Book Review: Atlas of AI

Review of: Kate Crawford, Atlas of AI (New Haven: Yale University Press, 2021).

Atlas of AI is a time­ly book by Kate Crawford cov­er­ing the var­i­ous im­pli­ca­tions of AI to­day and through­out his­to­ry. There is in­tense pub­lic scruti­ny on the is­sues cov­ered in this book: labour, en­er­gy con­sump­tion, cli­mate change, dis­crim­i­na­tion, pri­va­cy, and gov­ern­ment ac­count­abil­i­ty. However, these con­cerns are of­ten over­shad­owed by the gen­er­al en­am­our­ment with tech­no­log­i­cal ar­ti­facts like Amazon Alexa, Google Nest, Tesla, and mo­bile de­vices. This book sit­u­ates AI in the phys­i­cal and so­cial world. Crawford shows us the ways that AI ex­ploits and changes our re­la­tion­ships with the Earth and with each oth­er.

Crawford cov­ers many of the canon­i­cal ex­am­ples and is­sues flowing from AI. You will see many of the ex­am­ples from this book also ref­er­enced in any cur­rent writ­ing that views the field through a le­gal, so­cial, or po­lit­i­cal lens. The lat­ter chap­ters are an es­pe­cial­ly good pre­sen­ta­tion of AI as a tool of state pow­er. While I found this to be a valu­able, scope-defining work for peo­ple new­ly think­ing about the con­se­quences of AI use and de­vel­op­ment, I did find por­tions of it repet­i­tive, vague, or not well-con­nect­ed to a cen­tral the­sis. These are easy enough to avoid or skim over.

Summary

This book opens with an im­age of two peo­ple hold­ing up what ap­pears to be a main­frame com­put­er on their shoul­ders, them­selves crouched on top of a frag­ment of the Earth. The com­put­er is pro­ject­ing, through a lens, a dis­tort­ed im­age of the Earth’s sur­face. That pro­jec­tion sits above the scene.

This im­age neat­ly cap­tures the themes in this book. AI ex­ploits hu­man labour and re­sources ex­tract­ed from the Earth. The con­cep­tion of AI as dis­em­bod­ied, while some­times a help­ful ab­strac­tion—and high­lights one way in which AI might nev­er be Really Intelligent—serves to ob­scure the ma­te­r­i­al needs and effects of AI. This is the cen­tral theme of the first two chap­ters, “Earth” and “Labor.” The mid­dle of the book fo­cus­es on the com­put­er sci­ence of AI and ma­chine learn­ing. AI, es­pe­cial­ly ma­chine learn­ing, trans­forms data through its op­er­a­tion. It cre­ates pro­jec­tions of re­al­i­ty into dig­i­tal rep­re­sen­ta­tions. These are nec­es­sar­i­ly an ap­prox­i­ma­tion, with the goal that the rep­re­sen­ta­tion pre­serves fea­tures that are use­ful for the in­tend­ed task. In the im­age above, this ap­prox­i­ma­tion ap­pears as a butterfly map, a pro­jec­tion that presents more ac­cu­rate­ly the dis­tances be­tween lo­ca­tions in north­ern lat­i­tudes. The pro­jec­tion ex­ag­ger­ates the dis­tance and in­tro­duces stark dis­con­ti­nu­ities be­tween land mass­es in the south­ern hemi­sphere. Of course all map pro­jec­tions onto a flat sur­face must make tradeoffs, but Crawford’s se­lec­tion of this par­tic­u­lar pro­jec­tion reflects a theme de­vel­oped in the lat­ter por­tions of this book: “jin­go­is­tic idea of sov­er­eign AI”Kate Crawford, Atlas of AI (New Haven: Yale University Press, 2021) at 208. and ethics frame­works that are “over­whelm­ing­ly pro­duced by eco­nom­i­cal­ly de­vel­oped coun­tries, with lit­tle rep­re­sen­ta­tion from Africa, South and Central America, or Central Asia.”Ibid at 224.

