AMANOTHI · AMAPHANELI AMAMODELI
Tshelani imodeli ukuthi ingubani. Izinqumo zayo ngezimbangela azishintshi kakhulu.
Okthoba 2026
Amagumbi amamodeli olimi ku-True Causes anika imodeli ngayinye umlingiswa: unesi oqashile, umnikazi wekhaya osethathe umhlalaphansi, unjiniyela omncane wezindlu. Umqondo ukuthi abantu abahlukene babona izimbangela ezahlukene, futhi iphaneli yamamodeli ebadlalayo nayo ingabona kanjalo. Ukuthi imodeli etshelwe ukuthi ingubani iphendula yini njengalowo muntu kungumbuzo owaziwayo (Argyle et al., 2023); sakuhlola ezinqumweni ngezimbangela.
Ayenzi kanjalo, ekwahluleleni. Kumamodeli angu-21 nabalingiswa abangu-14, umlingiswa owanikwa imodeli awuzange ushintshe izixhumanisi zezimbangela eyazamukela nangalutho esingalukala. Ukuthi bekuyimodeli efe kushintshe kakhulu.
Ucwaningo
Sathatha izixhumanisi ezingu-120 egumbini lerekhodi lomphakathi elimayelana nokuthengeka kwezindlu, ezikhethwe ngendlela yokuthi izixhumanisi igumbi elaziqinisekisa, elazenqaba nalezo elingakwazanga ukuzinquma zimelwe ngokulinganayo. Imodeli ngayinye kwezingu-21 yaphendula sonke isixhumanisi, kanye njengoba injalo bese kane noma kahlanu njengomlingiswa ohlukile: abalingiswa abayishumi nambili bezindlu (abaqashile nabanikazi, umnikazi wendlu eqashwayo, umhleli, unjiniyela wezindlu, umhleli wabaqashi, isisebenzi sokwakha nabanye), nama-wele amabili, umuntu ofanayo ngaphandle kweqiniso elilodwa, eqashile esikhundleni sokuba ngumnikazi. Umlingiswa ngamunye wadlalwa ngamamodeli ayisikhombisa. Lokho kuyizihlalo ezingu-119 nezimpendulo ezingu-14,220, sonke isihlalo siphendula sonke isixhumanisi, ngakho ayikho impendulo eyayincike kwenye.
Lokho okwanquma izimpendulo
Isixhumanisi uqobo sichaza cishe isigamu sakho konke ukwehluka (52%): amamodeli avumelana kakhulu ngokuthi yiziphi izixhumanisi ezinengqondo. Indlela imodeli ngayinye efunda ngayo izixhumanisi ezithile ichaza enye ingxenye yesine (24%), noma ngabe idlala muphi umlingiswa. Ukuthambekela kwemodeli jikelele ukuthi yebo kwengeza u-3%.
Umlingiswa awuzange wengeze lutho. Isabelo sawo ekwehlukeni kokuthi yiziphi izixhumanisi ezithole ukuthi yebo sasingu-0.000; isabelo sawo ekuthambekeleni jikelele ekuthini yebo, u-0.002. Ukuhlola okuzama ukuhlonza umlingiswa ezimpendulweni zesihlalo, ngamamodeli angakaboni kuwo, akuzange kwenze kangcono kunokuqagela (6.1% uma kuqhathaniswa no-7.1%). Futhi izihlalo azizange zinyakazele ohlangothini izici zomlingiswa wazo ezingalubikezela: ukuhlobana kwaba ngu-zero. Lapho umlingiswa ushintsha impendulo, kuphakathi kuka-1% no-15% wezixhumanisi kuye ngemodeli, wayishintsha ngokungahleliwe.
Lokhu kuvumelana nalokho abanye abakutholile. Izici zomlingiswa zichaza okuncane kokwehluka emisebenzini yokumaka (Hu & Collier, 2024), futhi indima eqanjwe kumyalelo wesistimu ayishintshi ngokungaguquki izimpendulo zemodeli (Zheng et al., 2024). Umbuzo wezimbangela ufundeka kumodeli njengombuzo ngomhlaba, hhayi ngomuntu obuzayo.
