Hi Bob & team,

Really great work. Regardless of my specific disagreements, I do think calculating moral weights for animals is literally some of the highest value work the EA community can do, because without such weights we cant compare animal welfare causes to human-related global health/longtermism causes - and hence cannot identify and direct resources towards the most important problems. And I say this as someone who has always donated to human causes over animal ones, and who is not, in fact, vegan.

With respect to the post and the related discussion:

(1) Fundamentally, the quantitative proxy model seems conceptually sound to me.

(2) I do disagree with the idea that your results are robust to different theories of welfare. For example, I myself reject hedonism and accept a broader view of welfare (given that we care about a broad range of things beyond happiness,  e.g. life/freedom/achievement/love/whatever). If (a) such broad welfarist views are correct, (b) you place a sufficiently high weight on the other elements of welfare (e.g. life per se, even if neutral valenced), and (c) you don't believe animals can enjoy said elements of welfare (e.g. if most animals aren't cognitively sophisticated enough to have preferences over continued existence), then  an additional healthy year of human life would plausibly be worth a lot more than an equivalent animal year even after accounting for similar degrees of suffering and the relevant moral weights as calculated.

(3) I would like to say, for the record, that a lot of the criticism you're getting (and I don't exempt myself here) is probably subject to a lot of motivated reasoning. I am personally uncertain as to the degree to which I should discount my own conclusions over this reason.

(4) My main concern, as someone who does human-related cause prioritization research, is the meat eater argument and whether helping to save human lives is net negative from overall POV, given the adverse consequences for animal suffering. I am moderately optimistic that this is not so, and that saving human lives is net positive (as we want/need it to be) . Having very roughly run the numbers myself using RP's unadjusted moral weights (i.e. not taking into account point 2 above) and inputting other relevant data (e.g. on per capita consumption rate of meat), my approximate sense is that in saving lives we're basically buying 1 full week of healthy human life for around 6 days of chicken suffering or above 2 days of equivalent human suffering - which is worth it.

At risk of jeopardizing EA's hard-won reputation of relentless internal criticism:

Even setting aside its object-level impact-relevant criteria (truth, importance, etc), this is just enormously impressive both in terms of magnitude and quality. The post itself gives us readers an anchor on which to latch critiques, questions, and comments, so it's easy to forget that each step or decision in the whole methodology had to be chosen from an enormous space of possibilities. And this looks— at least on a first red—like very many consecutive well-made steps and decisions

I'm curating the post. I should note that I think I agree with a big chunk of Joel's comment.

I notice I'm quite confused about the symmetry assumption. For example: suppose we have two animals — M and N — and they're both at the worst end of their welfare ranges (~0th percentile) and have equal lifespans (and there are no indirect effects). M has double the welfare range of N. If we assume that their welfare ranges are symmetric around the neutral point, then replacing one M with one N is similar to moving M from the 0th percentile of its welfare range to the 25th. If, however, their welfare ranges aren't symmetric — say M's is skewed very positive and N's is skewed very negative — then we could actually be making the situation worse. In the BOTEC spreadsheet you linked, you seem to resolve this by requiring people to state the specific endpoints of the welfare ranges relative to the neutral point. If that's the main solution, it seems very important to be clear about where the neutral point is for different animals, and that seems really hard — I'm curious if you have thoughts on how to approach that. (Maybe you assume that welfare ranges are generally close to symmetric, or asymmetric in similar ways? If so, I would like to understand why you think that.) It's also very possible that I misunderstood something; I was reading things fast and haven't read all the linked posts and documents.

To make sure that I understand (the broad strokes of the rest of the framework) correctly; suppose I want to use this framework and these welfare range estimates to help me decide between two (completely hypothetical, unrealistic) options — assuming that every animal's welfare range is symmetric around the neutral point: (A) getting someone to buy the equivalent of a cage-free chicken instead of a caged chicken vs (B) getting someone to buy a farmed salmon instead of a farmed carp. Is it right that I'd now need to incorporate (estimates for) the following additional information? 

