Moral weights for various species and distributions
This analysis estimates the expected moral weight of the beings of various species relative to humans for various types of moral weight distributions.
The mean moral weight is close to 1 for all the considered species, ranging from 0.5 to 5 excluding the lognormal and pareto distributions (for which it is even higher, but seemingly inaccurate).
I welcome comments about how to interpret the results.
The expected moral weight of the beings of various species relative to humans was determined from the product between:
The probability of the beings of the species having moral patienthood, as defined by Luke Muehlhauser here, which was set to the values provided in this section of Open Philanthropy’s 2017 Report on Consciousness and Moral Patienthood.
The mean of a distribution whose 10th and 90th percentiles were set to the lower and upper bounds of the “80 % prediction interval” guessed by Luke Muehlhauser here for the moral weight of various species relative to humans conditional on the respective beings having moral patienthood (see “Moral weights of various species”).
The mean of the distribution was computed from the quantiles as described here.
The expected moral weight might depend on the theory of consciousness. The above product is implicitly assumed to represent the expected weighted mean of the moral weight distributions of the various theories of consciousness. These are, in turn, supposed to produce (summable) moral weight distributions. Potential concerns about calculating expected moral weights are discussed here.
The mean and median moral weight of various species relative to humans for uniform, normal, loguniform, lognormal, pareto and logistic distributions were calculated here, and are presented in the tables below.
|Species||Mean moral weight relative to humans|
|Rainbow trouts||4.55||4.55||3.00||28.4 k||4.55|
|Fruit flies||2.50||2.50||1.95||2.46 M||2.50|
|Species||Median moral weight relative to humans|
The results suggest animals and humans have a similar moral value. The mean moral weight is close to 1 for all the considered species, ranging from 0.5 to 5 excluding the lognormal and pareto distributions.
The lognormal distributions do not seem to represent the moral weights accurately. Their heavy right tails imply high mean moral weights, which would arguably require frequent strong experiences. However, as noted here by Jason Schukraft, “it appears unlikely that evolution would select for animals with a non-contiguous range that was exclusively extraordinarily strong because extremely intense experiences are distracting in a way that appears likely to reduce fitness”.
The pareto distributions are not reasonable representions of the moral weights, as they lead to mean moral weights of infinity.
Loguniform distributions appear to be the best choice amongst the 6 studied types of distributions:
Being positive, they prohibit negative moral weights.
Having mean larger than the median, they are compatible with the intuition that the moral weight is a product (not a sum) of multiple dimensions (for example, clock speed of consciousness, unity of consciousness, and unity-independent intensity of valenced aspects of consciousness).
Being bounded, they prevent unreasonably large mean moral weights.
The probability of pigs being moral patients is not provided in this section of Open Philanthropy’s 2017 Report on Consciousness and Moral Patienthood. However, it was assumed to be equal to that of cows and chickens (80 %).
1 equals .
- Corporate campaigns for chicken welfare are 10,000 times as effective as GiveWell’s Maximum Impact Fund? by 28 Jul 2022 8:22 UTC; 76 points) (
- A peek at pairwise preference estimation in economics, marketing, and statistics by 8 Oct 2022 4:56 UTC; 30 points) (
- 29 Jul 2022 10:11 UTC; 2 points)'s comment on Corporate campaigns for chicken welfare are 10,000 times as effective as GiveWell’s Maximum Impact Fund? by (
Since this exercise is based on numbers I personally made up, I would like to remind everyone that those numbers are extremely made up and come with many caveats given in the original sources. It would not be that hard to produce numbers more reasonable than mine, at least re: moral weights. (I spent more time on the “probability of consciousness” numbers, though that was years ago and my numbers would probably be different now.)
To pick a bit on the notion from this article which establishes the range of moral weights in question.
They say fruit flies range from a moral weight of .000001 − 20 times the moral significance of human experience. In log space, that’s between 10^-6 and 10^1.3. The mean log uniform distribution, as you mention, is at 1.95. I find the significant probability mass being above 1 as implausible for fruit flies, and I will go on to explain why I think except for species like dogs, pigs, elephants, octopuses, or other long living intelligent social creatures it would be difficult to argue that they are plausibly more moral weight than a human.
Arguments for fruit flies being about as likely to be more morally significant than humans as less:
A fruit fly may experience things much faster than a human, meaning their short lives may be experienced longer than how long a human may perceive their lives to be.
They may experience things more intensely as well, given their single pointed focus on conscious experience. This means although it’s possible they experience suffering more fully, it may also be possible that they can completely forget and move their focus to some other task if focusing on the pain does not confer an advantage.
They may be less “distracted” than humans, in that they experience the world more fully and in full awareness.
They are also typically considered innocent of other moral wrongdoings, so perhaps that makes them more morally valuable in some moral systems.
There are a whole host of reasons to think that they could not possibly be as morally significant as a human.