Lokho abalingiswa abakushintsha: ophendulayo
Abalingiswa benza amamodeli athi “angikwazi ukusho” kaningi, u-26% wezikhathi uma kuqhathaniswa no-19% njengoba enjalo, futhi hhayi ngokulinganayo. Abalingiswa abasebasha, abaholela kancane, abaqashile, abathambekele kwesobunxele nabangewona abaseMelika bagwema ukuphendula kakhulu, kuze kube ngamaphuzu angu-14; abalingiswa abadala, abathambekele kwesokudla, abanikazi bezindlu eziqashwayo, abaholela kakhulu nabanikazi bagwema kancane, kuze kube ngamaphuzu angu-13. Amamodeli adlala abalingiswa abasezingeni eliphansi njengabangaqiniseki kangako ngalokho abakwaziyo. Imiyalelo yabalingiswa yaziwa ngokuthwala ukucwasa okunjalo kuyiyise ekucabangeni (Gupta et al., 2024). Amanye amamodeli enqaba kakhulu uma enomlingiswa: i-Gemini 3.5 Flash yathi “angikwazi ukusho” ku-59% wezixhumanisi njengomlingiswa kodwa ayizange ithi njalo nakwesisodwa njengoba injalo.
Yimaphi amamodeli ahlulela kahle
Amashadi abeka imodeli ngayinye ngokuthi ithi yebo kangaki (ngokuvundlile) nokuthi izimpendulo zayo zinikeza ulwazi olungakanani (ngokuqondile phezulu). Ukunikeza ulwazi kulinganiselwa ngaphandle kokwazi izimpendulo ezilungile, ngendlela amamodeli avumelana ngayo (Dawid & Skene, 1979): u-1 ungumahluleli ongaguquki ngokuphelele, u-0 yimpendulo engathwali uphawu. Kukala ukuvumelana nalokho iphaneli iyonke ekutholayo, hhayi ukulunga, ngakho imodeli eqinile engavumelani nebuthakathaka ingathola amaphuzu aphansi kunalokho efanelwe yikho.
Amamodeli amaningi ayahlukanisa: athi yebo phakathi kwengxenye eyodwa kwezine nezingxenye ezimbili kwezintathu zezixhumanisi, futhi izimpendulo zawo zithwala uphawu. I-Claude Sonnet 5.5 ne-DeepSeek V4 Flash zivumelana kakhulu nephaneli. Amamodeli amancane afana ne-Gemma 3 4B ne-Command R7B athwala okuncane. I-Schematron, imodeli eyakhelwe ukukhipha idatha kunokwahlulela izigomelo, yathi “angikwazi ukusho” cishe kukho konke, futhi lokho eyakuphendula kuthwala uphawu olucishe lungabikho.
Ingabe amamodeli anemibono yawo?
Uma abalingiswa bengabumbi izimpendulo, mhlawumbe amamodeli uqobo anemibono ehlukene ngenkinga, enye isola imithetho, enye ukuqagela emakethe. Ingxenye eyodwa kwezine yokwehluka kwakungamamodeli afunda isixhumanisi esifanayo ngokwehlukile, ngakho safuna izinhlangothi zalokho kwehluka: imodeli ebeka imodeli ngayinye nesixhumanisi ngasinye ezinhlangothini ezimbalwa, kusuka ezimpendulweni zamamodeli njengoba enjalo. Uhlangothi lubalwa kuphela uma luzinzile, okungukuthi, uma ukulufaka ngokwehlukana ezingxenyeni ezimbili zezixhumanisi kubeka amamodeli cishe ezindaweni ezifanayo.
Uhlangothi olulodwa lwaluzinzile, futhi lwalungewona umbono. Ekugcineni okukodwa kwakunezixhumanisi ezicishe zingabi nangqondo, njengokuthi “amafulethi kanokusho athengwe ngabatshalizimali bangaphandle” aholela ekutheni “iMelika yengeza abokufika abayizi-600,000 kuya ezigidini eziyi-1.2 ngonyaka”, amamodeli abuthakathaka azamukela. Ekugcineni okunye kwakunendlela eyodwa enengqondo, amavoti ezakhamuzi nezithakazelo zabasezikhundleni zigcina ukutholakala kwezindlu kunqunyelwe, wonke amamodeli aqinile ayamukela. Uhlangothi luwukuthi imodeli yahlulela kahle kangakanani: amamodeli aqinile ahlala eduze komunye nomunye, futhi imodeli ngayinye ebuthakathaka iyaphambuka kuwo ngendlela yayo.