1. To understand the welfare impact on the animals in question
   1. Lifespans of the animals^[1]
   2. Where exactly on their respective welfare ranges they are, on average (in the situations I'm considering)^[2]
2. The other stuff
   1. Indirect effects
      1. E.g. how (many) other animals are affected by the farming processes — feed (insects/fish), how many die in the farming process, etc.
   2. Costs of the interventions

(In particular, I worry a bit that people might not be tracking 1a and 1b — you seem to worry about this, too, given the sections on things like "so you're saying that one person =~ three chickens?" — and I'd like to make sure that I actually understand correctly (and that others do, too).)

1. ^{^}

   Broiler chickens live for 5-7 weeks, apparently. Farmed carp apparently live for around a year, and farmed salmon live for around 1-3 years. (These numbers are from quick Google searches  —definitely don't trust them.) 

2. ^{^}

   A highly technical diagram is below. Note that the diagram represents the ranges as if they're all symmetric — as if each animal can experience as much bad as good —  whereas that isn't necessarily true. The welfare impact of choice (A) and (B) is the highlighted interval (assuming completely made-up numbers), multiplying by the lifespans of the animals, and adjusting for indirect effects.

   

   Given the lifespans of the animals in question, switching to salmon seems harmful ((even) without accounting for indirect effects or costs). 

“I can’t believe that bees beat salmon!”

We also find it implausible that bees have larger welfare ranges than salmon. But (a) we’re also worried about pro-vertebrate bias; (b) bees are really impressive; (c) there's a great deal of overlap in the plausible welfare ranges for these two types of animals, so we aren't claiming that their welfare ranges are significantly different; and (d) we don’t know how to adjust the scores in a non-arbitrary way. So, we’ve let the result stand. (We’d make similar points in response to: “I can’t believe that octopuses beat carp!”)

(I could believe octopuses beat carps, because octopuses seem unusually cognitively sophisticated among animals.)

I'd guess the main explanation for this (at least sentience-adjusted, if that's what's meant here), which may have biased your results against salmons and carps, is that you used the prior probability for crab sentience (43% mean, 31% median from table 3 in the doc) as the prior probability for salmon and carp sentience, and your posterior probabilities of sentience are generally very similar to the priors (compare tables 3 and 4 in the doc). Honeybees, fruit flies, crabs, crayfish, salmons and carps all ended up with similar sentience probabilities, but I'd assign higher probabilities to salmons and carps than to the others. You estimated octopuses to be about 2x as likely to be sentient as salmons and carps, according to both your priors and posteriors, with means and medians roughly between 73% and 78% for octopuses. On the other hand, your sentience-conditioned welfare ranges didn't differ too much between the fish, octopuses and bees. It's worth pointing out that Luke Muehlhauser had signficantly higher probabilities for rainbow trouts (70%, in the salmonid family like salmons) than Gazami crabs (20%) and fruit flies (25%), and you could use his prior for rainbow trouts for salmons and carps instead (or something in between). That being said, his probabilities were generated in different ways from yours, so that might introduce other biases. You could instead use your prior for octopuses (or something in between). Or, most consistent with your methodology, would be to have the authors of the original estimates for RP just estimate these probabilities directly, with or without the data you gathered for salmons and carps. Any of these would be relatively small fixes.

As an aside, should we interpret this sentience probability work as not primarily refining your old estimates (since the posteriors and priors are very similar), but as adding other species and further modelling your uncertainty?

There may be some other smaller potential sources of bias that contributed here, but I don't expect them to have been that important:

1. I'm guessing salmon and carp (and apparently zebrafish, which seem to often have been used when direct evidence wasn't available, maybe more for carp) are less well-studied than bees, so your conservative assumptions of assigning 0 to "unknown" for both probabilities of sentience and welfare ranges conditional on sentience may count more against them. For example, there were some studies found for "cognitive sophistication" for honeybees but not for salmons or carps, and more found for "mood state behaviors" for honeybees than salmons and carps in your new Sentience Table. For your Welfare Range Table, bees had fewer "unknowns" for cognitive proxies than salmons and carps, but more for hedonic proxies and representational proxies.
   1. One possible quick-ish fix would be to use a prior for the presence/absence of proxies across animals based on the ones for which there are studies (possibly just those you collected evidence for), although this may worsen other biases, like publication bias, and it requires you to decide how to weigh different animal species (but uniformly across those you collected evidence for is one way, although somewhat arbitrary).
   2. Another quick-ish fix could be to make more assumptions between species you gathered evidence for, e.g. if a fruit fly has some capacity, I'd expect fish to, as well, and if some mammal is missing some capacity, I'd expect salmon and carp to not have it either. This may be too strong, but you did use the crab sentience prior for the fish.
   3. Longer fixes could use more sophisticated missing data methods.
2. You may have underestimated salmon and carp neuron counts around 100x.