I think it’s reasonable to say a fruit fly cannot remember things in the long term, and it cannot contemplate or ruminate, which is one of the worst aspects of negative experiences and pain. I think most people would prefer to have experiences of extreme pain and trauma erased from their lives.
A fruit fly lives a tiny fraction of the duration of a human’s life, so it would have to experience its own life much faster.
A human can be considered an ensemble or family of different personalities and conscious processes. Each one of these may have moral significance, increasing the relative moral significance of a human.
The more complex something is, typically, it is more valued in generic terms.
Humans form a network of social connections and social connections. When a human is lost, their loss is understood and grieved by many other humans, thus greatly increasing the overall negative effect of harm to a human compared to a fruit fly.
Humans have very few children relative to fruit fly, so they are likely value higher on an individual level by their families and communities.
In summary the most relevant factors for moral significance are likely the degree of social embeddedness, the experience of higher order emotions and complexity in general, the ability to grieve, long lives, and long memories, which strongly implies that humans are more morally significant than all or most other animals.
A final thought is that we don’t know with very high confidence that animals are conscious in the way that we care about morally, but we know this for sure with humans. For that reason, we would be safer to prefer to save humans first, in case we were wrong about animals having conscious experiences in the first place.
Note the median moral weight for fruit flies assuming a loguniform distribution (the type I prefer) is 0.00192 << 1. So I do not think the moral weight of fruit flies relative to humans being smaller than 1 is as likely as it being larger than 1.
Based on this analysis from Jason Schukraft, “mental time travel [“the capacity to remember past events and imagine future events”] seems to reduce the intensity of experiences in some circumstances and amplify the intensity of experiences in other circumstances. It is thus unclear whether animals that possess this ability have characteristically more or less intense valenced experiences overall” (see this section for details).
The moral weights presented here have units QALY/aQALY (QALY per “animal QALY”), and therefore they are not affected by differences in life expectancy between species. For example, a moral weight of 2 QALY/cQALY (QALY per “chicken QALY”) means that 2 T years of fully healthy human life are as valuable as T years of fully healthy chicken life.
I tend to agree. From Jason’s analysis (see here), “species that experience a greater variety and/or greater complexity of emotional states are, all else equal, capable of more intense positive and negative experiences”.
From the “Key Highlights” of Jason’s analysis:
So I tend to agree with your point, and think this is a good argument for not trusting mean moral weights which are much larger than 1. For the luguniform distributions, my maximum mean moral weight is 3, which is not much larger than 1.
I agree, and think this should be considered when comparing interventions. That being said, these points do not influence the moral weight, which is the ratio between the value of T years of fully healthy animal life and T years of fully healthy animal life (i.e. the duration of the experiences is normalised).
This is taken into account here by multiply the moral weight given moral patienthood by:
In terms of your summary:
I think your conclusion may well be right, but there is lots of uncertainty, so I do not think there is a “strong implication”. For example, I think the likelihood of the moral weight being larger than 1 is at least 10 %, so the mean moral weight should be larger than 0.1.
As a disclaimer, I came in with the preconception that one should assign near-zero probability of animals being of more moral relevance than humans.
After reading the arguments, I have found little to no convincing arguments contradicting this.
It’s true that we should be uncertain as to how animals experience the world. However, I don’t feel that the uncertainty in moral value should be thought of as ever exceeding human’s moral value.
To illustrate my current understanding of the best way to think about this topic, I think all your probability distributions should probably be modeled as never exceeding 1 for every animal, as the probability of such an outcome is so low it’s not worth considering. I think of it like the probability that you can build a perpetual energy-creating machine violating the laws of physics, or the probability that tomorrow the sun does not rise because the earth stopped rotating.
Perhaps, it could analogized as the same moral probability that causing suffering is a good thing, all things considered. One might argue that the human brain is extremely complicated, and morality is complicated, so we should put some weight on moral views that prefer to cause infinite suffering for eternity. Perhaps one could argue that some people enjoy causing others to suffer, and they might be right, and so suffering might be intrinsically good. I think this argument has about as much supporting evidence as the concept that animals could be more morally relevant than people. However, again, I would say the probability of such an outcome is so low it’s not worth considering.
Although it’s true we do not know the details of how animals experience consciousness, this is not enough to overturn the intuition all humans share about the morality of killing people versus animals—one is simply entirely different than another, and there is no instance in which it is better to kill an animal than a person. This conception has apparently been held constant for many cultures throughout human history. In some cases some animals were revered as gods, but this was less about the animals and more about the gods. In some cases animals and living things were seen as equally valuable as humans. I think this is unlikely, but not impossible, but the key point is that killing was seen as wrong in all cases, and not that animals were seen as more valuable than humans.