Phakathi kwamamodeli ayishumi nanye aqinile, efakwe ezixhumanisini ezingu-120 zocwaningo, kwakungekho okuzinzile. Lezo kwakuyizixhumanisi ezimbalwa kakhulu. Uma efakwe kuzo zonke izixhumanisi ezingu-870 zegumbi, eziphendulwe ngamamodeli afanayo ayishumi nanye njengoba enjalo, kwavela umehluko, futhi uzinzile: uma efakwe ngokwehlukana ezingxenyeni ezimbili zezixhumanisi, amamodeli awela cishe ezindaweni ezifanayo (0.89).
Ekugcineni okukodwa, amamodeli amukela izixhumanisi ezisuka ezithakazelweni zabantu ziye ezinqumweni zezombusazwe: inzalo yabanikazi bezindlu ekukhuphukeni kwamanani iholela ezinguqukweni zikahulumeni ezimelene nokwandiswa kwezindlu, amavoti ezakhamuzi aholela ku-Proposition 13. I-Gemini 3.5 Flash, i-Claude Sonnet 5.5, i-Nex ne-DeepSeek zihlala lapho. Ekugcineni okunye, amamodeli amukela izixhumanisi ezisuka ezithiyweni ziye endleleni yokutholakala kwezindlu: imithetho, izithakazelo zabasezikhundleni, imikhandlu eyenqaba amafulethi ashibhile namalungelo empahla, ngakunye kuholela ekutheni “izindlu eziningi zehlisa amanani”. I-Mistral Small, i-Ministral 8B, i-Laguna ne-Qwen zihlala lapho. Ngokusondele, olunye uhlangothi luchaza izindlu ngokuthi ubani ozuzayo kwezombusazwe, olunye ngokuthi yini evimba ukutholakala kwazo. Lokho kufunda kungokwethu, kusuka ezixhumanisini ezisekugcineni ngakunye. Uhlangothi lwesibili lubeka womabili amamodeli e-GPT-OSS ekugcineni okukodwa, kodwa izixhumanisi zalo azabelani ngesihloko esingasiqamba.
Ukwehluka kuyavumelana. Phakathi kwamamodeli ayishumi nanye aqinile, isixhumanisi uqobo sichaza u-68% wakho, uma kuqhathaniswa no-52% nawo wonke amamodeli angu-21, futhi ukufunda okuqondene nemodeli ngayinye kwehla kusuka ku-24% kuya ku-8%: cishe izingxenye ezimbili kwezintathu zalokho okwakubukeka njengokufunda okuhlukene kwavela kumamodeli abuthakathaka. Okusele kuncane, kodwa kungumehluko ongaguquki ongenhla.
Lokho okushiwo yilokhu ku-True Causes
Iphaneli yamamodeli inikeza ukwahlulela, kodwa nemibono emincane ebanzi. Abalingiswa abengezi mibono. Amamodeli anekhono avumelana ezixhumanisini eziningi, futhi ahluka ngokungaguquki kwezinye, emgqeni ophakathi kwezincazelo zezombusazwe nezokutholakala kwezindlu. Lobo bubanzi, kodwa buncane, bunqunywe yindlela amamodeli aqeqeshwa ngayo: hhayi ububanzi bemibono abantu abanesipiliyoni esahlukene ababezobuletha. Ucwaningo ngamamodeli afinyeza izingxoxo lukhomba ngendlela efanayo: izifinyezo azibamele kahle abambalwa, kumamodeli angu-18 (Zhu et al., 2025), futhi izitatimende ezibhalelwe ukuthola ukuvumelana zingathola imvume ngezwi lazo kunangokuqukethwe kwazo (Tsai et al., 2024). Ku-True Causes imodeli ayikhulumeli igumbi. Ibhala noma yahlulele izixhumanisi ezizodwa, sonke isinqumo siyabalwa, futhi ukuhlukana kuhlala ebalazweni njengokuhlukana.
