QORAALLO · GUDDIYADA MOODELLADA
Moodel u sheeg cidda uu yahay. Go’aannadiisa sababeed ku dhowaad isma beddelaan.
Oktoobar 2026
Qolalka moodellada luqadda ee True Causes moodel kasta waxay siiyaan shakhsiyad: kalkaaliso kiraysata, mulkiile guri oo hawlgab ah, horumariye yar. Fikraddu waa in dad kala duwan ay arkaan sababo kala duwan, guddi moodello ah oo iyaga ciyaarayana sidaas oo kale. In moodel loo sheegay cidda uu yahay uu u jawaabo sida qofkaas waa su’aal la yaqaan (Argyle et al., 2023); waxaan ku tijaabinay go’aamada sababeed.
Ma jawaabo, go’aaminta. 21 moodel iyo 14 shakhsiyadood, shakhsiyadda moodel la siiyey waxba ma beddelin xiriirrada sababeed ee uu aqbalay, ilaa inta aan cabbiri karnay. Moodelka uu ahaa ayaa wax badan beddelay.
Daraasadda
Waxaan ka qaadannay 120 xiriir qol diiwaan dadweyne oo ku saabsan awoodda guriyeynta, oo loo doortay si xiriirrada qolku xaqiijiyey, diiday iyo go’aamin waayey ay si siman u matalaan. 21-ka moodel mid kastaa wuxuu ka jawaabay xiriir kasta, hal mar isagoo ah naftiisa iyo afar ama shan jeer isagoo ah shakhsiyad kale: laba iyo toban shakhsiyadood oo guriyeyn (kiraystayaal iyo milkiilayaal, mulkiile guri kireeya, qorsheeye, horumariye, abaabule kiraystayaal, shaqaale dhisme iyo kuwo kale), iyo laba mataano ah, isla qofka marka laga reebo hal xaqiiqo, kirayn halkii lahaansho. Shakhsiyad kasta waxaa ciyaaray toddoba moodel. Taasi waa 119 kursi iyo 14,220 jawaabood, kursi kastaa wuxuu ka jawaabay xiriir kasta, sidaas darteed jawaabna kuma tiirsanayn mid kale.
Waxa go’aamiyey jawaabaha
Xiriirka laftiisa ayaa sharraxa qiyaastii badh dhammaan kala duwanaanshaha (52%): moodelladu inta badan way isku raacaan xiriirrada macquulka ah. Akhrinta moodelka u gaarka ah ee xiriirro gaar ah ayaa sharraxda rubuc kale (24%), shakhsiyad kasta oo uu ciyaaro. U janjeedhka guud ee moodel inuu yiraahdo haa wuxuu ku daraa 3%.
Shakhsiyaddu waxba kuma darin. Qaybteeda kala duwanaanshaha xiriirrada helay haa waxay ahayd 0.000; qaybteeda u janjeedhka guud ee haa, 0.002. Tijaabo isku dayda inay shakhsiyadda ka aqoonsato jawaabaha kursi, moodello aan lagu arkin, ma dhaafin nasiib (6.1% marka loo eego 7.1%). Kuraastuna uma dhaqaaqin jihada ay sifooyinka shakhsiyaddoodu saadaalin lahaayeen: xiriirku wuxuu ahaa eber. Marka shakhsiyad ay beddeshay jawaab, inta u dhexeysa 1% iyo 15% xiriirrada iyadoo ku xiran moodelka, si nasiib ah ayay u beddeshay.
Tani waxay la jaan qaadaysaa waxa kuwa kale heleen. Doorsoomayaasha shakhsiyaddu waxay sharxaan wax yar oo kala duwanaanshaha hawlaha calaamadaynta (Hu & Collier, 2024), doorka lagu magacaabay tilmaamaha nidaamkuna si joogto ah uma beddelo jawaabaha moodel (Zheng et al., 2024). Su’aal sababeed moodelka waxay u akhrisantaa su’aal ku saabsan adduunka, ee kuma saabsana qofka weydiinaya.