Hi Bob and RP team,

I've been working on a comparative analysis of the knock-on effects of bivalve aquaculture versus crop cultivation, to try to provide a more definitive answer to how eating oysters/mussels compares morally to eating plants. I was hoping I could describe how I'd currently apply the RP team's welfare range estimates, and would welcome your feedback and/or suggestions. Our dialogue could prove useful for others seeking to incorporate these estimates into their own projects.

For bivalve aquaculture, the knock-on moral patients include (but are not limited to) zooplankton, crustaceans, and fish. Crop cultivation affects some small mammals, birds, and amphibians, though its effect on insect suffering is likely to dominate.

RP's invertebrate sentience estimates give a <1% probability of zooplankton or plant sentience, so we can ignore them for simplicity (with apologies to Brian Tomasik). The sea hare is the organism most similar to the bivalve for which  sentience estimates are given, and it is estimated that a sea hare is less likely to be sentient than an individual insect. Although the sign of crop cultivation's impact on insect suffering is unclear, the magnitude  seems likely to dominate the effect of bivalve aquaculture on the bivalves themselves, so we can ignore them too for simplicity.

The next steps might be:

1. Calculate welfare ranges:
   1. For bivalve aquaculture, use carp, salmon, crayfish, shrimp, and crabs to calculate a welfare range for the effect of bivalve aquaculture on marine populations.
   2. Use chickens as a model species to calculate a welfare range for the effect of crop cultivation on vertebrate populations.
   3. For the effect of crop cultivation on insect suffering, I might just toss this problem on to future researchers. I'm only doing this as a side project, and given the sheer complexity of the considerations at play, I'm worried I might publish something which inadvertently increases insect suffering instead of decreasing it.
2. For several moral views (negative utilitarianism, symmetric utilitarianism) and several perspectives of the value of a typical wild animal's life (net negative, net neutral, net positive), extract relevant conclusions. (e.g. if bivalve aquaculture is robustly shown to increase marine populations, given Brian's arguments that crop cultivation likely reduces vertebrate populations, a negative utilitarian who views wild animal lives as net negative may want to oppose bivalve consumption.)

(Of course, I'd have to mention longtermist considerations. The effect of  norms surrounding animal consumption on moral circle expansion could be crucial. So could the effect of these consumption practices on climate change or on food security.)

Question about uncertainty modeling (tagging @Laura Duffy here since she might be the best person to answer it): 

How do you think about the different models of welfare capacity that were averaged together to make the mixture model? Is your assumption that one of these models is really the true correct model in all species (and you don't yet know which one it is), or that the different constituent models might each be more or less true for describing the welfare capacity for each individual species? 

My context for asking this is in thinking about quantifying the uncertainty for a function that depends on the welfare ranges of two different species (e.g. y = f(welfare range of shrimp, welfare range of pigs)). It's tempting to just treat the welfare ranges of shrimp and pigs as independent variables and to then sample each of them from their respective mixture model distribution. But if we think there's one true model and the mixture model is just reflecting uncertainty as to what that is, the welfare ranges of shrimp and pigs should be treated as correlated variables. One might then obtain an estimate of the uncertainty in y by generating samples as follows:

1. Randomly pick one of the 9 models in the mixture model as the true model 
2. Sample the welfare range of both shrimp and pigs from their distributions for the selected constituent model
3. Compute y = f(welfare range of shrimp, welfare range of pigs)
4. Repeat steps 1-3 until the desired # of samples is obtained 

I could also imagine computing the covariance of the different species' welfare ranges and directly generating samples as correlated random variables. 

