Changes in conditions are a priori bad for average animal welfare

Sum­mary: What­ever your be­liefs about the ex­pected value of the av­er­age welfare of an­i­mals in the wild un­der con­stant con­di­tions and at equil­ibrium, I think you should ex­pect, a pri­ori (weakly, with­out fur­ther ev­i­dence), their av­er­age welfare to be lower af­ter con­di­tions change. If you be­lieve that un­der con­stant con­di­tions and at equil­ibrium, the ex­pected value of the av­er­age welfare in the wild is at most (or, per­haps by re­source effi­ciency and sym­me­try, equal to) 0, then you should be­lieve, a pri­ori, that un­der chang­ing con­di­tions, it is nega­tive, and so, with it, the to­tal welfare would also be nega­tive in ex­pec­ta­tion. Since con­di­tions are con­stantly chang­ing, you should ex­pect, a pri­ori, the net welfare in the wild to be nega­tive.

The same ar­gu­ment should ap­ply a pri­ori to farmed an­i­mals, as well as sen­tient AIs de­vel­oped with­out welfare in mind and op­ti­mized for a spe­cific pur­pose through some op­ti­miza­tion al­gorithm, if their af­fec­tive (re­ward and pun­ish­ment) sys­tems are also op­ti­mized for that pur­pose.

Dis­claimers: I have no for­mal back­ground in evolu­tion or ecol­ogy past high school, so there’s a good chance I’m wrong or even very wrong here. My for­mal back­ground is in math and com­puter sci­ence. I also have suffer­ing-fo­cused views, so this might bias me to­wards mak­ing wild an­i­mal welfare look worse to those with more sym­met­ric views, and that was one of my mo­ti­va­tions for writ­ing this post.


This is effec­tively the Anna Karen­ina prin­ci­ple (or The Prin­ci­ple of Frag­ility of Good Things).

I think this con­clu­sion is already pretty ob­vi­ous enough, but it’s worth ex­pand­ing.

The a pri­ori analy­ses of the sign of the to­tal/​net welfare of an­i­mals in the wild that I’m aware of take as im­plicit the as­sump­tion that pop­u­la­tions are in equil­ibrium and un­chang­ing, and the con­di­tions in which they live aren’t chang­ing ei­ther. If our prior be­liefs about the net welfare were cap­tured by a dis­tri­bu­tion with ex­pected value 0 (by sym­me­try) or nega­tive, could we not have rea­son to be­lieve that un­der chang­ing con­di­tions (per­haps in­clud­ing cy­cles, al­though you’d ex­pect some adap­ta­tion in many cases, e.g. to weather cy­cles), an­i­mal welfare will tend to be lower on av­er­age? Of course, this will in each case de­pend on the par­tic­u­lar changes — many changes can in­deed be good for welfare — but we might ex­pect them to usu­ally be bad, be­cause changes are of­ten away from the con­di­tions for which the pop­u­la­tion is adapted, since an­i­mals are adapted to spe­cific con­di­tions which might have “sweet spots”.

The argument

Here is my ex­panded ar­gu­ment, which de­pends on some in­tu­itive (but not nec­es­sar­ily tech­ni­cal) un­der­stand­ing of the ba­sics of con­tin­u­ous op­ti­miza­tion:

Evolu­tion is an op­ti­miza­tion al­gorithm with a pos­si­bly mov­ing tar­get: the ul­ti­mate tar­get is the pro­lifer­a­tion of genes, but we can use the evolu­tion­ary fit­ness of a pop­u­la­tion as a proxy and a func­tion of ge­net­ics, and this goal is speci­fied un­der par­tic­u­lar con­di­tions, so with chang­ing con­di­tions, which solu­tions are bet­ter can change, too. Be­cause evolu­tion is an op­ti­miza­tion al­gorithm, you ex­pect it to pre­fer small changes to its pop­u­la­tions’ ge­net­ics which in­crease fit­ness to small changes which de­crease fit­ness. You would ex­pect it to be more likely to pass through sad­dle points to­wards in­creased fit­ness than to de­creased fit­ness, and to avoid pro­duc­ing lo­cal min­ima for the cur­rent con­di­tions (ex­cept pos­si­bly in very flat re­gions of the func­tion).

Now, sup­pose evolu­tion had pro­duced a solu­tion (not nec­es­sar­ily op­ti­mal) of pop­u­la­tion genes un­der equil­ibrium for a given fixed set of con­di­tions. Sup­pose now that the con­di­tions change slightly. I ex­pect these chang­ing con­di­tions, a pri­ori, to be bad for the av­er­age welfare of that pop­u­la­tion. I’ll illus­trate with two ex­am­ples, for which I con­sider what’s good or bad for the welfare of in­di­vi­d­u­als who live through the changes in con­di­tions (and not the gen­er­a­tion’s offspring, who may be bet­ter adapted), and I as­sume av­er­age fit­ness and av­er­age welfare for a given pop­u­la­tion un­der differ­ent con­di­tions cor­re­late lo­cally (enough to use them in­ter­change­ably):

a. Change in nu­tri­tion qual­ity/​abun­dance. If it in­creases, this is good. If it de­creases, this is bad.

b. Change in tem­per­a­ture. If it in­creases, this is bad, since it in­creases the risk of hy­per­ther­mia. If it de­creases, this is bad, since it in­creases the risk of hy­pother­mia. This is be­cause evolu­tion will aim to tune the body tem­per­a­tures and heat reg­u­la­tion of a pop­u­la­tion to the cur­rent con­di­tions (or a given range of con­di­tions, given cycli­cal weather pat­terns), and un­der differ­ent con­di­tions, the solu­tion may do too lit­tle (e.g. not enough fur) or too much (e.g. too much fur). Later gen­er­a­tions may be bet­ter or worse off, though, since they may need to use fewer or more re­sources for tem­per­a­ture reg­u­la­tion.