Earth

The first chap­ter high­lights the phys­i­cal dam­age to the Earth when we ex­tract the ma­te­ri­als need­ed for our tech­nol­o­gy: e.g. lithi­um mines in Nevada, rare earth el­e­ments of­ten lo­cat­ed in emerg­ing mar­kets and conflict zones, and la­tex need­ed for transat­lantic ca­bles that is ex­tract­ed from a species of tree in Southest Asia. That is just some of what is re­quired to cre­ate our phys­i­cal tech­no­log­i­cal tools. Crawford also high­lights the on­go­ing en­er­gy needs to use these tools. Training very large deep neur­al net­works is a promi­nent ex­am­ple.

Labor

The chap­ter on Labor shows us how hu­man labour is both con­trolled by AI and ex­ploit­ed in the de­vel­op­ment of AI. AI is used to mon­i­tor, track, and pre­dict de­mand and ca­pac­i­ty. It is used to make hir­ing de­ci­sions. AI also re­lies on labour-in­ten­sive, crowd­sourced, man­u­al la­belling of data that is need­ed for its train­ing data. Crawford re­views Amazon’s Mechanical Turk and Google’s re­CAPTCHA tech­nolo­gies.

Data and Classification

I have grouped these chap­ters to­geth­er in this re­view be­cause they en­com­pass the gen­er­al fo­cus of cur­rent-day re­search into AI fair­ness and they can’t be con­sid­ered sep­a­rate from each oth­er. These two chap­ters are a high­light of the book for me. One of my only crit­i­cisms of these chap­ters is how much of each might pos­si­bly have been also dis­cussed in the oth­er chap­ter.

The chap­ter on data is the most co­her­ent in the en­tire book, with a great col­lec­tion of on-point ex­am­ples. It does slight­ly un­der­play the cur­rent aware­ness of dataset is­sues, though.See e.g. the NeurIPS 2021 Paper Checklist which asks authors to report on the copyright, licence, and consent relating to the creation of the dataset(s) that are used. The data chap­ter has a good dis­cus­sion of pri­va­cy and own­er­ship is­sues and the prac­tice of us­ing pub­lic data to cre­ate pri­vate mod­els.

The classification chap­ter high­lights the cir­cu­lar­i­ty and cy­cle of feed­back be­tween data gen­er­a­tion, train­ing, classification, and back to data gen­er­a­tion. Classification risks repli­cat­ing in the fu­ture the bi­as­es ex­hib­it­ed in the past and this is tied up both with the data used to train an AI and the sys­tems sur­round­ing an AI.

Making these choic­es about which in­for­ma­tion feeds AI sys­tems to pro­duce new classifications is a pow­er­ful mo­ment of de­ci­sion mak­ing: but who gets to choose and on what ba­sis? The prob­lem for com­put­er sci­ence is that jus­tice in AI sys­tems will nev­er be some­thing that can be cod­ed or com­put­ed. It re­quires a shift to as­sess­ing sys­tems be­yond op­ti­miza­tion met­rics and sta­tis­ti­cal par­i­ty and an un­der­stand­ing of where the frame­works of math­e­mat­ics and en­gi­neer­ing are caus­ing the prob­lems. This also means un­der­stand­ing how AI sys­tems in­ter­act with data, work­ers, the en­vi­ron­ment, and the in­di­vid­u­als whose lives will be affected by its use and de­cid­ing where AI should not be used.Crawford, supra note 1 at 147–78.

Affect

This chap­ter is a fo­cused cri­tique on a par­tic­u­lar ap­pli­ca­tion of AI: affect (rough­ly, emo­tion) recog­ni­tion. It echoes Crawford’s cri­tique from AI Now’s 2019 Report and else­where. It points out the du­bi­ous foun­da­tions un­der­ly­ing these at­tempts at clas­si­fy­ing peo­ple’s emo­tion­al state based on their fa­cial ex­pres­sions. This is an­oth­er good chap­ter in terms of co­her­ence and writ­ing style. I would not be sur­prised if it had ex­ist­ed as a stand­alone piece be­fore be­ing used in this book.