Bob Fischer @ 2023-09-27T21:00 (+6)

Thanks a bunch for your question, Matt. I can speak to the philosophical side of this; Laura has some practical comments below. I do think you're right that---and in fact our team discussed the possibility that---we ought to be treating the welfare range estimates as correlated variables. However, we weren't totally sure that that's the best way forward, as it may treat the models with more deference than makes sense.
Here's the rough thought. We need to distinguish between (a) philosophical theories about the relationship between the proxies and welfare ranges and (b) models that attempt to express the relationship between proxies and welfare range estimates. We assume that there's some correct theory about the relationship between the proxies and welfare ranges, but while there might be a best model for expressing the relationship between proxies and welfare range estimates, we definitely don't assume that we've found it. In part, this is because of ordinary points about uncertainty. Additionally, it's because the philosophical theories underdetermine the models: lots of models are compatible with any given philosophical theory; so, we just had to choose representative possibilities. (The 1-point-per-proxy and aggregation-by-addition approaches, for instance, are basically justified by appeal to simplicity and ignorance. But, of course, the philosophical theory behind them is compatible with many other scoring and aggregation methods.) So, there's a worry that if we set things up the way you're describing, we're treating the models as though they were the philosophical theories, whereas it might make more sense not to do that and then make other adjustments for practical purposes in specific decision contexts if we're worried about this.

Laura's practical notes on this:

1. A change like the one you're suggesting would likely decrease the variance in the estimates of f(), since if you assume the welfare ranges are independent variables, you'd get samples where the undiluted experiences model is dominating the welfare range for, say, shrimp, and the neuron count model is dominating the welfare range for pigs. I suggest a quick practical way of dealing with this would be to cut off values of f() below the 2.5th percentile and 97.5th percentile.
2. Or, even better, I suggest sorting the welfare ranges from least to greatest, then using pairs of the ith-indexed welfare ranges for the ith estimate of f(). Since each welfare model is given the same weight, I predict this'll most accurately match up welfare range values from the same welfare model. (e.g. the first 11% will be neuron count welfare ranges, etc.)
3. Ultimately, however, given all the uncertainty in whether our models are accurately tracking reality, it might not be advisable to reduce the variance as such.

MHR @ 2023-10-04T00:05 (+16)

Thanks, this is great information! The concern you raised regarding distinguishing between philosophical theories and models makes a lot of sense. With that said, I don't currently feel super satisfied with the practical steps you suggested. 

On the first note, the impact of the correlation depends on the structure of $f$. Suppose I'm trying to estimate the total harms of eating chicken/pork, so we have something like $y=c_1\ast welfarerangeofpigs+c_2\ast welfarerangeofchickens$. In this case, treating the welfare ranges of chickens and pigs as correlated will increase the variance of $y$. On the flip side, if we're trying to estimate the welfare impact of switching from eating chicken to eating pork, we have something like $y=c_3\ast welfarerangeofchickens-c_4\ast welfarerangeofpigs$. In that case,  treating the welfare ranges of pigs and chickens as correlated will decrease the variance of $y$. Trying to address this in an ad-hoc manner seems like it's pretty challenging. 

On the second note, I think that's basically treating the welfare capacities of e.g. pigs and chickens as perfectly correlated with one another. That seems extreme to me, since I think a substantial portion of the uncertainty in the welfare rages is coming from uncertainty as to which traits each species has, not which philosophical theory of welfare is correct. 

I come away still thinking that the procedure I suggested seems like the most workable of the approaches mentioned so far. To put a little more rigor to things, here are some examples of plotting the welfare range estimates of chickens and pigs against one another with the different methods (uncorrelated sampling from the respective mixture distributions, sampling from the ordered distributions, and pair-wise sampling from the constituent models). In addition, there are some plots showing the impact of the different sampling methods on some toy analyses of the welfare impact of eating chicken/pork and the impact of switching from eating chicken to eating pork (note that the actual numbers are not intended to be very representative). You can see that the trimming approach only make sense in the second case, and that the paired sampling from constituent models approach produces distributions in between those for the uncorrelated case and those for the ordered case. 

