인공지능의 법적 규율 II: 판별모델
Legal Oversight of AI, Part Two: Discriminative Model
0
조회
0
다운로드
3
인용
0
참고문헌
키워드
초록
판별모델(discriminative model)은 주로 지도학습 기반의 분류기 또는 채점기이다. 그간 판별모델의 법적 통제는 불투명성과 자동화에 의한 인간소외, 존엄성 훼손에 초점이 맞춰졌다. 이는 2020년 신용정보법상 “자동화평가” 설명요구⋅이의제기권, 2021년 행정기본법상 “자동적 처분”의 재량행위 활용 금지와 기속행위 활용 법률유보, 2023년 개인정보법상 “자동화된 결정” 거부권⋅설명등요구권⋅공개의무 조항으로 이어진다. 그러나 소외⋅존엄은 자동성보다는 자율성의 문제이고, 불투명성⋅편향이 사람의 판단에 현저한 경우도 많으니, 자동화의 금지⋅제약이 제대로 된 해법이 될 수 없다. 특히, 취업준비생 등 판별대상자들이 현실에서 겪는 어려움을 덜려면, 소외⋅존엄 등 추상 개념에 천착할 것이 아니라, 완전자동⋅부분자동⋅수동을 불문하고 판별 작업의 정확성, 가치정렬성, 비차별성, 다원성을 기해야 한다. 따라서 입법적 과제는 의사결정 자동화의 억지⋅지연에 있는 것이 아니라, (1) 기존의 사람에 의한 결정의 성실성까지 성찰하면서, (2) 귀납⋅실증⋅데이터에 근거한 판별모델의 도입을 오히려 촉진하고, (3) 이로써 채점기준(rubric)의 공개 등 더 나은 설명을 기하되, (4) 판별대상자들이 부당하게 차별받고 배제되지 않도록 안전장치를 설정함에 있다. 이 관점 하에 다음 단계를 거쳐 기존 입법의 대안을 강구한다. 첫째, 판별모델을 사람과의 상호작용 방식에 따라 유형화하여 (1) 배분모델(채용⋅입학사정⋅대출심사⋅신용평가⋅보험계약인수⋅사회보장⋅랭킹 등)은 차별, (2) 제재모델(부정계정탐지⋅사기탐지⋅보험손해사정⋅재범예측 등)은 오판(정확성의 부족), (3) 인지모델(의료영상진단⋅센서⋅생체인증⋅추천 등)은 안전을 해하거나 배분⋅제재모델의 특성값에 오류를 초래할 정도의 정확성(견고성 포함)의 부족(품질편차 문제 포함), 프라이버시를 해할 정도의 정확성의 과잉을 핵심 위해로 지목한다. 모델에 대한 설명⋅공개는 배분⋅제재모델에서 판별대상자에게 대칭적 정보 하에서의 개선의 기회 또는 차별에 문제제기할 근거를 부여하는 장치로 구성한다. 둘째, 법적 개입을 정당화할 만큼 (1) 공적 영역에서 활용되거나 판별대상자의 삶을 좌우하는 배분⋅제재모델, (2) 안전, 프라이버시 리스크가 있거나 배분⋅제재모델에 투입될 특성값을 매기는 인지모델을 식별한다. 셋째, 이같이 식별된 모델들에 대해 앞서 지목된 핵심 위해를 측정⋅경감하는 방법과 이를 뒷받침할 법적 수단을 검토한다. 넷째, 상기 검토를 종합하여 현 자동화 평가⋅처분⋅결정 규제의 대안을 제시한다.
