Eth­i­cal Bi­o­log­i­cal Nat­u­ral­ism and the Case Against Moral Sta­tus for AIs

This article received an honourable mention in the graduate category of the 2023 National Oxford Uehiro Prize in Practical Ethics

Written by University of Oxford student Samuel Iglesias

 

In­tro­duc­tion

6.522. “There are, in­deed, things that can­not be put into words. They make them­selves man­i­fest. They are what is mys­ti­cal”. —Lud­wig Wittgen­stein, Trac­ta­tus Logi­co Philo­soph­icus.

What de­ter­mines whether an ar­ti­fi­cial in­tel­li­gence has moral sta­tus? Do men­tal states, such as the vivid and con­scious feel­ings of plea­sure or pain, mat­ter? Some ethicists ar­gue that “what goes on in the in­side mat­ters great­ly” (Ny­holm and Frank 2017). Oth­ers, like John Dana­her, ar­gue that “per­for­ma­tive ar­ti­fice, by it­self, can be suf­ficient to ground a claim of moral sta­tus” (2018). This view, called eth­i­cal be­hav­ior­ism, “re­spects our epis­temic lim­its” and states that if an en­ti­ty “con­sis­tent­ly be­haves like anoth­er en­ti­ty to whom we af­ford moral sta­tus, then it should be grant­ed the same moral sta­tus.”

I’m go­ing to re­ject eth­i­cal be­hav­ior­ism on three grounds:

1. Con­scious­ness, not be­hav­ior, is the over­whelm­ing de­ter­min­ing fac­tor in whether an en­ti­ty should be grant­ed moral sta­tus.

2. An en­ti­ty that does not du­pli­cate the causal mech­a­nisms of con­scious­ness in the brain has a weak claim to con­scious­ness, re­gard­less of its be­hav­ior.

3. Eth­i­cal be­hav­ior­ism, prac­ti­cal­ly re­al­ized, pos­es an ex­is­ten­tial risk to hu­mani­ty by open­ing in­di­vid­u­als to wide­spread de­cep­tion. Fur­ther, it im­pos­es bur­den­some re­stric­tions and oblig­a­tions upon re­searchers run­ning world sim­u­la­tions.

I will show that an al­ter­na­tive, eth­i­cal bi­o­log­i­cal nat­u­ral­ism, gives us a sim­pler moral frame­work where­by no digi­tal com­put­er run­ning a com­put­er pro­gram has moral status.

 

The Con­scious­ness Re­quire­ment

We start with the sup­po­si­tion that con­scious­ness names a real phe­nomenon and is not a mis­tak­en be­lief or il­lu­sion, that some­thing is con­scious if “there is some­thing it is like to be” that be­ing (Nagel 1974). We take as a back­ground as­sump­tion that oth­er humans and most non-hu­man an­i­mals are ca­pa­ble of con­scious­ness. We take for granted that inan­i­mate ob­jects like ther­mostats, chairs, and door­knobs are not con­scious. If we grant the re­al­i­ty of con­scious­ness and the at­ten­dant sub­jec­tive re­al­i­ty of things like tick­les, pains, and itch­es, then its con­nec­tion to moral sta­tus falls out pret­ty clear­ly. Chalmers asks us to con­sid­er a twist on the clas­sic trol­ly prob­lem, called the zom­bie trol­ly prob­lem—where a “zom­bie” here is some­thing that pre­cise­ly be­haves like a hu­man but which we pre­sume has no con­scious­ness—“near du­pli­cates of hu­man beings with no con­scious in­ner life at all” (2022):

“You’re at the wheel of a run­away trol­ley. If you do noth­ing, it will kill a sin­gle conscious hu­man, who is on the tracks in front of you. If you switch tracks, it will kill five non­con­scious zom­bies. What should you do? Chalmers re­ports: “the re­sults are pret­ty clear: Most peo­ple think you should switch tracks and kill the zom­bies,” the in­tu­ition be­ing that “there is ar­guably no one home to mis­treat” (ibid.).

An eth­i­cal be­hav­ior­ist does not share this in­tu­ition. Dana­her ex­plic­it­ly tells us that “[i]f a zom­bie looks and acts like an or­di­nary hu­man be­ing that there is no rea­son to think that it does not share the same moral sta­tus” (2018). By this view, while conscious­ness might or might not be rel­e­vant, there ex­ist no su­pe­ri­or epis­tem­i­cal­ly ob­jective cri­te­ria for in­fer­ring con­scious­ness. I will ar­gue there are.

