Kwamfuta da AIKafin bugawaNazarin bayanaiKaratun minti 4

AI TA SAUYA WASU KALMOMIN KIMIYYA, AMMA BA KIMIYYAR KANTA BA — TUKUNA

Tun da manyan samfuran harshe (large language models) suka shiga hannun jama’a a ƙarshen 2022, masu bincike suna amfani da su wajen rubutu, gyaran rubutu da kuma bincike a cikin rubuce-rubucen kimiyya. Bincike-bincike suna danganta su da riba ta gaske ga masana kimiyya ɗaiɗai, a fannin yawan aiki da kuma shahara (visibility). Amma damuwa ta ƙaru game da kimiyya baki ɗaya: wani “sabanin AI” (AI paradox) inda yawan aiki ke zuwa a kan kuɗin inganci, ambaliyar “sharar AI” (AI slop), da kuma wani tsarin iri ɗaya (monoculture) da ke shigowa sannu-sannu wanda zai taƙaita fagen ra’ayoyi.

Ron Bekkerman, Nichole Li, Bahareh Kargar da Mark Somers, a New Jersey Institute of Technology, sun tashi don auna ko wani abu ya karye — ba idan aka kwatanta da wani zamani na dā da ake ɗauka cikakke ba, sai dai idan aka kwatanta da yanayin tafiya (trends) da suke gudana tun tuni.

Shekaru goma na cikakkun rubutu

Tarin bayanansu (corpus) ya ƙunshi kusan cikakkun takardu miliyan 2.5 da aka wallafa tsakanin 2015 da 2025, a cikin fannoni biyar: kimiyyar kwamfuta, lissafi da physics (daga arXiv), da kimiyyar lafiya da ta rayuwa (daga PubMed Central). Maimakon duba taƙaitattun bayanai (abstracts) kawai, sun yi nazarin dukkan takardu, bisa wani matakala na tambayoyi: waɗanne kalmomi ake amfani da su, yaya wadata da zurfin fasahar harshen yake, yaya bambancin batutuwan yake, yadda ake gina takardu, da kuma su waye ke rubuta su.

Hanyar iri ɗaya ce a ko’ina: kwatanta kowane yanayin tafiya kafin da bayan 2022, da kuma neman karyewa.

Kalmomin da suka tashi suka faɗi

Farawa daga kalmomin Turanci 20,000 da aka fi amfani da su, waɗanda aka rage zuwa asalin siffofi (base forms) 14,484, ƙungiyar ta sami 796 da suka yi tashin gwauron zabo bayan 2022, 199 da suka ragu da kuma 71 da suka tashi sannan suka faɗi. Masu tashin kalmomi ne na nazari masu goge-goge da samfuran harshe suka fi so, kamar “establish” (tabbatar) da “underscore” (jaddada); masu raguwar kuma kalmomi ne masu sauƙi kamar “good” (mai kyau) da “happen” (faru). Shahararriyar kalmar “delve” (zurfafa) tana nuna salon tashi-da-faɗuwa: kusan babu ita kafin 2023, ta kai ƙololuwa a 2024 kuma ta faɗi sosai a 2025 — marubutan suna ganin hakan ya faru ne yayin da aka sabunta samfuran kuma masu karatu suka fara yin hattara.

Alamar ta fi ƙarfi a kimiyyar kwamfuta kuma ta fi rauni a lissafi, wanda alamomin lissafi na ƙa’ida (formal notation) ke takura rubutunsa. Kuma ba ta taƙaita ga taƙaitattun bayanai ba: kalmomin da suka yi tashin gwauron zabo suna bayyana a cikin jikin takardu ma.