 

Note that when using the pair-wise sampling from constituent models approach, pigs and chickens are more strongly correlated with one another than many other pairs of species are. Here is what the correlation between chickens and shrimp looks like, for example: 

Laura Duffy @ 2023-10-04T00:30 (+7)

Hey, thanks for this detailed reply! 
When I said "practical", I more meant "simple things that people can do without needing to download and work directly with the code for the welfare ranges." In this sense, I don't entirely agree that your solution is the most workable of them (assuming independence probably would be). But I agree--pairwise sampling is the best method if you have the access and ability to manipulate the code! (I also think that the perfect correlation you graphed makes the second suggestion probably worse than just assuming perfect independence, so thanks!)

MHR @ 2023-10-04T00:47 (+4)

Yeah that makes complete sense, it was a pain to get the pairwise sampling working.

Love this type of research, thank you very much for doing it!

I'm confused about the following statement:

While carp and salmon have lower scores than pigs and chickens, we suspect that’s largely due to a lack of research.

Is this a species-specific suspicion? Or does a lower amount of (high-quality) research on a species generally reduce your welfare range estimate? 
On average I'd have expected the welfare range estimate to stay the same with increasing evidence, but the level of certainty about the estimate to increase. 

If you have reason to believe that the existing research is systematically biased in a way that would lead to higher welfare range estimates with more research,  do you account for this bias in your estimates?

This is really valuable work, and I look forward to seeing the discussion that it generates and to digging into it more closely myself. I did have one immediate question about the neuron count model specifically, though I recognize that it's a  a small contributor to the overall weights.  I'd be curious to understand how you arrived at 13 million neurons as your estimate for salmon. The reference in the spreadsheet is: 

The teleost brain is capable of adult neurogenesis, with neural proliferation zones in dozens of locations within the brain (e.g. Zupanc et al. 2005, Zupanc 2009). This makes a definitive count of total neurons within the brain difficult, since the number of neurons may be continuously in flux. For example, Zupanc (2009) summarizes: “the continuous production of new cells, together with the longterm persistence of a large portion of them, leads to a permanent growth of the brain and its individual structures... This growth by a net increase in the total number of brain cells is characteristic of at least some, but likely most, of the estimated 30,000 species of teleost fish.” Therefore, reports of total neuron counts for salmon and carp are rare, but Hinsch & Zupanc (2007) report that “By labeling S-phase cells with the thymidine analog 5-bromo-2-deoxyuridine (BrdU), quantitative analysis demonstrated that, on average, 6000 new cells were generated in the entire adult brain within any 30 min period. This corresponds to roughly 0.06% of the total number of brain cells” in an adult zebrafish (Danio rerio, a model cyprinid) brain. As part of their study, Hinsch & Zupanc (2007) report that, for adult zebrafish, the total number of brain cells varied between 0.8 x 107 and 1.3 x 107 (mean: 1.0 x 107 ± S.E.M. 8 x 105). They also report that “approximately 46% of the cells present at 10 days persisted in the adult zebrafish brain” meaning that “​​at least half of the cells generated in the adult zebrafish brain develop into neurons and are likely to persist for the rest of the fish’s life.” This pattern is reflected in other species of teleosts, for example in adult gymnotiform fish (Apteronotus leptorhynchus) who generate 100 000 new brain cells (corresponding to approximately 0.2% of the total population of cells in the brain) within a period of 2 hours (Zupanc & Horschke 1995). Thus the teleost brain is constantly growing and likely increasing in terms of total number of neurons, and counts are only representative of snapshots through time.

I don't easily see how that translates to 13 million neurons. When I previously looked at this issue myself, I came away thinking it was possible that salmon had substantially more neurons than you're estimating. 

Hello to all,

Have you contacted the Integrated Information Theory group about this project? In my (dualistic naturalist) viewpoint their work is the most advanced in the area of consciece detection. 

https://www.amazon.com/Sizing-Up-Consciousness-Objective-Experience/dp/0198728441

Of course, conscience is absolutely noumenal and the best part of their work is focused in the case where self reported conscience experience is possible [humans], but they tried to extrapolate into mathematical models of application to any material system.