State

While state use of AI was dis­cussed tan­gen­tial­ly through­out the book (COMPAS, law en­force­ment gen­er­al­ly, bor­der cross­ings, mug shot analy­sis), this chap­ter is os­ten­si­bly fo­cused on state in­tel­li­gence, sur­veil­lance, and de­ci­sion-mak­ing. However, it un­avoid­ably weaves to­geth­er a sto­ry that in­cludes pri­vate re­search (e.g. Google, Microsoft) and de­ci­sion-mak­ing span­ning NSA in­tel­li­gence gath­er­ing to apps like “Neighbors, Citizen, and Nextdoor.” Crawford dis­cuss­es the Snowden doc­u­ments, the goal of mil­i­taries around the world to gain the AI edge, and en­list­ment of pri­vate cor­po­ra­tions in these efforts. The en­am­our­ment with AI sur­veil­lance and de­ci­sion-mak­ing has also worked its way down to lo­cal po­lice forces, mu­nic­i­pal­i­ties, and ad­min­is­tra­tive bod­ies. Crawford presents a com­mon­ly ref­er­enced ex­am­ple of gov­ern­ment de­ci­sion-mak­ing gone wrong: a sys­tem used in Michigan to at­tempt iden­ti­fy un­em­ploy­ment-in­sur­ance fraud. It “in­ac­cu­rate­ly identified more than forty thou­sand Michigan res­i­dents of sus­pect­ed fraud.”Ibid at 206.

Power

This chap­ter is the book’s con­clu­sion but it presents a new idea, too: en­chant­ed de­ter­min­ism. This is the para­dox­i­cal idea that AI sys­tems are both “en­chant­ed” in that they can do what we haven’t told them to do, but also de­ter­min­is­tic enough to be safe and eth­i­cal for de­ploy­ment in high-stakes de­ci­sion-mak­ing. I have also ar­gued that you can’t have it both ways, es­pe­cial­ly when we ex­pect a de­ci­sion-mak­er to act with dis­cre­tion. An al­go­rithm will ei­ther be fol­low­ing de­ter­min­is­tic in­struc­tions, which is a fet­ter­ing of dis­cre­tion, or it will in­cor­po­rate ran­dom­ness, and that also isn’t what we ex­pect from dis­cre­tion.

I like Crawford’s re­minder that we must at­tend to the com­mitt­ments we are mak­ing as we de­vel­op and de­ploy AI sys­tem in pri­vate and through the state.

Critiques

“AI is nei­ther artificial nor in­tel­li­gent.”Ibid at 8. This is a nice pithy point, but I don’t think it is one that any­one dis­agrees with, at least with the sense that Crawford sets up through con­trast in the fol­low­ing sen­tence.The quote goes on: “Rather, artificial intelligence is both embodied and material, made from natural resources, fuel, human labour...” While some peo­ple have op­ti­mism that we might some­day cre­ate real in­tel­li­gence, many peo­ple doubt that this is pos­si­ble, and no­body to­day sug­gests that we have done so. This is why it is called artificial in­tel­li­gence. The term artificial is used not to im­ply that there are no real ar­ti­facts or effects of AI, but mere­ly to in­di­cate that they are not tru­ly in­tel­li­gent. They are artifices of in­tel­li­gence. They are also “artificial” in an­oth­er sense: that they are cre­at­ed by us, rather than oc­cur­ing nat­u­ral­ly. I think this is whol­ly con­sis­tent with the point that Crawford is try­ing to make with this apho­rism but with­out the need for Crawford’s lin­guis­tic at­tack on the term it­self.