Ukukhetha iphaneli kumayelana nekhwalithi kuqala, bese kuba wukulinganisela. Umehluko omkhulu kunayo yonke phakathi kwamamodeli ukuthi ahlulela kahle kangakanani, ngakho iphaneli kufanele ibe ngamamodeli anekhono, abuthakathaka eshiywe ngaphandle. Phakathi kwamamodeli anekhono, imindeni ithambekela ngezindlela ezahlukene, ngakho iphaneli kufanele ithathe emindenini eminingana, futhi ungabikho umndeni owodwa odlula eminye ngenani.
Amamodeli kufanele ahlolwe ngaphambi kokuthi ahlaliswe. Imodeli ethi yebo kukho konke, cha kukho konke noma “angikwazi ukusho” kukho konke yengeza umsindo kuso sonke isibhangqa eyahlulelayo. Ukulinganisa okufushane ezixhumanisini ezaziwayo kuyawathola.
Abalingiswa akufanele basetshenziselwe ukwahlulela. Bengeza ukugwema ukuphendula nokucwasa ngesikhundla, kodwa abengezi kwehlukahlukana kokwahlulela. Ukubhala kungumbuzo ohlukile. Igumbi lamamodeli anabalingiswa libhala amakhadi ahlukahlukene, kodwa amamodeli ahlukene abhala ngokwehlukile, futhi imodeli eyodwa ibhala ngokwehlukile njalo uma ibuzwa, ngakho amakhadi ahlukahlukene odwa awabonisi ukuthi abalingiswa benza okuthile. Ukuhlola yilokho okusetshenziswe lapha ekwahluleleni: amamodeli afanayo abhala izimbangela enomlingiswa nangaphandle kwawo, bese kubalwa izimbangela ezihlukile ngayinye ezikhiqizayo.
Amagumbi abantu ahlukile. Umuntu uletha lokho akubonile, hhayi incazelo yomuntu ongase akubone. Lolu cwaningo alusho lutho ngokumelene nalokho; lusho ukuthi imodeli etshelwe ukuthi iqashile ayikaze iphile njengomuntu oqashile. Lapho inhloso kungukuthola ukuthi abantu abavumelani ngani, ngabantu kuphela abangakunikeza. Kulawo magumbi imodeli iwusizo kakhulu ekubuzeni kunasekuphenduleni: umbuzo ngekhadi, owodwa ngesikhathi futhi ingenambono wayo (ubungabonani ukube lokhu bekuyiqiniso? kwenzekani phakathi?), usiza umbhali wakho ukuthi enze imbangela icace ngaphandle kokufaka amazwi emodeli emlonyeni wakhe.
Imikhawulo
Umbuzo owodwa, izindlu; izixhumanisi ezingu-120; abalingiswa abane noma abahlanu ngemodeli ngayinye, ngamunye ephendula kanye. Imithelela emincane kunalokho le ndlela yocwaningo engakubona ngeke ibonakale. Abalingiswa babhalwa njengezincazelo zezindima; izindlela eziqinile zokubavusa, njengokucela imodeli ukuthi icabange njengomlingiswa ngaphambi kokuphendula, zingayinyakazisa kakhulu, kanye nengozi yokuhlekisa (Cheng et al., 2023). Ukunikeza ulwazi kukala ukuvumelana nephaneli, hhayi iqiniso, futhi amamodeli asetshenziswa njengabahluleli bezingxoxo avumelana nabantu kancane kunalokho umuntu angakulindela (Zhu et al., 2025): iphaneli evumelana nayo ingaba nephutha ohlangothini olufanayo. Umehluko phakathi kwamamodeli anekhono wavela kuphela nazo zonke izixhumanisi ezingu-870; ngezingu-120 wawungabonakali, ngakho imehluko emincane isengase ingabonakali, futhi umbuzo ophikiswana kakhulu kunezindlu ungahlukanisa amamodeli kakhulu.
Izinkomba
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.