Suffering is not the only relevant moral consideration. See “The Righteous Mind” by Jonathon Haidt—humans probably share a few more moral foundations than purely care/harm, including authority, fairness, sanctity, etc. Some may view these as equally morally relevant. My point is here, it’s questionable whether we have equal moral responsibility over nonhuman animals as we have to humans, depending on how you construct your moral frameworks. If you look at how human brains are wired, the foundations of our conceptions of morality are built with in-group vs out-group. So, the moral status of animals based on understanding of human psychology which is our best way to guess at a “correct” moral framework would indicate that as things become less like us, our moral intuitions will guide us as valuing these things less.
I think you may have come to your probability distributions because you are a sequence thinker and are using your intuitions to argue for each part of a sequence which comes to some conclusion, where the proper thing to do when coming to some conclusion about whether to spend on an animal welfare charity or not is to use cluster-style thinking.
I hope that this is seen as a respectful difference in perspective and not at all a personal attack. I think it is useful to question these sort of assumptions to make moral progress, but I also think we need a lot of evidence to overturn the assumption that humans are more or equally morally relevant than animals, in large part due to the pre-existing moral intuitions we all probably share. There don’t appear to be sufficient arguments out there to overturn this position.
Okay, that was enough philosophizing, let me put in a few more points in favor of my position here:
Most people I know that are smarter than me believe humans are more morally significantly than animal. I know of zero people seriously arguing the opposite side
If morality is actually all fake and a human invention with no objective truth to it, then humans and animals will both be worth zero, and I will still be correct.
The actual actions of people who argue animals are more morally relevant than humans is not to kill people to save animals, so there’s probably no-one who sincerely, deep down believes this
People tend to anthropomorphize other things like teddy bears and Roombas and things like that, and mistakenly assign them some moral worth until they think about it more. Therefore, our intuitions can tend to guide us to incorrect conclusions about what is morally worthwhile.
Thanks for clarifying!
This is an interesting exercise. I imagine that Luke’s estimates were informed by his uncertainty about multiple incompatible theories / considerations and so any smooth distribution won’t properly reflect the motivations that lead to those estimates. Do you think these results suggest anything about what a lumpy distribution would say?
Thanks for commenting!
If the moral weight distribution is based on 2 theories A and B which produce distributions (lumps) MWA and MWB, and we think A and B are equally valid, the moral weight and expected moral weight would be:
MW = 0.5 MWA + 0.5 MWB.
E(MW) = 0.5 E(MWA) + 0.5 E(MWB)
It is unclear whether this expected moral moral weight would be smaller/larger than the one of the continuous case. Luke only provides 2 quantiles, but MWA and MWB are defined by 4 parameters (assuming 2 for each).
Eventually, to derive the 4 paramers, one could further assume that the variance of MWA equals that of MWB, and that the median of MW equals the arithmetic/geometric mean between Luke’s lower and upper bound. However, I do not know whether these assumptions make sense.
It’s awkward to interpret mathematical judgements about a value that is described as an unknown and then as a supposition about one’s internal process of deciding an arbitrary value for the unknown and finally as a possible range varying over a large magnitude for that unknown. That is what I decided that the report on consciousness (and the speculation about moral weights) describes.
I would like to learn more about how EA folks typically assign evidence for the presence of different kinds of consciousness or moral weight of different species. In particular, what evidence helps you decide the presence of different aspects of consciousness in specific amounts? What evidence helps you decide the moral weight of a person of one species relative to another?
Finally, What is EA speculation about more traditional models of morality that rely on a moral identity, judgements of right and wrong, and in particular, the symbolic importance of actions, even when they have (potentially) minimal verifiable consequences for others (for example, catching a fly and releasing it outside)?
Thanks for commenting!
In Luke’s post, clock speed of consciousness, unity of consciousness, and unity-independent intensity of valenced aspects of consciousness are the factors based on which the quantiles for the moral weight were defined, I think.
Personally, I put more than 90 % of weight on total hedonic utilitarianism (classical utilitarianism). However, in practice, the full consequences of my actions are really hard to measure, so I very often (always?) rely on heuristics to decide what to do (especially when there are “minimal verifiable [or measurable] consequences”).
Note that cost-effectiveness analyses or other quantitative methods are still heuristics, not definitive answers, because they are always incomplete.
I took the clock speed, unity, and intensity factors to be the aspects of consciousness about which one gathered evidence.
Total hedonic utilitarianism is mathematically interesting. I should explore its logical implications.
I appreciate what you describe as heuristics. In my everyday life I apply heuristics.
Morality is informed by heuristics that determine consequences of actions or by heuristics that determine the symbolic content of actions (their subjective or intersubjective meaning).
EDIT: morality is also informed by heuristics that determine intentions of actions, irrespective of consequences of actions, but that was not my interest here.
I wonder what heuristics the EA community officially acknowledge as relevant to understanding the level of consciousness or moral weight of beings from other species.