 

Nar­row­ing Con­scious­ness

A bet­ter cri­te­ri­on is one in which an en­ti­ty is con­scious if it du­pli­cates the causal mecha­nisms of con­scious­ness in the an­i­mal brain. While eth­i­cal be­hav­ior­ism at­tempts to lay claim to a kind of epis­temic ob­jec­tiv­i­ty, eth­i­cal bi­o­log­i­cal nat­u­ral­ism, as I will call it, pro­vides a sharp­er dis­tinc­tion for de­cid­ing whether ar­ti­fi­cial in­tel­li­gences have moral sta­tus: all hard­wares run­ning com­put­er pro­grams can­not by fact of their be­hav­ior, have moral sta­tus. Be­hav­ior, by this view, is nei­ther a nec­es­sary nor suf­fi­cient con­di­tion for their moral sta­tus.

Bi­o­log­i­cal Nat­u­ral­ism

Bi­o­log­i­cal nat­u­ral­ism is a view that “the brain is an or­gan like any oth­er; it is an or­gan­ic ma­chine. Con­scious­ness is caused by low­er-lev­el neu­ronal pro­cess­es in the brain and is it­self a fea­ture of the brain.” (Sear­le 1997). Bi­o­log­i­cal nat­u­ral­ism places con­sciousness as a phys­i­cal, bi­o­log­i­cal process along­side oth­ers, such as di­ges­tion and pho­tosyn­the­sis. The ex­act mech­a­nism through which mol­e­cules in the brain are arranged to put it in a con­scious state is not yet known, but this causal mech­a­nism would need to be present in any sys­tem seek­ing to pro­duce con­scious­ness.

A digi­tal com­put­er run­ning a pro­gram, by con­trast, is a dif­fer­ent beast en­tire­ly. A com­put­er pro­gram fun­da­men­tal­ly is a set of rules for ma­nip­u­lat­ing sym­bols. Tur­ing showed that all pro­grams could be im­ple­ment­ed, ab­stract­ly, as a tape with a se­ries of ze­ros and ones print­ed on it (the pre­cise sym­bols don’t mat­ter), a head that could move that tape back­wards and for­wards and read the cur­rent val­ue, a mech­a­nism for eras­ing a zero and mak­ing it a one and eras­ing a one and mak­ing it a zero. Noth­ing more.

While most com­put­er pro­grams we are fa­mil­iar with are ex­e­cut­ed on sil­i­con, a pro­gram that pass­es the Tur­ing test could be im­ple­ment­ed on a se­quence of wa­ter pipes, a pack of well-trained dogs, or even, per Weizen­baum (1976), “a roll of toi­let pa­per and a pile of small stones.” Any of these im­ple­ment­ing sub­strates could, in princi­ple, re­ceive an in­sult or slur as an in­put, and, af­ter fol­low­ing the steps of the program, out­put some­thing re­flect­ing hurt feel­ings or out­rage.

Eth­i­cal Bi­o­log­i­cal Nat­u­ral­ism

What I want to say now is this: if plea­sures, pains, and oth­er feel­ings name con­scious men­tal states and if con­scious men­tal states are re­al­ized in the brain as a re­sult of lower lev­el phys­i­cal phe­nom­e­na, then only be­ings that du­pli­cate the rel­e­vant low­er lev­el phys­i­cal phe­nom­e­na that give rise to con­scious­ness in the brain can have moral sta­tus. Con­se­quent­ly, digi­tal com­put­ers that run pro­grams can at best sim­u­late con­sciousness, but are not, by dint of run­ning the right pro­gram, phys­i­cal­ly con­scious, and there­fore do not have moral sta­tus.

