conspiracy theory

社媒推送因人而异，驱动政治两极分化

作者：罗伯特·艾略特·史密斯

来源：《英语世界》2020年第09期

Theelectionasoniswindingup，andmysocialmediaisonceagainawashwithpolitical

nesstream：“WarrenandBernie’sawkwardtruce...”，“Trumpeshisba

growing...”and“TheFed’srealmessage...”.ThisistheAmericaIetoday.

Thetroubleis，it’personally-curated

antlyshiftingmirage，evolvinginreal-time，dependingonmylikesand

dislikes，whatIclickon，andwhatIshare.

ArecentPewRearchCenterstudyfoundblacksocialmediaursaremorelikelytoerace-

llerreportsuggestsRussianeffortsagainstHillaryClintontargetedBernie

ber2016，BradParscale，thenPresidentTrump’s2016digital

director，toldBloombergNewsthathetargetedFacebookandmediapostsatpossibleClinton

supporterssothattheywouldsittheelectionout.

Parscale―who，asofearlyAugust，hasspentmore（$9.2million）onFacebookadsfor

Trump2020thanthefourtopDemocraticcandidatescombined―saidthatin2016hetypicallyran

50，000advariationachday，micro-targetingdifferentgmentsoftheelectorate.

Algorithmsareprejudiced

Whilepoliticaloperativexploitingyellowjournalismisnothingnew，thecouplingoftheir

manipulatthmsarenow

themostpowerfulcuratorsofinformation，whoactionnablesuchmanipulationbycreatingour

fracturedinformationalmultiver.

ysoundextreme，butletmeexplain.

InanalysconductedbymylfandcolleaguesatUniversityCollegeLondon（UCL），we

modeledthebehaviorofsocialnetworks，usingbinarysignals（1sand0s）pasdbetween

simplified“agents”thatreprentedpeoplesharingofopinionsaboutadivisiveissue（saypro-life

versuspro-choiceorthemeritsofbuildingawallornot）.

Most“agents”inthismodeldeterminethesignalstheybroadcastbadonthesignalsthey

receivefromthosurroundingthem（aswedosharingnewsandstoriesonline）.Butweaddedina

smallnumberofagentswecalled“motivatedreasoners，”who，regardlessofwhattheyhear，only

broadcasttheirownpre-determinedopinion.

Ourresultsshowedthatineveryca，motivatedreasonerscametodominatethe

conversation，drivingallotheragentstofixedopinions，ggests

that“echochambers”areaninevitableconquenceofsocialnetworksthatincludemotivated

reasoners.

Itgoesdeeperthanyouthink：TwoyearsafterCharlottesville1，I’mfightingtheconspiracy

theoryindustrialcomplex.

Sowhoarethemotivatedreasoners？Youmightassumetheyarepoliticalcampaigners，

，inreality，themostmotivated

reasonersonlinearethealgorithmsthatcurateouronlinenews.

Howtechnologygeneralizes

Intheonlinemediaeconomy，theartificialintelligenceinalgorithmsaresingle-mindedin

achievingtheirprofit-drivenagendasbyensuringthemaximumfrequencyofhumaninteractionby

arenotonlyeconomicallysingle-minded，they

arealsostatisticallysimple-minded.

Take，forexample，the2016storyinTheGuardianaboutGooglearchesfor

“unprofessionalhair”returningimagespredominantlyofblackwomen.

Doesthisrevealadeepsocialbiastowardsracismandxism？Toconcludethis，onewould

havetobelievethatpeopleareusingtheterm“unprofessionalhair”inclocorrelationwithimagesof

blackwomentosuchanextentastosuggestmostpeoplefeeltheirhairstylesdefine“unprofessional.”

Regardlessofsocietalbias（whichcertainlyexists），thisemsdoubtful.

Itisn’tallbadnewsfornewspapers：I’majournalismstudentinaneraofclosing

newsrooms，‘fakenews.’ButIstillwantin.

HavingworkedinAIfor30years，Iknowitisprobablymorestatisticallyreliablefor

algorithmstorecognizeblackwomen’shairstylesthanthoofblackmen，whitewomen，

issimplyanaspectofhowalgorithms“e，”byusingoverallfeaturesofcolor，shape，andsize.

Justaswithreal-worldracism，resortingtosimplefeaturesiasierforalgorithmsthanderivingany

ifythiffect.

Tobeprejudicedmeanstopre-judgeonsimplifiedfeatures，andthendrawgeneralizations

wtheypar

theincomprehensible“BigData”neers

likemeexplicitlyprogramgeneralizationasagoalofthealgorithmswedesign.

Giventhesimplifyingfeaturesthatalgorithmsu（gender，race，politicalpersuasion，

religion，age，etc.）andthestatisticalgeneralizationstheydraw，thereal-lifeconquenceis

informationalgregation，notunlikepreviousracialandsocialgregation.