Hey, I thought I'd make a Bayesian adjustment to the results of this post. To do this, I am basically ignoring all nuance. But I thought that it might still be interesting. You can see it here: https://nunosempere.com/blog/2023/02/19/bayesian-adjustment-to-rethink-priorities-welfare-range-estimates/

I skimmed the piece on axiological asymmetries that you linked and am quite puzzled that you seem to start with the assumption of symmetry and look for evidence against it. I would expect asymmetry to be the more intuitive, therefore default, position. As the piece says

At just the first-order level, people tend to assume that (the worst) pain is worse than (the best) pleasure is pleasurable. The agonizing ends for non-human animals in factory farms and in the wild seem far worse than the best sort of life they could realize would be good. [...]  it’s hard to find any organisms that risk the worst pains for the greatest pleasures and vice versa.

I would expect that a difference in magnitude between the best pleasure and worst possible is the most obvious explanation, but the piece concludes that these judgments are "far more plausibly explained by various cognitive biases".

As far as I can tell this would suggest that either:

• Someone who has recently experienced or is currently experiencing intense suffering (and therefore has a better understanding of the stakes) would be more willing to take the kind of roulette gamble described in the piece. This seems unlikely.
• People's assessments of hedonic states are deeply unreliable even if they have recent experience of the states in question. I don't like this much because it means we have to fall back on physiological evidence for human pleasure/suffering, which, as shown by the mayonnaise example, can't give us the full picture.

On a slightly separate note, I played around with the BOTEC to check the claim that assuming symmetry doesn't change the numbers much and I was convinced. The extreme suffering-focused assumption (where perfect health is merely neutral) resulted in double the welfare gain of the symmetric assumption (when the increase in welfare as a percentage of the animals' negative welfare range is held constant). 

My main question on this last point is: why use "percentage of the animals' negative welfare range" when "percentage of the animals' total welfare range"  seems more relevant and would not vary at all across different (a)symmetry assumptions?

If these estimates will be used as multipliers for a hedonistic/suffering scale based on WFP's pain intensity levels (as was done here recently), then the undiluted experience model might contradict the definition of disabling pain, and probably contradicts the definition of excruciating pain, because these can't be ignored and they take up most or ~all of an animal's attention, by definition. Furthermore, I think what you'd want to do instead anyway, if using WFP's pain scale, is just use an equality model and assess more carefully where an animal is on WFP's pain scale, taking into account potential distractions. Dilution wouldn't change the badness of a given level of suffering (affective component of physical and psychological pain, which is what I think WFP's scale is supposed to capture); it would reduce the level of suffering, and so move the experience towards the milder end of WFP's pain scale. I'm confident that excruciating pain in humans is never or rarely significantly diluted (just through distraction by things other than similarly intense pain), and I doubt that disabling pain is significantly diluted, too.

WFP also has a post on the role of attention here, and, related to this, they wrote (bold mine):

Additionally, the potential for positive welfare may be also overestimated if factors other than attention are not considered. For example, pain caused by traumatic injury or pathological processes may lead to immobility, restricted movement or impaired behavioral responsiveness to potentially pleasurable opportunities [30]. Similarly, sickness, weakness, nausea, dizziness and other debilitating affects may demotivate animals from engaging in physically active, gregarious and positive behaviors [30].

Finally, positive and negative affective states may interact in complex ways other than those considered. For instance, evidence indicates that in environments where animals can engage in motivated behaviors the perceived intensity of pain is reduced. In chickens, experiments conducted by Mike Gentle two decades ago [21,31] have shown that the higher the motivation to engage in a behavior (hence attention diverted to it), the higher the degree of endogenous analgesia mediated by opioids. The possibility to express positive behaviors may therefore inhibit pain that would otherwise be felt as Hurtful or Annoying (pain of higher intensity cannot, by definition, be eliminated with distraction).