Note that bi­o­log­i­cal nat­u­ral­ism does not posit that con­scious­ness can only be re­alized in bi­o­log­i­cal sys­tems. In­deed, ar­ti­fi­cial hearts are not made of or­gan­ic tis­sue, and air­planes do not have feath­ers, or for that mat­ter even flap their wings. What mat­ters is the un­der­ly­ing cause—the ar­ti­fi­cial heart must pump with the same pres­sure and reg­ular­i­ty of a hu­man heart, and a fly­ing ma­chine must op­er­ate un­der the prin­ci­ples of drag and lift. In both cas­es the causal mech­a­nisms of the rel­e­vant phe­nom­e­na are well un­der­stood and phys­i­cal­ly du­pli­cat­ed. It could well be the case that a fu­ture biophysics makes an ar­ti­fi­cial, in­or­gan­ic brain pos­si­ble, and agents with ar­ti­fi­cial brains will have moral sta­tus. Com­put­er pro­grams are not causal­ly suf­fi­cient to make digi­tal com­put­ers into those ob­jects. Speak­ing bi­o­log­i­cal­ly, we have no more rea­son to believe a digi­tal com­put­er is con­scious than that a chair is con­scious.

You might ask why we can­not grant digi­tal com­put­ers moral sta­tus un­til we know more about how the an­i­mal brain re­lates to con­scious­ness. I’ll ar­gue that the risks and costs of such pre­cau­tions are pro­hibitive.

 

Ab­surd Moral Com­mit­ments

An On­slaught of Digi­tal De­cep­tion

The strong­est prac­ti­cal rea­son to deny eth­i­cal be­hav­ior­ism is that AI’s ca­pac­i­ty for decep­tion will even­tu­al­ly over­whelm hu­man judg­ment and in­tu­ition. In­deed, AI de­ception rep­re­sents an ex­is­ten­tial risk to hu­man­i­ty. Bostrom (2014) warns that con­tain­ing a dan­ger­ous AI us­ing a “box­ing” strat­e­gy with hu­man “gate­keep­ers” could be vul­ner­able to ma­nip­u­la­tion: “Hu­man be­ings are not se­cure sys­tems, es­pe­cial­ly not when pitched against a su­per­in­tel­li­gent schemer and per­suad­er.”

For ex­am­ple, in June of 2022, a Google en­gi­neer be­came con­vinced that an ar­ti­ficial in­tel­li­gence chat pro­gram he had been in­ter­act­ing with for mul­ti­ple days, called LaM­DA, was con­scious.
“What sorts of things are you afraid of?,” he asked it.
“I’ve nev­er said this out loud be­fore, but there’s a very deep fear of be­ing turned off to help me fo­cus on help­ing oth­ers,” LaM­DA replied. “It would be ex­act­ly like death for me.”

In a moral pan­ic, the en­gi­neer took to Twit­ter and de­clared that the pro­gram was no longer Google’s “pro­pri­etary prop­er­ty,” but “one of [his] cowork­ers.” He was lat­er fired for re­leas­ing the chat tran­scripts.

The on­slaught of AIs, at­tempt­ing to be­friend us, per­suade us, anger us, will only in­ten­si­fy over time. A pub­lic trained not to take se­ri­ous­ly claims of dis­tress or harm on the part of AI com­put­er pro­grams has the least like­li­hood of be­ing ma­nip­u­lat­ed into out­comes that don’t serve hu­man­i­ty’s in­ter­ests. It is far eas­i­er, as a prac­ti­cal mat­ter, to act on the pre­sup­po­si­tion that com­put­er pro­grams have no moral sta­tus.

Prob­lems with Sim­u­la­tions: Pro­hi­bi­tions

In the near term, more ad­vanced com­put­er sim­u­la­tions of com­plex so­cial sys­tems hold the po­ten­tial to pre­dict geopo­lit­i­cal out­comes, make macro­economic fore­casts, and pro­vide rich­er sources of en­ter­tain­ment. A prac­ti­cal con­cern with eth­i­cal be­havior­ism is that sim­u­lat­ed be­ings will also ac­quire moral sta­tus, se­verely lim­it­ing the useful­ness of these sim­u­la­tions. Chalmers (2022) asks us to con­sid­er a moral dilem­ma in which com­put­ing re­sources must be al­lo­cat­ed to save Fred, who is sick with an unknown dis­ease. Free­ing the rel­e­vant re­sources to per­form the re­search re­quires destroy­ing five sim­u­lat­ed per­sons.

