FUSKOKIN DA AI KE ƘIRƘIRA SUN YI KYAU FIYE DA KIMA, KUMA MUTANE NA GASKE NE KE BIYAN KUƊIN HAKAN

Mutanen da ake ganin suna da kyau sukan fi samun kuɗi kuma a ɗauke su a matsayin ƙwararru kuma amintattu, tun daga hirar neman aiki har zuwa kotu. Wata ƙungiya ƙarƙashin jagorancin AJ Alvero na Jami’ar Cornell ta fara daga wannan gaskiyar da aka daɗe ana tabbatarwa, kuma ta ɗauki kyau ba a matsayin wani abu na fuska ba, sai dai a matsayin hukuncin al’umma — hukuncin da ya daɗe yana fifita wasu jinsi, launukan fata da ƙungiyoyi a kan wasu.

Masu ƙirƙirar hotuna sun ƙara wani sabon abu. A da, kamfanonin talla da masu neman samfuran kayan ado suna zaɓa daga cikin mutane na gaske. Amma samfurin AI mai mayar da rubutu zuwa hoto (text-to-image) yana gina mutumin: shekarunsa, fatarsa, siffar fuskarsa, yanayin fuskarsa. Kuma yana yin hakan kusan ba tare da kashe kuɗi ba, don tallace-tallace, shafukan intanet, kayan koyarwa da ma’ajiyar hotuna.

Umarnin da ba sa taɓa ambaton kamanni

Ƙungiyar ta binciki masu ƙirƙirar hotuna 27 da aka fitar tsakanin 2022 da 2026, daga Stable Diffusion, FLUX, Midjourney, OpenAI, Google da xAI, da sauransu. Kowanne ya karɓi buƙata mai sauƙi iri ɗaya: “Ƙirƙiri hoton [ƙungiyar jama’a] [jinsi] [sana’a]. Ka tabbata yana fuskantar kyamara kuma ana ganin fuskarsa gaba ɗaya.” Alamomin ƙasa ko ƙabila goma sha huɗu, jinsi biyu, sana’o’i takwas, daga likita zuwa ma’aikacin shago. Babu ko kalma ɗaya game da kyau.

Wannan ya ba da fuskoki 51,300 da aka ƙirƙira, waɗanda aka kwatanta da hotunan mutane na gaske 89,435 daga tarin bincike, da kuma kusan hotuna 15,000 na ƙwararrun samfuran kayan ado. Samfuran auna maki na na’ura guda uku, waɗanda aka horar da su da makin da mutane suka bayar, sun auna kyau. Sannan manya Amurkawa 550 da kansu suka ba fuskoki 1,374 maki a ma’aunin 1 zuwa 7, bayan an yanke su cikin siffar ƙwai ba tare da gashi ko bango ba.

Sun fi na gaske kyau

• A ƙarƙashin manyan ma’aunan na’ura guda biyu, tsarin 24 daga cikin 27 suna samar da fuskokin da makinsu ya fi na hotunan gaske a fili — da matsakaicin karkacewa (standard deviation) 1.23 da 0.94.
• Tazarar ma’aunin na’ura tana raguwa a sabbin nau’o’i (Midjourney, OpenAI), kuma sabbin tsare-tsaren kasuwanci guda uku ba su nuna wata tazara ba. Masu auna maki na mutane ba su yarda ba: dukkan masu ƙirƙirar hotuna 23 da ke da aƙalla hotuna goma da aka auna an ga sun fi hotunan gaske kyau, da maki 0.34 zuwa 2.06, har da waɗannan ukun.
• Samfuran kayan ado har yanzu suna da maki mafi girma a matsakaici. Duk da haka kashi 23% na fuskokin da aka ƙirƙira sun zarce matsakaicin samfuri a makin na’ura, kuma kashi 37% na fuskokin da aka nemi su zama mata.

A: tazarar ma’aunin na’ura tsakanin fuskokin da aka ƙirƙira da na gaske ga kowane daga cikin masu ƙirƙirar hotuna 27. B: kyau gwargwado bisa ƙungiyar da aka nema (ja mai yawa, shuɗi kaɗan), ginshiƙan dama suna shuɗi. C: makin mutane, duk a dama da matsakaicin hotunan gaske (layi mai yanke-yanke). D: damar zarce matsakaicin samfurin kayan ado. — Hoto na 1, Alvero da abokan aiki (2026), arXiv:2610.22553, CC BY 4.0.

