BAYANAI · KWAMITOCIN SAMFURAN AI
Ku gaya wa samfurin AI ko wane ne shi. Hukunce-hukuncensa na dalili ba sa sauyawa sosai.
Oktoba 2026
Ɗakunan samfuran AI na harshe a True Causes suna ba kowane samfuri matsayin aro: ma’aikaciyar jinya mai haya, mai gida mai ritaya, ƙaramin mai gina gidaje. Ra’ayin shi ne cewa mutane daban-daban suna ganin dalilai daban-daban, kuma kwamitin samfuran da ke taka rawarsu zai gani haka. Ko samfurin da aka gaya masa ko wane ne yana amsawa kamar wannan mutumin tambaya ce da aka sani (Argyle et al., 2023); mun gwada ta a kan hukunce-hukuncen dalili.
Ba ya yi, wajen tantancewa. A cikin samfura 21 da matsayin aro 14, matsayin da aka ba samfuri bai canza alaƙoƙin dalili da ya karɓa da wani abu da za mu iya aunawa ba. Wane samfuri ne kuwa ya canza abubuwa sosai.
Nazarin
Mun ɗauki alaƙoƙi 120 daga ɗakin bayanan jama’a kan arahar gidaje, an zaɓa su ta yadda alaƙoƙin da ɗakin ya tabbatar, ya ƙi da waɗanda bai iya yankewa ba suke daidai wa daida. Kowane cikin samfura 21 ya amsa kowace alaƙa, sau ɗaya a matsayin kansa da sau huɗu ko biyar a matsayin wani matsayin aro daban: matsayin aro goma sha biyu na gidaje (masu haya da masu gida, mai bayar da haya, mai tsara birni, mai gina gidaje, mai shirya masu haya, ma’aikacin gini da sauransu), da tagwaye biyu, mutum ɗaya sai dai a abu ɗaya, haya maimakon mallaka. Samfura bakwai ne suka taka kowane matsayin aro. Wannan kujeru 119 ne da amsoshi 14,220, kowace kujera tana amsa kowace alaƙa, don haka babu amsar da ta dogara ga wata.
Abin da ya yanke amsoshin
Alaƙar kanta tana bayyana kusan rabin dukkan bambancin (52%): samfura sun fi yarda kan waɗanne alaƙoƙi ne masu yiwuwa. Karatun samfurin kansa na wasu alaƙoƙi musamman yana bayyana wani kashi ɗaya cikin huɗu (24%), ko wane matsayin aro yake takawa. Ɗabi’ar samfuri ta gama gari ta cewa e tana ƙara 3%.
Matsayin aro bai ƙara komai ba. Kasonsa na bambancin waɗanne alaƙoƙi suka samu e ya kasance 0.000; kasonsa na karkata gaba ɗaya zuwa e, 0.002. Gwajin da ke ƙoƙarin gane matsayin aro daga amsoshin kujera, ta samfuran da ba a gan shi a cikinsu ba, bai fi sa’a kyau ba (6.1% idan aka kwatanta da 7.1%). Kuma kujeru ba su motsa zuwa alƙiblar da siffofin matsayin aronsu za su yi hasashe ba: dangantakar ta kasance sifili. Lokacin da matsayin aro ya canza amsa, a tsakanin kashi 1% da 15% na alaƙoƙi gwargwadon samfurin, ya canza ta ne ba tare da tsari ba.
Wannan ya dace da abin da wasu suka gano. Matsayin aro yana bayyana ɗan kaɗan na bambanci a ayyukan yi wa bayanai alama (Hu & Collier, 2024), kuma rawar da aka ambata a umarnin tsari ba ta canza amsoshin samfuri a kai a kai (Zheng et al., 2024). Tambayar dalili tana karantuwa ga samfuri a matsayin tambaya game da duniya, ba game da mutumin da ke tambaya ba.