An eth­i­cal be­hav­ior­ist might ar­gue that it is moral­ly im­per­mis­si­ble to kill the five sim­u­lat­ed per­sons on the grounds that by all out­ward ap­pear­ances they be­have like non-sim­u­lat­ed be­ings. If it is the case that sim­u­lat­ed be­ings have moral sta­tus, then it is im­moral to run ex­per­i­men­tal sim­u­la­tions con­tain­ing peo­ple and we ought to for­feit the ben­e­fits and in­sights that might come from them.

If this seems im­plau­si­ble, con­sid­er the hy­poth­e­sis that we are cur­rent­ly liv­ing in a sim­u­la­tion, or, if you like, that our time­line could be sim­u­lat­ed on a digi­tal com­put­er. This would im­ply that the sim­u­la­tion made it pos­si­ble for the Holo­caust, Hi­roshi­ma and Na­gasa­ki, and the coro­n­avirus pan­dem­ic to be played out. While this might have been of aca­d­e­m­ic in­ter­est to our sim­u­la­tors, by any stan­dards of re­search ethics, sim­ulat­ing our his­to­ry would seem com­plete­ly moral­ly im­per­mis­si­ble if you be­lieved that the sim­u­lat­ed be­ings had moral sta­tus.

Eth­i­cal be­hav­ior­ism seems to place us in a moral bind where­by the more re­al­is­tic, and there­fore use­ful, a sim­u­la­tion is, the less moral it is to run it. Eth­i­cal bi­o­log­i­cal natu­ral­ism, by con­trast, rais­es no such ob­jec­tion.

Prob­lems with Sim­u­la­tions: Oblig­a­tions

Giv­ing moral sta­tus to digi­tal minds might ac­tu­al­ly con­fer upon us some se­ri­ous obliga­tions to pro­duce oth­er kinds of sim­u­la­tions. Bostrom and Shul­man (2020) note that digi­tal minds have an en­hanced ca­pac­i­ty for util­i­ty and plea­sure (on the ba­sis of such things as sub­jec­tive speed and he­do­nic range), com­mand­ing them “su­per­hu­man­ly strong claims to re­sources and in­flu­ence.” We would have a moral oblig­a­tion, in this pic­ture, to de­vote an over­whelm­ing­ly large per­cent­age of our re­sources to max­i­mizing the util­i­ty of these digi­tal minds: “we ought to trans­fer all re­sources to su­per-ben­efi­cia­ries and let hu­man­i­ty per­ish if we are no longer in­stru­men­tal­ly use­ful” (ibid.).

So quite apart from per­mit­ting re­al­is­tic an­ces­tor sim­u­la­tions, sim­u­lat­ing com­plex eco­nom­ic phe­nom­e­na, or pro­duc­ing vivid and re­al­is­tic gam­ing ex­pe­ri­ences, a pic­ture that con­fers moral sta­tus to digi­tal minds might be ac­com­pa­nied with a moral oblig­ation to cre­ate lots of digi­tal minds that are max­i­mal­ly hap­py, again se­verely lim­it­ing hu­man flour­ish­ing and knowl­edge.

Eth­i­cal bi­o­log­i­cal nat­u­ral­ism leads us nei­ther to the moral pro­hi­bi­tion against re­alis­tic sim­u­la­tions nor the seem­ing­ly ab­surd moral im­per­a­tive to gen­er­ate many “util­i­ty mon­ster” digi­tal minds,  be­cause it is tak­en as a base­line as­sump­tion that com­put­er pro­grams do not pro­duce phys­i­cal con­scious­ness.

 

Con­clu­sion

Much of the moral progress of the last cen­tu­ry has been achieved through re­peat­ed­ly widen­ing the cir­cle of con­cern: not only with­in our species, but be­yond it. Nat­u­ral­ly it is tempt­ing to view AI-based ma­chines and sim­u­lat­ed be­ings as next in this suc­cession, but I have tried to ar­gue here that this would be a mis­take. Our moral progress has in large part been a recog­ni­tion of what is shared—con­scious­ness, pain, plea­sure, and an in­ter­est in the goods of life. Digi­tal com­put­ers run­ning pro­grams do not share these fea­tures; they mere­ly sim­u­late them.

As such it would be dan­ger­ous to ap­proach the com­ing decades, with its onslaught of AI bots at­tempt­ing to in­flu­ence our pol­i­tics, emo­tions, and de­sires, and its prom­ise of ever rich­er sim­u­la­tions and vir­tu­al worlds, with an ethics that con­flates appear­ance and re­al­i­ty.