Babu narkewar harshen kimiyya

Duk da haka kalmomin da abin ya shafa — ɗan sama da dubu — ɗigo ne kawai a cikin tarin kalmomi dabam-dabam masu yawan miliyan goma-goma. Bambancin kalmomi (lexical diversity), wanda aka auna a kan tagogi masu zamewa na kalmomi 50 (sliding windows), bai nuna wani sauyi mai muhimmanci na gangara a 2022 a kowane fanni ba. Kuma yawan kalmomin kimiyya na musamman (density of scientific terminology) yana ci gaba da ƙaruwa a dukkan fannoni biyar. A kalaman marubutan, “takardun kimiyya suna ƙara bayyana a matsayin na kimiyya yayin da lokaci ke tafiya”.

Batutuwa: gargaɗi biyu na ƙarya

Shin takardu suna ƙara magana a kan abubuwa iri ɗaya? A physics da lissafi, bambancin batutuwa (topic diversity) bai sauya ba. A kimiyyar kwamfuta, yana faɗuwa kusan a madaidaiciyar layi — amma raguwar ta fara tun kafin 2022, yayin da fannin ya taru a kan zurfin koyo (deep learning) da AI mai samarwa (generative AI), har ma ta ɗan yi jinkiri tun 2021.

Kimiyyar lafiya da ta rayuwa sun nuna raguwa mafi kaifi bayan 2022. Amma yawancin kamanceceniyar da ke tsakanin takardu ta fito ne daga tsarinsu (format): daidaitattun tubala game da lasisi, bayanan bayyanawa (disclosures) da matsayin masu ba da gudummawa (contributor roles), yayin da rubuce-rubucen ilimin likitanci da halittu (biomedical) suka koma zuwa tsari ɗaya bai ɗaya. Tace irin waɗannan daidaitattun jimloli (boilerplate phrases) 105 ya cire yawancin raguwar.

Takardun da aka gina bisa samfuri

Tsarin takardun kimiyyar kwamfuta, wanda aka yi nazari a kan kusan 438,000 daga cikinsu, ya sauya sosai — amma a hankali kuma ba tsayawa. Takardun da ke da sassa ɗaya zuwa uku kawai sun ragu daga 17% zuwa 4%, yayin da waɗanda ke da sassa takwas ko fiye suka ƙaru daga 5% zuwa 30%. Gabatarwa ta ragu daga kalmomi 2,141 zuwa 1,236 a matsakaici, yayin da jerin manazarta suka kusan ninka tsawonsu. Babu ɗaya daga cikin waɗannan yanayin tafiya da ya karye a 2022.

Manyan ƙungiyoyi, ƙarin takardu — galibi kamar dā

Ƙungiyoyin marubuta tare (co-authors) suna ta girma a kowane fanni kafin 2022, kuma sun ci gaba da girma kawai. Akwai keɓantacce guda ɗaya da ya fito fili: ƙungiyoyin kimiyyar kwamfuta daga Asiya sun girma da sauri bayan 2022. Adadin takardu ma ya ƙaru da sauri a kimiyyar kwamfuta da physics bayan 2022, abin da marubutan ke dangantawa da samfuran harshe. Tsallen da aka samu a yawan wallafe-wallafen ilimin likitanci da halittu a 2023 ya fi yiwuwa ya kasance farfaɗowa ne daga koma-bayan 2022 da rikicin zamanin COVID ya haifar.

Ba tukuna ba

Marubutan ba su sami shaida mai yawa ta “sabanin AI” ba a matakin kimiyya baki ɗaya. Suna gargaɗi game da ɗaukar ’yan kalmomin zamani a matsayin hujjar juyin juya hali, kuma suna jayayya cewa manufofi (policy) — game da bayyanawa, game da bitar takwarori (peer review) — su ne kayan aiki da ya dace don sarrafa wannan tankiya. Amma tabbacin da suke bayarwa yana da iyakar lokaci. Binciken ya ƙunshi shekaru uku ne kawai bayan 2022 kuma yana auna sifofin rubutu da ake iya gani, kuma marubutan suna ganin “saurin da LLMs ke shiga cikin tsarin bincike” abin damuwa ne.