Waxa shakhsiyaduhu beddeleen: cidda jawaabta
Shakhsiyaduhu waxay moodellada ka dhigeen inay marar badan yiraahdaan “lama oga”, 26% waqtiga marka loo eego 19% iyagoo ah naftooda, mana ahayn si siman. Shakhsiyadaha dhallinyarada ah, dakhliga hooseeya, kiraysta, bidix u janjeedha iyo kuwa aan Maraykan ahayn way ka aamuseen in ka badan, ilaa 14 dhibcood; shakhsiyadaha waayeelka ah, midig u janjeedha, mulkiilaha guriga kireeya, dakhliga sare iyo milkiilayaashu way ka aamuseen in ka yar, ilaa 13 dhibcood. Moodelladu shakhsiyadaha maqaamka hooseeya waxay u ciyaareen kuwo aan ka hubin waxa ay yaqaanaan. Tilmaamaha shakhsiyadda waxaa la og yahay inay fikradahan nooca ah u gudbiyaan sababaynta (Gupta et al., 2024). Moodello qaar ayaa aad uga diiday shakhsiyad: Gemini 3.5 Flash wuxuu yiri “lama oga” 59% xiriirrada isagoo shakhsiyad ah, midna isagoo naftiisa ah.
Moodellada si fiican u go’aamiya
Shaxuhu moodel kasta waxay dhigaan inta jeer ee uu yiraahdo haa (dhinac) iyo inta jawaabahiisu xog ka bixiyaan (kor). Xog-bixinta waxaa lagu qiyaasaa iyadoon la aqoon jawaabaha saxda ah, sida moodelladu isugu raacaan (Dawid & Skene, 1979): 1 waa go’aamiye si buuxda u joogto ah, 0 waa jawaab aan calaamad sidin. Waxay cabbirtaa ku raacidda waxa guddiga oo dhan helo, ee ma aha saxnaanta, sidaas darteed moodel xooggan oo ku khilaafsan kuwa daciifka ah wuxuu heli karaa dhibco ka yar inta uu mudan yahay.
Moodellada badankood way kala saaraan: waxay haa ku yiraahdaan inta u dhexeysa rubuc iyo laba meelood saddex meelood xiriirrada, jawaabahooduna calaamad bay sidaan. Claude Sonnet 5.5 iyo DeepSeek V4 Flash ayaa ugu badan ku raaca guddiga. Moodellada yaryar sida Gemma 3 4B iyo Command R7B calaamad ka yar ayay sidaan. Schematron, moodel loo dhisay soo saaridda xogta halkii uu go’aamin lahaa sheegashooyin, wuxuu ku dhowaad wax walba ku yiri “lama oga”, waxa uu ka jawaabayna ku dhowaad calaamad ma sido.
Moodelladu ma leeyihiin aragtiyo u gaar ah?
Haddii shakhsiyaduhu aysan qaabayn jawaabaha, waxaa laga yaabaa in moodellada laftoodu ay leeyihiin aragtiyo kala duwan oo dhibaatada ah, mid eedeeya nidaaminta, mid kalena mala-awaalka suuqa. Rubuc kala duwanaanshaha wuxuu ahaa moodello si kala duwan u akhriyay isla xiriirka, sidaas darteed waxaan raadinnay jihooyinka kala duwanaanshahaas: moodel dhiga moodel kasta iyo xiriir kasta dhowr cabbir, oo laga soo qaaday jawaabaha moodellada iyagoo naftooda ah. Cabbir wuxuu tirsamaa oo keliya haddii uu xasilloon yahay, taas oo ah, haddii si gooni ah loogu habeeyo labada badh ee xiriirrada uu moodellada dhigo ku dhowaad isla meelaha.
Hal cabbir ayaa xasilloonaa, mana ahayn aragti. Hal dhinac waxaa ku jiray xiriirro aan macne badan lahayn, sida “guryo raaxo oo ay iibsadeen maalgashadayaal shisheeye” oo u horseeda “Maraykanku sannadkii wuxuu ku daraa 600,000 ilaa 1.2 milyan oo muhaajiriin ah”, kuwaas oo moodellada daciifka ahi aqbaleen. Dhinaca kale waxaa ku jiray hal hab oo isku xiran, codadka dadka deegaanka iyo danaha kuwa haysta oo xaddida bixinta guryaha, kaas oo moodellada xooggan dhammaantood aqbaleen. Cabbirku waa sida uu moodel si fiican u go’aamiyo: moodellada xooggan way isu dhow yihiin, moodel kasta oo daciif ahina si u gaar ah ayuu uga leexdaa.
Kow iyo tobanka moodel ee xooggan, oo lagu habeeyey 120-ka xiriir ee daraasadda, waxba ma xasilloonayn. Xiriirradu aad bay u yaraayeen. Marka lagu habeeyo dhammaan 870-ka xiriir ee qolka, oo ay ka jawaabeen isla kow iyo tobanka moodel iyagoo naftooda ah, kala duwanaansho ayaa soo baxay, waana xasilloon yahay: marka si gooni ah loogu habeeyo labada badh ee xiriirrada, moodelladu waxay ku dhacaan ku dhowaad isla meelaha (0.89).