No­tice that in the first case, it can be ei­ther good or bad, but in the sec­ond, it only looks bad. Of course, other con­sid­er­a­tions might lead us to be­lieve that a change in tem­per­a­ture is ac­tu­ally good in one di­rec­tion, but bad in the other, e.g. it might af­fect nu­tri­tion qual­ity/​abun­dance or al­low an­i­mals to use their en­ergy more effi­ciently. For cases like a., we should a pri­ori ex­pect the pos­si­ble good from it to match the bad, on av­er­age, but this will de­pend on speci­fics and per­haps even parametriza­tion and scales. Here, “di­men­sion” and “di­rec­tion” can be com­bi­na­tions of differ­ent fac­tors, e.g. tem­per­a­ture and food abun­dance.

How­ever, and this is the cru­cial claim: for a given sta­ble solu­tion for a fixed set of con­di­tions, we should ex­pect small changes in the con­di­tions across one di­men­sion that are good in both di­rec­tions to be less likely than small changes in the con­di­tions across one di­men­sion which are bad in both di­rec­tions (like b). This breaks the sym­me­try be­tween good and bad changes, and im­plies changes should a pri­ori be bad in ex­pec­ta­tion.

Per­son­ally, I have not been able to even think of a di­men­sion in con­di­tions ac­cord­ing to which a change in ei­ther di­rec­tion would be good, but this could just be my own ig­no­rance. Please com­ment with some if you do think of them. I also sus­pect that such a solu­tion would be less sta­ble from a pop­u­la­tion ge­net­ics stand­point, if fit­ness-im­prov­ing changes in ge­net­ics can al­ign with changes in con­di­tions. I sus­pect it’s pos­si­ble to make this claim more for­mal and prove a form of it math­e­mat­i­cally.


1. If con­di­tions de­crease pop­u­la­tion sizes, even if the av­er­age welfare de­creases, the to­tal welfare may in­crease or de­crease.

2. I think we can do much bet­ter than rely­ing on a sym­met­ric prior and mak­ing judge­ments about the net bal­ance of plea­sure and suffer­ing in the world (or par­tic­u­lar cases) ap­peal­ing to it and lit­tle else (when plea­sure and pain are not used for guid­ing ac­tion but sim­ply in­duced in ar­tifi­cial minds, we might think this kind of sym­me­try in en­ergy effi­ciency could hold). Life his­tory clas­sifi­ca­tion can bet­ter in­form our judge­ments, see:

“In­fant Mor­tal­ity and the Ar­gu­ment from Life His­tory” by Ozy Bren­nan
”Life his­tory clas­sifi­ca­tion” by Kim Cud­ding­ton
”In­sect her­bivores, life his­tory and wild an­i­mal welfare” by Kim Cuddington

Also, (sub­jec­tively) ag­gre­gat­ing differ­ent welfare in­di­ca­tors as in:

“From hu­mans in Canada to bat­tery caged chick­ens in the United States, which an­i­mals have the hard­est lives: re­sults” by Char­ity Entrepreneurship

3. An op­po­site prin­ci­ple might be an­tifrag­ility. I think this would only hold for small changes, and if con­di­tions con­tinue to change, they can out­pace pop­u­la­tion adap­ta­tion, and this would still be bad.

4. “How Much Do Wild An­i­mals Suffer? A Foun­da­tional Re­sult on the Ques­tion is Wrong.” by Zach Groff (EA Fo­rum post, EA Global talk, tran­scripts) for dis­cus­sion of the two pa­pers:

The cor­rec­tion to the origi­nal: “Does suffer­ing dom­i­nate en­joy­ment in the an­i­mal king­dom? An up­date to welfare biol­ogy” (2019) by Zach Groff and Yew‑Kwang Ng”

The origi­nal: “Towards welfare biol­ogy: Evolu­tion­ary eco­nomics of an­i­mal con­scious­ness and suffer­ing” (1995) by Yew-Kwang Ng.

5. I spent a bit of time think­ing about this in terms of a fit­ness func­tion of con­di­tions and pop­u­la­tion ge­net­ics, small changes in con­di­tions and ge­net­ics, par­tial deriva­tives and di­rec­tional deriva­tives. This might still be a promis­ing ap­proach, and I might try again even­tu­ally, but it’s not a pri­or­ity for me now, and it would prob­a­bly be bet­ter off in the hands of some­one with more back­ground in ecol­ogy or evolu­tion.