I also worry about most of the qualitative/non-quantitative models basically double-counting animals' responses, if used as multipliers for WFP's pain scale. Some animals may just not be capable of experiencing excruciating pain at all, but that should just be captured in the probability that they are in fact experiencing excruciating (or disabling) pain under given conditions, not as a multiplier for the badness of excruciating pain, except possibly for reasons that really do stack on top of excruciating pain. Maybe the number of JNDs or conscious subsystems stack on top, which are reflected in the quantitative models, but few if any of the qualitative indicators seem like they should stack on top.

I would personally shift the probabilities assigned to the qualitative models to the equality model, when you want to use the welfare ranges estimates as multipliers for a WFP pain intensity scale.^[1]

1. ^{^}

   But then, this also makes a uniform prior across the original subset of models look weird/suspicious.

   Each of the eight models was assigned an equal probability of being correct.

   Should we instead use a uniform prior over the new subset of models for multiplying WFP scales? Your credences in the models shouldn't be sensitive to something like this.

Do the estimates for black soldier flies primarily reflect adults? If we wanted to use an estimate for BSF larvae or mealworms, should we use the BSF estimates, the silkworm estimates (which presumably reflect the larvae, or else you'd call them silkmoths?), something in-between (an average?) or something else?

So, given our methodological commitment to letting the empirical evidence drive the results, we decided not to include this hypothesis in our calculations

I'm not sure I understand this reasoning. If our interpretation of the empirical evidence depends on whether we accept different philosophical hypotheses, it seems like the results should reflect our uncertainty over those hypotheses. What would it mean for claims about weights on potential conscious experiences to be driven purely by empirical evidence, if questions about consciousness are inherently philosophical?

@Laura Duffy @Bob Fischer 
A question about your methodology : If I understand correctly, your placeholders are probability-of-sentience-adjusted, but your key takeaways are not (since they are "conditional on sentience").
Why having adjusted for sentience in your placeholders but not in your key takeaways ?

Hi Bob,

Could you clarify how you aggregated the welfare range distributions from the 8 models you considered? I understand you gave the same weight to all of these 8 models, but I did not find the aggregation method here.

I would obtain the final cumulative distribution function (CDF) of the welfare range aggregating the CDFs of the 8 models with the geometric mean of odds, as Epoch did to aggregate judgement-based AI timelines. I think Jaime Sevilla would suggest using the mean in this case:

If you are not aggregating all-considered views of experts, but rather aggregating models with mutually exclusive assumptions, use the mean of probabilities.

However, I would say the 8 welfare range models are closer to the "all-considered views of experts" than to "models with mutually exclusive assumptions". In addition:

• The mean ignores information from extremely low predictions, and overweights outliers.
• The weighted/unweighted geometric mean of odds (and also the geometric mean) performed better than the weighted/unweighted mean on Metaculus' questions.
• Samotsvety aggregated predictions differing a lot between them from 7 forecasters^[1] using the geometric mean after removing the lowest and highest values (and the geometric mean is more similar to the geometric mean of odds than to the mean).

1. ^{^}

   For the question "What is the unconditional probability of London being hit with a nuclear weapon in October?", the 7 forecasts were 0.01, 0.00056, 0.001251, 10^-8, 0.000144, 0.0012, and 0.001. The largest of these is 1 M (= 0.01/10^-8) times the smallest.

Hi Bob,

Do you have any thoughts on the feasibility of extending your framework to estimate the welfare range of non-biological systems, namely advanced AI models like GPT-4? It naively looks like some of the models you considered to estimate the welfare ranges could apply to AI systems. I wish discussions about artificial sentience moved from "is this AI system sentient" to "what is the expected welfare range of this AI system"...

Thanks and congratulations to the RP team for your work on this. This is incredibly thorough and useful!

Having looked at the whole Moral Weight Project sequence in some detail, I have some uncertainties around the following question/objection that you list above:
“Your literature review didn’t turn up many negative results. However, there are lots of proxies such that it’s implausible that many animals have them. So, your welfare range estimates are probably high.”

In your response you write that this is a good objection.

However, as I understand it, whenever proxies were unknown, you assumed these to be zero (i.e. not present). For instance, in your methodology writeup, I read: "Assigning proxies labeled “Unknown” zero probability of being present is certainly leading to underestimates of the welfare ranges and probabilities of sentience."

