When Algorithms Discriminate
Algoritmlar kamsitganda
Sarlavha va pastdagi so‘zlarga qarang. Matn nima haqida bo‘lishi mumkin? Bir daqiqa o‘ylab ko‘ring — bu miyani o‘qishga tayyorlaydi.
Sun’iy intellekt tizimlari nega xolis bo‘lmasligi va buni qanday tuzatish mumkin.
Kalit so‘zlar
Bu so‘zlar matnda uchraydi va matnda ajratib ko‘rsatiladi. Bosing — talaffuzini eshitasiz.
⏱ O‘qish boshlanganda vaqt o‘lchanadi — oxirida tezligingizni ko‘rasiz.
👆 Istalgan so‘zni bosing — tarjimasi yoki lug‘atga havola chiqadi.
Algorithms are increasingly used to make or inform decisions with profound consequences: who is shortlisted for a job, who receives a loan, which neighbourhoods are patrolled by police. Their appeal lies partly in a presumption of objectivity; a machine, it is assumed, cannot be prejudiced.
Algoritmlar tobora chuqur oqibatlarga ega qarorlarni qabul qilish yoki ularga ta’sir ko‘rsatish uchun ishlatilmoqda: kim ishga tanlov ro‘yxatiga kiritiladi, kim kredit oladi, qaysi mahallalarni politsiya kuzatadi. Ularning jozibasi qisman xolislik taxminiga asoslangan; mashina, deb taxmin qilinadi, g‘arazli bo‘la olmaydi.
This assumption has proved naive. Machine-learning systems learn patterns from historical data, and where that data reflects past discrimination, the system may reproduce it. In one widely reported case, a large technology company abandoned an experimental recruitment tool after discovering that it downgraded applications containing the word “women’s”, as in “women’s chess club captain”, because it had been trained on a decade of CVs submitted predominantly by men.
Bu taxmin sodda bo‘lib chiqdi. Mashinali o‘qitish tizimlari tarixiy ma’lumotlardan qonuniyatlarni o‘rganadi va bu ma’lumotlar o‘tmishdagi kamsitishni aks ettirsa, tizim uni takrorlashi mumkin. Keng yoritilgan bir holatda yirik texnologiya kompaniyasi tajriba tariqasidagi ishga yollash vositasidan voz kechdi: u “ayollar shaxmat klubi sardori” kabi “ayollar” so‘zi bor arizalarni pastroq baholagani aniqlangan, chunki u asosan erkaklar topshirgan o‘n yillik rezyumelarda o‘qitilgan edi.
Bias can also enter more subtly. Even when protected characteristics such as race or gender are excluded, other variables, such as postcode or browsing history, may act as proxies for them. Moreover, many systems are opaque: their internal logic cannot easily be explained, even by their designers, making errors difficult to detect and challenge.
Xolislik buzilishi yanada nozik yo‘llar bilan ham kirib kelishi mumkin. Irq yoki jins kabi himoyalangan xususiyatlar chiqarib tashlanganda ham, pochta indeksi yoki brauzer tarixi kabi boshqa o‘zgaruvchilar ularning o‘rinbosari vazifasini bajarishi mumkin. Bundan tashqari, ko‘p tizimlar shaffof emas: ularning ichki mantiqini hatto yaratuvchilari ham oson tushuntira olmaydi, bu esa xatolarni aniqlash va ularga e’tiroz bildirishni qiyinlashtiradi.
Remedies are emerging. Independent audits can test systems for disparate outcomes before deployment; regulations such as the European Union’s AI Act impose stricter requirements on “high-risk” uses. Ultimately, however, technical fixes cannot resolve what are, at heart, questions of values. Deciding what counts as a fair outcome is a political and ethical judgement that should not be delegated to engineers alone.
Davolash choralari paydo bo‘lmoqda. Mustaqil auditlar tizimlarni joriy etishdan oldin notekis natijalarga tekshirishi mumkin; Yevropa Ittifoqining sun’iy intellekt to‘g‘risidagi qonuni kabi qoidalar “yuqori xavfli” foydalanishlarga qat’iyroq talablar qo‘yadi. Biroq oxir-oqibat texnik tuzatishlar mohiyatan qadriyatlar haqidagi savollarni hal qila olmaydi. Qaysi natija adolatli ekanini belgilash faqat muhandislarga topshirilmasligi kerak bo‘lgan siyosiy va axloqiy mulohazadir.
Facial recognition provides another instructive case. Independent testing in 2018 found that several commercial systems misidentified the gender of darker-skinned women far more often than that of lighter-skinned men, largely because the images used to train them were unrepresentative. Following public criticism, some developers improved their training data and error rates fell substantially, demonstrating that such problems can be measured and reduced.
Yuzni tanish yana bir ibratli holat. 2018-yildagi mustaqil sinovlar bir nechta tijorat tizimlari qora tanli ayollarning jinsini oq tanli erkaklarnikiga qaraganda ancha tez-tez noto‘g‘ri aniqlashini ko‘rsatgan, asosan ularni o‘qitishda ishlatilgan tasvirlar vakillik qilmagani sababli. Jamoatchilik tanqididan so‘ng ba’zi ishlab chiquvchilar o‘qitish ma’lumotlarini yaxshiladi va xatolar darajasi sezilarli kamaydi, bu esa bunday muammolarni o‘lchash va kamaytirish mumkinligini ko‘rsatdi.
A further complication is that different mathematical definitions of fairness can be mutually incompatible. A system cannot, in general, simultaneously ensure equal error rates across groups and equal accuracy of its positive predictions when the underlying base rates differ. Choosing between these definitions is unavoidable, and it is precisely the kind of choice that should be made transparently rather than buried in code.
Yana bir murakkablik shundaki, adolatning turli matematik ta’riflari o‘zaro mos kelmasligi mumkin. Asosiy ko‘rsatkichlar farq qilganda tizim odatda guruhlar bo‘ylab teng xato darajasini va ijobiy bashoratlarining teng aniqligini bir vaqtda ta’minlay olmaydi. Bu ta’riflar orasida tanlov muqarrar va aynan shunday tanlov kodga ko‘milmasdan, shaffof tarzda qilinishi kerak.
Tushundingizmi?
Javob topa olmasangiz, «O‘qish» bosqichiga qaytib, matnga qarang — bu ham mashq.
-
Asosiy g‘oya
1 What is the main argument of the article?
-
2 Why did the recruitment tool downgrade applications mentioning “women’s”?
-
3 What does the article mean by “proxies”?
-
4 Why is opacity a problem?
True / False / Not Given
True — matnda aynan shunday deyilgan. False — matn aksini aytadi. Not Given — matnda bu haqda ma’lumot yo‘q.
-
1 Removing race and gender from the data always eliminates bias.
-
2 The technology company was fined for its recruitment tool.
-
3 The EU’s AI Act sets stricter requirements for high-risk uses.
-
4 All definitions of fairness can always be satisfied at the same time.
-
5 Some facial recognition systems improved after criticism.
-
6 Facial recognition has been banned in the United States.
So‘zni ma’nosi bilan juftlang
Chapdan so‘zni, o‘ngdan uning ma’nosini tanlang. Hammasi juftlansa, natija chiqadi.