In the Earth chap­ter, Crawford cites the work of Emma Strubell et al. for the propo­si­tion that “run­ning only a sin­gle NLP mod­el pro­duced more than 660,000 pounds of car­bon diox­ide emis­sions.”Crawford, supra note 1 at 42, citing Emma Strubell, Anaya Ganesh & Andrew McCallum, “Energy and Policy Considerations for Deep Learning in NLP” (2019) arX­iv.org. Strubell et al. were re­port­ing the cost to train such a net­work. It is now well-rec­og­nized that train­ing large neur­al net­works has a large cost to our en­vi­ron­ment, but also that the cost of run­ning them at in­fer­ence time is not near­ly cost­ly.See e.g. Tom B Brown et al, “Language Models are Few-Shot Learners” (2020) arX­iv.org, s 6.3 (Energy us­age).

I found the con­tin­ued ref­er­ence to log­ics to be un­help­ful­ly vague. Crawford of­ten does not elab­o­rate on what a par­tic­u­lar log­ic ac­tu­al­ly en­tails.E.g. “the logics of capital, policing, and militarization”; “the logics of extraction, a constant drawdown of minerals, water and fossil fuels”; “Logics of efficiency, surveillance, and automation”; “the response is to accept regulation at the margins but to leave untouched the underlying logics of production”; “new logics of information at a planetary level”; “cross-breeding corporate and state logics”; “Silicon Valley had already built and monetized the logics and infrastructures of AI.” It seems that log­ics is used vari­ably as a syn­onym for “meth­ods,” “effects,” or “val­ues.” But the in­tend­ed mean­ing isn’t clear and seems to shift de­pend­ing on the us­age.

Last, it seems like the ed­i­tor may have de­mand­ed ad­di­tion­al con­tent to fill out some of the chap­ters, es­pe­cial­ly the chap­ter on labour. The Google TrueTime ex­am­ple seems shoe­horned into this chap­ter and it is not a good fit. Google TrueTime is a dis­trib­uted, con­sis­tent clock that is avail­able on Google servers. This is pre­sent­ed in a chap­ter de­vot­ed to high­light­ing the ways in which AI re­lies on and en­ables ex­ploita­tive labour prac­tices. It is hard for me to see the ar­gu­ment that TrueTime does this. TrueTime is used to con­sis­tent­ly cre­ate time­stamps for events that oc­cur across mul­ti­ple servers and dat­a­cen­ters, sep­a­rat­ed by thou­sands of kilo­me­tres. This kind of con­sis­ten­cy is crit­i­cal to pro­grams that deal with any sort of ac­count­ing of re­sources: e.g. bank ac­counts, sell­ing con­cert tick­ets. Many peo­ple might be at­tempt­ing to in­ter­act with such a sys­tem si­mul­ta­ne­ous­ly through sep­a­rate servers and there has to be a way for those sep­a­rate servers to de­ter­mine which in­ter­ac­tion hap­pened first. Otherwise, two peo­ple might be sold the same final tick­et to a con­cert. Crawford fails to con­nect this tech­no­log­i­cal ex­am­ple to the the­sis de­vel­oped in this chap­ter: that AI serves the goals of con­trol­ling “bod­ies in space and time”Crawford, supra note 1 at 16. and of “con­trol­ling the po­lit­i­cal or­der.”Ibid at 82. Other ex­am­ples that Crawford pro­vides are com­pelling; this one is not.

Again, these are mi­nor points that can be skimmed past or set aside while you read this work. The big­ger pic­ture is worth see­ing.

Conclusion

For a gen­er­al au­di­ence un­versed on the full scope of the use and de­vel­op­ment of AI, this is a help­ful book. Much of the ma­te­r­i­al trans­fers to con­cerns with oth­er large-scale tech­nol­o­gy: bat­ter­ies, blockchain, dig­i­tal cur­ren­cy, NFTs. And AI as a tool of state pow­er will need pub­lic reg­u­la­tion.