Itgoesdeeperthanyouthink：TwoyearsafterCharlottesville1，I’mfightingtheconspiracy

theoryindustrialcomplex.

Sowhoarethemotivatedreasoners？Youmightassumetheyarepoliticalcampaigners，

，inreality，themostmotivated

reasonersonlinearethealgorithmsthatcurateouronlinenews.

Howtechnologygeneralizes

Intheonlinemediaeconomy，theartificialintelligenceinalgorithmsaresingle-mindedin

achievingtheirprofit-drivenagendasbyensuringthemaximumfrequencyofhumaninteractionby

arenotonlyeconomicallysingle-minded，they

arealsostatisticallysimple-minded.

Take，forexample，the2016storyinTheGuardianaboutGooglearchesfor

“unprofessionalhair”returningimagespredominantlyofblackwomen.

Doesthisrevealadeepsocialbiastowardsracismandxism？Toconcludethis，onewould

havetobelievethatpeopleareusingtheterm“unprofessionalhair”inclocorrelationwithimagesof

blackwomentosuchanextentastosuggestmostpeoplefeeltheirhairstylesdefine“unprofessional.”

Regardlessofsocietalbias（whichcertainlyexists），thisemsdoubtful.

Itisn’tallbadnewsfornewspapers：I’majournalismstudentinaneraofclosing

newsrooms，‘fakenews.’ButIstillwantin.

HavingworkedinAIfor30years，Iknowitisprobablymorestatisticallyreliablefor

algorithmstorecognizeblackwomen’shairstylesthanthoofblackmen，whitewomen，

issimplyanaspectofhowalgorithms“e，”byusingoverallfeaturesofcolor，shape，andsize.

Justaswithreal-worldracism，resortingtosimplefeaturesiasierforalgorithmsthanderivingany

ifythiffect.

Tobeprejudicedmeanstopre-judgeonsimplifiedfeatures，andthendrawgeneralizations

wtheypar

theincomprehensible“BigData”neers

likemeexplicitlyprogramgeneralizationasagoalofthealgorithmswedesign.

Giventhesimplifyingfeaturesthatalgorithmsu（gender，race，politicalpersuasion，

religion，age，etc.）andthestatisticalgeneralizationstheydraw，thereal-lifeconquenceis

informationalgregation，notunlikepreviousracialandsocialgregation.

Itgoesdeeperthanyouthink：TwoyearsafterCharlottesville1，I’mfightingtheconspiracy

theoryindustrialcomplex.

Sowhoarethemotivatedreasoners？Youmightassumetheyarepoliticalcampaigners，

，inreality，themostmotivated

reasonersonlinearethealgorithmsthatcurateouronlinenews.

Howtechnologygeneralizes

Intheonlinemediaeconomy，theartificialintelligenceinalgorithmsaresingle-mindedin

achievingtheirprofit-drivenagendasbyensuringthemaximumfrequencyofhumaninteractionby

arenotonlyeconomicallysingle-minded，they

arealsostatisticallysimple-minded.

Take，forexample，the2016storyinTheGuardianaboutGooglearchesfor

“unprofessionalhair”returningimagespredominantlyofblackwomen.

Doesthisrevealadeepsocialbiastowardsracismandxism？Toconcludethis，onewould

havetobelievethatpeopleareusingtheterm“unprofessionalhair”inclocorrelationwithimagesof

blackwomentosuchanextentastosuggestmostpeoplefeeltheirhairstylesdefine“unprofessional.”

Regardlessofsocietalbias（whichcertainlyexists），thisemsdoubtful.

Itisn’tallbadnewsfornewspapers：I’majournalismstudentinaneraofclosing

newsrooms，‘fakenews.’ButIstillwantin.

HavingworkedinAIfor30years，Iknowitisprobablymorestatisticallyreliablefor

algorithmstorecognizeblackwomen’shairstylesthanthoofblackmen，whitewomen，

issimplyanaspectofhowalgorithms“e，”byusingoverallfeaturesofcolor，shape，andsize.

Justaswithreal-worldracism，resortingtosimplefeaturesiasierforalgorithmsthanderivingany

ifythiffect.

Tobeprejudicedmeanstopre-judgeonsimplifiedfeatures，andthendrawgeneralizations

wtheypar

theincomprehensible“BigData”neers

likemeexplicitlyprogramgeneralizationasagoalofthealgorithmswedesign.

Giventhesimplifyingfeaturesthatalgorithmsu（gender，race，politicalpersuasion，

religion，age，etc.）andthestatisticalgeneralizationstheydraw，thereal-lifeconquenceis

informationalgregation，notunlikepreviousracialandsocialgregation.