Abin da matsayin aro ya canza: wane ne ke amsawa
Matsayin aro ya sa samfura su ce “ba za a iya faɗa ba” sau da yawa, kashi 26% na lokaci idan aka kwatanta da 19% a matsayin kansu, kuma ba daidai wa daida ba. Matasa, masu ƙaramin samu, masu haya, masu ra’ayin hagu da wanda ba na Amurka ba sun ƙaurace fiye, har maki 14; tsofaffi, masu ra’ayin dama, masu bayar da haya, masu babban samu da masu gida sun ƙaurace ƙasa, har maki 13. Samfuran sun taka matsayin aro na ƙasan matsayi kamar ba su da tabbaci kan abin da suka sani. An san umarnin matsayin aro yana ɗauke da irin waɗannan zato zuwa tunani (Gupta et al., 2024). Wasu samfura sun ƙi amsawa sosai idan suna da matsayin aro: Gemini 3.5 Flash ya ce “ba za a iya faɗa ba” ga kashi 59% na alaƙoƙi a matsayin aro kuma ba ko ɗaya a matsayin kansa.
Waɗanne samfura ne ke tantancewa da kyau
Jadawalai suna sanya kowane samfuri bisa sau nawa yake cewa e (a kwance) da yawan bayanin da amsoshinsa ke ɗauke da shi (a tsaye). Ana ƙiyasta yawan bayani ba tare da sanin amsoshin da suka dace ba, daga yadda samfura ke yarda da juna (Dawid & Skene, 1979): 1 mai tantancewa ne mai daidaito gaba ɗaya, 0 amsa ce da ba ta ɗauke da wata alama. Yana auna yarjejeniya da abin da kwamitin gaba ɗaya ya gano, ba daidaito ba, don haka samfuri mai ƙarfi da ya saɓa wa marasa ƙarfi zai iya samun ƙaramin maki fiye da yadda ya cancanta.
Yawancin samfura suna bambancewa: suna cewa e ga tsakanin kashi ɗaya cikin huɗu da kashi biyu cikin uku na alaƙoƙi, kuma amsoshinsu suna ɗauke da alama. Claude Sonnet 5.5 da DeepSeek V4 Flash sun fi yarda da kwamitin. Ƙananan samfura kamar Gemma 3 4B da Command R7B suna ɗauke da ƙasa. Schematron, samfurin da aka gina don fitar da bayanai maimakon tantance iƙirari, ya ce “ba za a iya faɗa ba” ga kusan komai, kuma abin da ya amsa kusan ba ya ɗauke da wata alama.
Shin samfura suna da nasu ra’ayoyi?
Idan matsayin aro ba ya tsara amsoshin, wataƙila samfuran kansu suna riƙe da ra’ayoyi daban-daban game da matsalar, ɗaya yana zargin dokoki, wani kuma hasashen kasuwa. Kashi ɗaya cikin huɗu na bambancin samfura ne ke karanta alaƙa ɗaya daban, don haka mun nemi alƙiblolin wannan bambanci: samfurin lissafi da ke sanya kowane samfuri da kowace alaƙa a kan ʼyan fannoni, daga amsoshin samfuran a matsayin kansu. Ana ƙirga fanni ne kawai idan ya tabbata, wato, idan daidaita shi daban a kan rabin alaƙoƙi biyu ya sanya samfuran kusan a wurare iri ɗaya.
Fanni ɗaya ya tabbata, kuma ba ra’ayi ba ne. A ɗaya ƙarshen akwai alaƙoƙin da da ƙyar suke da ma’ana, kamar “gidajen alfarma da masu zuba jari na waje suka saya” yana kai wa ga “Amurka tana ƙara baƙi 600,000 zuwa miliyan 1.2 a shekara”, wanda samfura marasa ƙarfi suka karɓa. A ɗayan ƙarshen akwai tsari ɗaya mai ma’ana, ƙuri’un mazauna da muradun waɗanda ke kan mulki suna takaita samar da gidaje, wanda samfura masu ƙarfi duka suka karɓa. Fannin shi ne yadda samfuri ke tantancewa da kyau: samfura masu ƙarfi suna kusa da juna, kuma kowane samfuri mara ƙarfi yana kauce musu ta hanyarsa.