Hal dhinac, moodelladu waxay aqbalaan xiriirro ka socda danaha dadka una socda go’aannada siyaasadeed: danta milkiilayaasha guryaha ee kor u kaca qiimaha oo u horseeda isbeddellada dawladda ee ka soo horjeeda bixinta, codadka dadka deegaanka oo u horseeda Proposition 13. Gemini 3.5 Flash, Claude Sonnet 5.5, Nex iyo DeepSeek ayaa halkaas jooga. Dhinaca kale, moodelladu waxay aqbalaan xiriirro ka socda caqabadaha una socda habka bixinta: nidaaminta, danaha kuwa haysta, golayaasha diida guryo jaban iyo xuquuqda hantida, mid kastaa oo u horseeda “guryo badan ayaa qiimaha dhimaya”. Mistral Small, Ministral 8B, Laguna iyo Qwen ayaa halkaas jooga. Qiyaas ahaan, hal dhinac wuxuu guriyeynta ku sharxaa cidda siyaasad ahaan ka faa’iidaysa, kan kalena waxa xannibaya bixinta. Akhrintaasi waa tayada, oo laga soo qaaday xiriirrada dhinac kasta. Cabbirka labaad wuxuu labada moodel ee GPT-OSS dhigaa hal dhinac, laakiin xiriirradiisu ma wadaagaan mawduuc aan magacaabi karno.
Kala duwanaanshuhu wuu isku raacsan yahay. Kow iyo tobanka moodel ee xooggan, xiriirka laftiisa ayaa sharraxa 68%, marka loo eego 52% oo leh dhammaan 21-ka moodel, akhrinta u gaarka ah ee moodelkuna waxay ka dhacdaa 24% ilaa 8%: qiyaastii laba meelood saddex meelood waxa u ekaa akhrin kala duwan waxay ka yimaadeen moodellada daciifka ah. Inta kale way ka yar tahay, laakiin waa kala duwanaanshaha joogtada ah ee kore.
Waxa tani u macnayso True Causes
Guddi moodello ahi wuxuu bixiyaa go’aan, iyadoo aragtiyuhu cidhiidhi yihiin. Shakhsiyaduhu kuma daraan aragtiyo. Moodellada awoodda leh waxay isku raacaan xiriirrada badankood, si joogto ahna ugu kala duwan yihiin inta kale, khad u dhexeeya sharraxaadaha siyaasadda iyo bixinta. Taasi waa kala duwanaansho, laakiin mid cidhiidhi ah, oo uu dejiyey sida moodellada loo tababaray: ma aha kala duwanaanshaha aragtiyaha ay dad khibrad kala duwan leh keeni lahaayeen. Cilmi-baarista moodellada soo koobaya wada-tashiyada sidaas oo kale ayay tilmaamaysaa: soo koobiddu waxay yareysaa matalaadda mowqifyada laga tirada badan yahay, 18 moodel oo dhan (Zhu et al., 2025), weedhaha loo qoray in lagu helo is-afgaradna waxay ku heli karaan oggolaansho codkooda in ka badan nuxurkooda (Tsai et al., 2024). True Causes, moodel uma hadlo qol. Wuxuu qoraa ama go’aamiyaa xiriirro keli ah, go’aan kastaa waa la tiriyaa, kala qaybsanaantuna khariidadda ayay ku ahaanaysaa kala qaybsanaan.
Doorashada guddi waa tayada marka hore, kadib dheellitirka. Kala duwanaanshaha ugu weyn ee moodellada waa sida ay si fiican u go’aamiyaan, sidaas darteed guddigu waa inuu ahaado moodello awood leh, kuwa daciifka ahna laga reebo. Moodellada awoodda leh, qoysaskoodu dhinacyo kala duwan ayay u janjeedhaan, sidaas darteed guddigu waa inuu ka soo qaato dhowr, qoysna yuusan ka badnaan inta kale.
Moodellada waa in la baaro ka hor inta aan la fadhiisin. Moodel wax walba ku yiraahda haa, wax walba ku yiraahda maya ama wax walba ku yiraahda “lama oga” wuxuu buuq ku daraa lammaane kasta oo uu go’aamiyo. Habayn gaaban oo lagu sameeyo xiriirro la yaqaan ayaa helaysa.