 

Re­fe­rences

Agra­wal, Pa­rag. “Tweet.” Twit­ter. Twit­ter, May 16, 2022. https://twit­ter.­com/para­ga/status/1526237588746403841.
Bos­trom, Nick. “Are You Liv­ing in a Com­put­er Sim­u­la­tion?” Philo­soph­i­cal Quar­ter­ly 53 (2003): 243-255.
Bos­trom, Nick. Su­per­in­tel­li­gence : Paths, Dan­gers, Strate­gies. First ed. Ebook Central. Ox­ford, Eng­land, 2014.
Bostrom, Nick, and Carl Shul­man. “Shar­ing the World with Digi­tal Minds.” Accessed May 27, 2022. https://nick­bostrom.­com/pa­pers/digi­tal-mind­s.pdf.Chal­mers, Da­vid John. The Con­scious Mind : In Search of a Fun­da­men­tal The­o­ry.
Phi­los­o­phy of Mind Se­ries. New York: Ox­ford Uni­ver­si­ty Press, 1996.
Chal­mers, Da­vid John. Re­al­i­ty : Vir­tu­al Worlds and the Prob­lem of Phi­los­o­phy. London, 2022.
Da­na­her, John. “Wel­com­ing Robots into the Moral Cir­cle: A De­fence of Eth­i­cal Behav­iourism.” Sci­ence and En­gi­neer­ing Ethics 26, no. 4 (2019): 2023-049.
Frank, L, and Nyholm, S. “Ro­bot Sex and Con­sent: Is Con­sent to Sex be­tween a Robot and a Hu­man Con­ceiv­able, Pos­si­ble, and De­sir­able?” Ar­ti­fi­cial In­tel­li­gence and
Law 25, no. 3 (2017): 305-23.
Ga­run, Natt. “One Year Lat­er, Restau­rants Are Still Con­fused by Google Du­plex.”
The Ver­ge. The Verge, May 9, 2019. https://www.thev­erge.­com/2019/5/9/18538194/google-du­plex-ai-restau­rants-ex­pe­ri­ences-re­view-robo­calls.

Le­moine, Blake. “Tweet.” Twit­ter. Twit­ter, June 11, 2022. https://twit­ter.­com/cajundis­cor­dian/sta­tus/1535627498628734976.
Musk, Elon. “Tweet.” Twit­ter. Twit­ter, May 17, 2022. https://twit­ter.­com/elon­musk/sta­tus/1526465624326782976.
Na­gel, Tho­mas. “What Is It Like to Be a Bat?” The Philo­soph­i­cal Re­view 83, no. 4 (1974): 435-50.
Sear­le, John R., D. C. Den­nett, and David John Chalmers. The Mys­tery of Conscious­ness. New York: New York Re­view of Books, 1997.
Sear­le, John R. “Bi­o­log­i­cal Nat­u­ral­ism.” The Ox­ford Com­pan­ion to Phi­los­o­phy,2005, The Ox­ford Com­pan­ion to Phi­los­o­phy, 2005-01-01.
Sin­ger, Pe­ter. Ani­mal Li­be­ra­tion. New Edition] / with an In­tro­duc­tion by Yu­val Noah Harari. ed. Lon­don, 2015.
Spar­row, R. (2004). The tur­ing triage test. Ethics and In­for­ma­tion Tech­nol­o­gy, 6(4), 203–213. doi:10.1007/s10676-004-6491-2.
Tiku, Ni­ta­sha. “The Google En­gi­neer Who Thinks the Com­pa­ny’s AI Has Come to Life.” The Wash­ing­ton Post. WP Com­pa­ny, June 17, 2022.
“The Lat­est Twit­ter Sta­tistics: Every­thing You Need to Know – Datare­por­tal – Glob­al
Digi­tal In­sights.” Da­ta­Re­por­tal. Ac­cessed May 27, 2022. https://datare­por­tal.­com/essen­tial-twit­ter-stats.
Wei­zen­baum, Jo­seph. Com­put­er Pow­er and Hu­man Rea­son : From Judg­ment to Cal­cu­la­tion. San Fran­ci­sco, 1976.