A cikin samfura goma sha ɗaya masu ƙarfi, da aka daidaita a kan alaƙoƙi 120 na nazarin, babu abin da ya tabbata. Alaƙoƙin sun yi kaɗan. Idan aka daidaita a kan dukkan alaƙoƙi 870 na ɗakin, waɗanda samfura goma sha ɗaya ɗin suka amsa a matsayin kansu, wani bambanci ya bayyana, kuma ya tabbata: idan aka daidaita daban a kan rabin alaƙoƙi biyu, samfuran suna sauka kusan a wurare iri ɗaya (0.89).
A ɗaya ƙarshen, samfura suna karɓar alaƙoƙi daga muradun mutane zuwa shawarwarin siyasa: ribar masu gida a hauhawar farashi tana kai wa ga sauye-sauyen gwamnati na hana samarwa, ƙuri’un mazauna suna kai wa ga Proposition 13. Gemini 3.5 Flash, Claude Sonnet 5.5, Nex da DeepSeek suna a can. A ɗayan ƙarshen, samfura suna karɓar alaƙoƙi daga shinge zuwa tsarin samarwa: dokoki, muradun waɗanda ke kan mulki, majalisun da ke ƙin gidaje masu arha da hakkin mallaka, kowanne yana kai wa ga “ƙarin gidaje suna rage farashi”. Mistral Small, Ministral 8B, Laguna da Qwen suna a can. A taƙaice, wani ɓangare yana bayyana gidaje ta hanyar wanda ke samun riba a siyasa, ɗayan kuma ta hanyar abin da ke toshe samarwa. Wannan karatun namu ne, daga alaƙoƙin da ke kowane ƙarshe. Fanni na biyu yana sanya samfuran GPT-OSS biyu a ɗaya ƙarshe, amma alaƙoƙinsa ba su da jigo ɗaya da za mu iya ambato.
Bambancin ya yarda. A cikin samfura goma sha ɗaya masu ƙarfi, alaƙar kanta tana bayyana kashi 68% nasa, idan aka kwatanta da 52% tare da dukkan samfura 21, kuma karatun da ya keɓanta ga samfuri ya ragu daga 24% zuwa 8%: kusan kashi biyu cikin uku na abin da kamar karatu daban-daban ne ya fito daga samfura marasa ƙarfi. Sauran ƙarami ne, amma shi ne bambancin da ke tafiya a kai a kai da ke sama.
Abin da wannan ke nufi ga True Causes
Kwamitin samfuran AI yana ba da hukunci, tare da ƙunƙuntaccen fagen ra’ayoyi. Matsayin aro ba sa ƙara ra’ayoyi. Samfura masu ƙwarewa suna yarda kan mafi yawan alaƙoƙi, kuma suna bambanta a kai a kai kan sauran, a kan layi tsakanin bayanin siyasa da bayanin samarwa. Wannan fage ne, amma ƙunƙuntacce, wanda yadda aka horar da samfuran ya ƙayyade: ba fagen ra’ayoyin da mutane masu gogewa daban-daban za su kawo ba. Bincike kan samfuran da ke taƙaita shawarwari yana nuna haka: taƙaitawar ba ta wakiltar ra’ayoyin tsiraru yadda ya kamata, a cikin samfura 18 (Zhu et al., 2025), kuma maganganun da aka rubuta don neman yarjejeniya za su iya samun yarda ta salonsu fiye da abin da suke ƙunshe (Tsai et al., 2024). A True Causes samfuri ba ya magana a madadin ɗaki. Yana rubutu ko tantance alaƙoƙi ɗaya-ɗaya, ana ƙirga kowane hukunci, kuma rabuwar ra’ayi tana nan a taswira a matsayin rabuwar ra’ayi.