Shakhsiyadaha looma isticmaalo go’aaminta. Waxay ku daraan ka aamusnaan iyo fikrad maqaam, mana keenaan kala duwanaansho go’aan. Qoristu waa su’aal gooni ah. Qol moodello ah oo leh shakhsiyado wuxuu qoraa kaarar kala duwan, laakiin moodello kala duwan si kala duwan ayay wax u qoraan, hal moodelna si kale ayuu wax u qoraa mar kasta oo la weydiiyo, sidaas darteed kaarar kala duwan oo keliya ma muujinayaan in shakhsiyaduhu wax sameeyeen. Tijaabadu waa tan halkan go’aaminta loogu isticmaalay: isla moodellada waxay qoraan sababo shakhsiyad leh iyo mid la’aan, sababaha gaarka ah ee mid kastaa soo saarana waa la tiriyaa.
Qolalka dadku way ka duwan yihiin. Qofku wuxuu keenaa waxa uu arkay, ee ma keeno sharraxaad qof laga yaabo inuu arkay. Daraasaddani waxba kama sheegto taas; waxay sheegaysaa in moodel loo sheegay inuu yahay kiraysta uusan u noolaan sidiisa. Meesha ujeeddadu tahay in la ogaado waxa dadku isku khilaafsan yihiin, dad oo keliya ayaa keena. Qolalkaas moodelku wuxuu waxtar badan yahay marka uu weydiiyo halkii uu ka jawaabi lahaa: su’aal ku saabsan kaar, mid mid oo aragti u gaar ah aan lahayn (maxaad arki lahayd haddii tani run tahay? maxaa dhexda ka dhaca?), ayaa qoraageeda ka caawisa inuu sababta caddeeyo iyadoon ereyada moodelka afkiisa la gelin.
Xaddidaadaha
Hal su’aal, guriyeyn; 120 xiriir; afar ama shan shakhsiyadood moodelkiiba, mid kastaa hal mar ka jawaabay. Saameyno ka yar inta nashqaddani arki karto ma muuqan lahaayeen. Shakhsiyadaha waxaa loo qoray sharraxaad door; habab xooggan oo lagu keeno, sida in moodel laga codsado inuu shakhsiyadda ku sababeeyo ka hor intaanu jawaabin, ayaa laga yaabaa inay sii dhaqaajiyaan, iyo khatarta sawir-qaloocinta (Cheng et al., 2023). Xog-bixintu waxay cabbirtaa ku raacidda guddiga, ee ma aha runta, moodellada loo isticmaalo go’aamiyeyaal wada-tashina dadka kuma raacsana inta la rajayn karo (Zhu et al., 2025): guddi naftiisa ku raacsan weli wuu ku khaldami karaa isla jihada. Kala duwanaanshaha moodellada awoodda leh wuxuu soo baxay oo keliya dhammaan 870-ka xiriir; 120 lama arki karin, sidaas darteed kala duwanaansho ka yar weli waa laga yaabaa in la arki waayo, su’aal ka murran guriyeyntana moodellada way sii kala qaybin kartaa.
Tixraacyada
Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3), 337–351.
Cheng, M., Piccardi, T., & Yang, D. (2023). CoMPosT: Characterizing and evaluating caricature in LLM simulations. Proceedings of EMNLP 2023.
Dawid, A. P., & Skene, A. M. (1979). Maximum likelihood estimation of observer error-rates using the EM algorithm. Journal of the Royal Statistical Society, Series C, 28(1), 20–28.
Gupta, S., Shrivastava, V., Deshpande, A., Kalyan, A., Clark, P., Sabharwal, A., & Khot, T. (2024). Bias runs deep: Implicit reasoning biases in persona-assigned LLMs. Proceedings of ICLR 2024.
Hu, T., & Collier, N. (2024). Quantifying the persona effect in LLM simulations. Proceedings of ACL 2024.
Tsai, L. L., Pentland, A., Braley, A., Chen, N., Enríquez, J. R., & Reuel, A. (2024). Generative AI for pro-democracy platforms. An MIT Exploration of Generative AI, MIT.
Zheng, M., Pei, J., Logeswaran, L., Lee, M., & Jurgens, D. (2024). When “a helpful assistant” is not really helpful: Personas in system prompts do not improve performances of large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024.
Zhu, S., Yang, S., Bakker, M. A., Pentland, A., & Pei, J. (2025). Can AI truly represent your voice in deliberations? A comprehensive study of large-scale opinion aggregation with LLMs. arXiv:2510.05154.