Somehow I cannot square these two statements. Can you solve that seeming contradiction for me?

Thanks for the writeup. Not an area I know much about. Interested to hear what you think the priorities are for further research in this area. 

I liked the common questions & responses section - very helpful for someone like me who is new to this topic.

What surprised me - perhaps it shouldn't have done - is that you think it's plausible that some animals have a welfare higher than humans... 

so although I'm not worth only 3 chickens, the key takeaway is that I'm worth around 50 chickens, is that the deal?

I love this research! Thank you so much for doing it!

My gut reaction to the results is that it's odd that humans are so high up in terms of their capacity for welfare. Just as an uninformative prior, I would've expected us to be somewhere in the middle. Less confidently, I would've expected a similar number of orders of magnitude deviation from the human baseline in either direction, within reason. E.g. +/- ~.5 OOM.

Plus, we are humans, so there's a risk that we're biased in our favor. It could be simply a bias from our ability to emphasize with other humans. But it could also be the case that there are countless more markers of sentience that humans don't have (but many other sentient animals do) that we are prone to overlook.

Have you investigated what the sources of this effect might be? There might be any number of biases at work as I mentioned, but perhaps our lives have become so comfy most of the time that we perceive slight problems very strongly (e.g., a disapproving gaze). If then something really bad happens, it feels enormously bad?

(I've in the past explicitly assumed that most beings with a few (million) neurons have a roughly human capacity for welfare – not because I thought that was likely but because I couldn't tell in which direction it was off. Do you maybe already have a defense of the results for people like me?)

In any case, I'll probably just adopt your results into my thinking now. I don't expect them to change my priorities much given all the other factors.

Thank you again! <3

Update: When I mentioned this to a friend on a hike, I came up with two ways in which the criteria might be amended to include nonhuman ones: (1) In may cases, we probably have a theory for why a particular behavior or feature is likely to be indicative of conscious experience. Understanding this mechanism, we can look for other systems that might implement the same mechanism, sort of how the eyes of humans, eagles, and flies are very different but we infer that they are probably all for the purpose of vision. (2) Maybe a number of animals that show certain known criteria for consciousness also share suspiciously consistently some other features. One could then investigate whether these features are also indicative of consciousness and whether there are other animals that have these new features at the expense of the older, known ones. (The analysis could cluster features that usually co-occur to not overweight causally related features in cases where many of them are observable.)

Hi Bob,

Great work!

I think it would be nice to have all the estimates in the table here with 3 significant digits, in order not to propagate errors. I understand more digits may give a sense of false precision, but you provide the 5th and 95th percentiles in the same table, so I suppose the uncertainty is already being conveyed.

Why do you give estimates for the median moral weight, instead of the mean moral weight? Normally, we care about expectations...

What follows are some probability-of-sentience- and rate-of-subjective-experience-adjusted welfare range estimates.

The probability of sentience is multiplied through here, right? Some of these animals are assigned <50% probability of sentience but have nonzero probability of sentience-adjusted welfare ranges at the median. Another way to present this would be to construct the random variable that's 0 if they're not sentient, and then equal to the random variable representing their moral weight conditional on sentience. This would be your actual distribution of welfare ranges for the animal, accounting for their probability of sentience. That being said, what you have now might be more useful to represent a range of expected moral weights for (approximately) risk-neutral EV-maximizing utilitarians, to represent deep uncertainty or credal fragility.

The use of expected value doesn't seem useful here. Your confidence intervals are huge (95% confidence interval for pig suffering capacity relative to humans is between 0.005 to 1.031). Because the implications are so different across that spectrum (varying from basically "make the cages even smaller, who cares" at 0.005 to "I will push my nan down the stairs to save a pig" at 1.031) it really doesn't feel like I can draw any conclusions from this.

I'm curious whether you've indicated parental care is "present" or "absent" in bees, however, I have briefly checked the documents linked and couldn't find where that lives but maybe I missed it. Can anyone link to that documentation?

(Bees provide care to young, but it's primarily done by siblings, not parents, so it's considered alloparental care, not parental care. I should think that probably counts, but wasn't sure.)