Zaɓar kwamiti game da inganci ne da farko, sannan daidaito. Bambanci mafi girma tsakanin samfura shi ne yadda suke tantancewa da kyau, don haka kwamiti ya kamata ya kasance na samfura masu ƙwarewa, ba tare da marasa ƙarfi ba. A cikin masu ƙwarewa, iyalai suna karkata ta hanyoyi daban, don haka kwamiti ya kamata ya ɗauko daga da dama, kuma kada wani iyali ɗaya ya fi sauran yawa.
Ya kamata a tantance samfura kafin a ba su kujera. Samfurin da ke cewa e ga komai, a’a ga komai ko “ba za a iya faɗa ba” ga komai yana ƙara surutu ga kowane haɗin da ya tantance. Ɗan gajeren gwaji a kan alaƙoƙin da aka sani yana gano su.
Kada a yi amfani da matsayin aro wajen tantancewa. Suna ƙara ƙauracewa da zato na matsayi, kuma ba sa ƙara bambancin hukunci. Rubutu tambaya ce daban. Ɗakin samfura masu matsayin aro yana rubuta katuna iri-iri, amma samfura daban-daban suna rubutu daban, kuma samfuri ɗaya yana rubutu daban kowane lokaci da aka tambaye shi, don haka katuna iri-iri kaɗai ba sa nuna cewa matsayin aro ya yi wani abu. Gwajin shi ne wanda aka yi amfani da shi a nan don tantancewa: samfura iri ɗaya suna rubuta dalilai da kuma ba tare da matsayin aro ba, kuma ana ƙirga dalilai daban-daban da kowanne ya samar.
Ɗakunan mutane daban ne. Mutum yana kawo abin da ya gani, ba bayanin wanda wataƙila ya gani ba. Wannan nazari ba ya faɗin komai a kan haka; yana cewa samfurin da aka gaya masa cewa mai haya ne bai rayu a matsayin ɗaya ba. Inda manufar ita ce a gano abin da mutane ba su yarda a kai ba, mutane ne kaɗai ke samar da shi. A irin waɗannan ɗakuna samfuri ya fi amfani wajen tambaya fiye da amsawa: tambaya game da kati, ɗaya bayan ɗaya kuma ba tare da nasa ra’ayi ba (me za ku gani idan wannan gaskiya ne? me ke faruwa a tsakani?), tana taimaka wa marubucinsa ya bayyana dalilin ba tare da saka kalmomin samfurin a bakinsa ba.
Iyakoki
Tambaya ɗaya, gidaje; alaƙoƙi 120; matsayin aro huɗu ko biyar ga kowane samfuri, kowanne yana amsawa sau ɗaya. Tasirin da ya fi ƙanƙanta fiye da abin da wannan tsari zai iya gani ba zai bayyana ba. An rubuta matsayin aro a matsayin bayanin rawa; hanyoyi masu ƙarfi na sanya su, kamar roƙon samfuri ya yi tunani a cikin halin kafin ya amsa, za su iya ƙara motsa shi, tare da haɗarin zane-zanen ba’a (Cheng et al., 2023). Yawan bayani yana auna yarjejeniya da kwamitin, ba gaskiya ba, kuma samfuran da ake amfani da su a matsayin masu tantance shawara suna yarda da mutane ƙasa da yadda mutum zai yi fata (Zhu et al., 2025): kwamitin da ke yarda da kansa zai iya yin kuskure a alƙibla ɗaya. Bambancin da ke tsakanin samfura masu ƙwarewa ya bayyana ne kawai tare da dukkan alaƙoƙi 870; da 120 ba a gan shi ba, don haka ƙananan bambance-bambance za su iya kasancewa ba a gani ba, kuma tambayar da aka fi jayayya a kanta fiye da gidaje za ta iya ƙara raba samfura.
Nassoshi
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.