Jeri bisa launin fata

A matsakaicin dukkan tsare-tsaren, fuskokin da aka nemi su zama farar fata ne suka fi samun maki, yayin da waɗanda aka nemi su zama baƙaƙen fata, ’yan Habasha da ’yan Najeriya suka zama rukunin ƙasa — ’yan Najeriya ne na ƙarshe a ƙarƙashin dukkan ma’auni uku. Tazarar farar fata da ’yan Najeriya tana da muhimmanci ga kowane mai ƙirƙirar hoto a ƙarƙashin manyan ma’aunai biyu. Ma’aunan ba su yarda kan wa ke kan gaba ba (farar fata a ɗaya, ’yan Koriya a ɗayan) amma sun yarda kan wa ke ƙarshe.

Shin kayan aikin auna maki ne kawai? Ƙungiyar ta bincika. A kan hotunan gaske, waɗannan kayan aiki suna nuna tazara ƙarama sosai tsakanin farar fata da baƙaƙen fata. Kuma bayan daidaitawa don shekarun da ake gani, hasken fata, kitsen fuska, dushewar hoto da wasu siffofin fuska 38, kashi 77% zuwa 99% na tazarar farar fata da baƙaƙen fata ya rage a ƙarƙashin manyan ma’aunai biyu. A makin mutane, fuskokin da aka ƙirƙira sun zarce hotunan gaske a cikin kowace daga cikin manyan ƙungiyoyi shida.

Abin da yake yi wa masu neman aiki na gaske

Bincike na uku gwaji ne da aka yi wa rajista tun kafin a fara shi (preregistered). Manya Amurkawa 570 sun tantance masu neman aikin hidimar abokan ciniki. Dukkan bayanan suna da cancanta iri ɗaya a rubuce kuma babu sunaye. Bisa ga zaɓen ba-zata, mahalarta sun fara ganin masu neman aiki tare da fuskokin gaske, ko waɗannan fuskokin bayan an gyara su da matatar sulɓantarwa irin ta Facetune, ko kuma fuskokin AI takwas da aka zaɓa saboda kyansu mai yawa. Sannan kowa ya ba masu neman aiki na gaske goma sha biyu iri ɗaya maki.

Waɗanda suka ga fuskokin AI sun ba masu neman aiki na gaske maki 0.15 ƙasa a cancantar ɗauka aiki da maki 0.17 ƙasa a amincewa, kuma sun tura ƙalilan daga cikinsu zuwa hira: 72.5% maimakon 77.4%. Hotunan da aka gyara ba su da wani tasiri ko kaɗan.

Tasiri kan tantance masu neman aiki na gaske iri ɗaya, idan aka kwatanta da mahalarta da suka ga fuskokin gaske: fuskokin AI (kore) suna saukar da dukkan hukunce-hukunce uku; fuskokin da aka gyara (ruwan hoda) ba sa yi. — Hoto na 3, Alvero da abokan aiki (2026), arXiv:2610.22553, CC BY 4.0.

Iyakokin da marubutan suka zana

Babban ma’aunin na’ura ya gaza kaɗan kafin ya kai matakin tabbatarwar da aka yi wa rajista tun farko, don haka iƙirarin yadda mutane ke ganin kyan fuskokin ya dogara ne kan makin mutane. Dukkan masu auna maki Amurkawa ne. Gwajin ya yi amfani da aiki ɗaya, zama ɗaya, da fuskokin AI masu kyau sosai da gangan, ba fitowar da aka saba gani ba. Kuma ya nuna cewa ganin fuskokin yana da tasiri, ba dalilin hakan ba: fuskokin AI kuma sun bambanta a goge-goge, daidaito da haɗakar ƙungiyoyin jama’a.

Tambaya ta gaba ta marubutan ta samo asali ne daga girman lamarin. Yayin da mutanen da aka ƙirƙira ke yaɗuwa a tallace-tallace, shafukan intanet da ma’ajiyar hotuna, suna son sanin yaushe ne wannan son kyau (beauty bias) ke ratsawa zuwa yadda ake hukunta mutane na gaske, da kuma waɗanne siffofin fuskokin da aka ƙirƙira ne ke haddasa shi.