (the outside world exists, there’s some fact of the matter about better and worse), since they don’t seem to affect my decisions.
Save for sponges and the like?