Abstract
A discriminative model is largely a supervised learning-based classifier or regressor/interpolator. Legal responses thus far have primarily targeted bias and opacity arising from “black boxes” or “algorithmic alienation.” These notions have led to regulations such as the explanation requirement for automated evaluation under the Credit Information Act of 2020, the prohibition of discretionary actions in automated disposition under the Framework Act on Administration of 2021, and the explanation and disclosure requirement for automated decision-making under the Personal Information Protection Act of 2023. However, bias, opacity, and alienation are not exclusive to AI. They already exist in human judgment processes that discriminative models seek to replace or complement, sometimes to a greater extent. Viewing the emergence of AI as an opportunity to reflect on existing human decisions, the legal system must be revamped to ensure fairness and acceptability in decisions impacting others’ lives, regardless of being automated or not. From this perspective, the following steps are taken to examine the problems of existing legislation and seek alternatives. First, discriminative models are categorized based on how they interact with humans, and different harms are mapped for each category: (1) discrimination in allocative models (such as hiring, admission, loan review, credit scoring, insurance underwriting, public assistance, ranking, etc.); (2) misjudgment (lack of accuracy) in punitive models (such as fraud detection, insurance claim adjustment, recidivism prediction, etc.); and (3) either lack of accuracy (including robustness) jeopardizing safety or excessive accuracy compromising privacy in cognitive models (such as medical imaging and diagnosis, sensors, biometric authentication, recommendations, etc.). Instead of conceptualizing opacity as a stand-alone harm, explanations and disclosures about the model are structured as means to provide affected individuals with opportunities for improvement under symmetric information or grounds for challenging discrimination in allocative models or quality disparity in punitive models. Second, (1) allocative or punitive models either deployed in the public domain or decisive of individual lives and (2) cognitive models posing safety or privacy risks are identified to justify legal intervention. Third, tools for measuring and mitigating the identified core concerns of these models and appropriate legal means are reviewed. Fourth, alternatives to existing automated decision-making rules are explored.
참고문헌 (74)
- 고학수, "인공지능과 차별", 저스티스, 통권 : 171 (2019)
- 고학수, "인공지능과 고용차별의 법경제학: 블라인드 채용과 베일의 역설을 중심으로", 법경제학연구, 16(1) (2019)
- 채용분야 인공지능(AI)활용실태 및 공정성 확보방안 연구, "권오성", 한국산업인력공단, ~, pp.한국산업인력공단 위탁연구 2023-12 (2023)
- 권은정, "자동적 처분의 행정법적 제 문제: 처분의 형식･절차 및 구제에 관한 쟁점을 중심으로", 법학논총, 38(3) (2021)
- 김동희, "행정법 Ⅰ" (2021)
- 김병필, "대규모 언어모형 인공지능의 법적 쟁점", 정보법학, 26(1) (2022)
- 김병필, "인공지능 공정성 심사 기준에 관한 연구", 박사, pp.2023
- 김예진, "채용절차에 있어서 인공지능의 차별 문제: 미국과 한국의 비교를 중심으로", 노동법연구, (55) (2023)
- 김혜진, "알고리듬 행정결정의 법적 쟁점: 프랑스 대학입학플랫폼(Parcoursup)의 사례를 중심으로", 행정법연구, (69) (2022)
- 김현섭, "민주주의의 퇴행이란 무엇이고, 어떻게 발생, 진행되며, 왜 문제인가?: 조망과 평가", 한국정치연구, 32(3) (2023)
- 모정훈, "알고리즘 공개에 관하여", 디지털사회, (36) (2021)
- 박노형, "자동화된 결정에 관한 개인정보보호법 정부 개정안 신설 규정의 문제점: EU GDPR과의 비교 분석", 사법, 1(62) (2022)
- 박상철, "금융AI의 활용과 금융소비자 보호: 차별금지, 설명요구권, 6대 판매규제 준수를 중심으로", BFL, (107) (2021)
- 박상철, "플랫폼법", 저스티스, 통권 : 188 (2022)
- 유동훈, "자동적 재량행위 가능성에 관한 시론적 연구", 법제, (697) (2022)
- 이상용, "개인정보보호법에 의한 자동화된 결정의 규율", 선진상사법률연구, 통권 : 105 (2024)
- 이상용, "개인정보보호법상 자동화된 결정 조항의 해석", 법조, 73(1) (2024)
- 임성훈, "인공지능 행정이 행정절차⋅행정소송에 미치는 영향에 대한 시론적 고찰", 행정법연구, (62) (2020)
- 정남철, "인공지능 시대의 도래와 디지털화에 따른 행정자동결정의 법적 쟁점: 특히 행정기본법상 자동적 처분의 문제점을 중심으로", 공법연구, 50(2) (2021)
- 조성규, "인공지능에 기반한 자동화된 행정결정의 행정법적 쟁점", 동북아법연구, 16(4) (2023)
- Adamson, A. S., "Machine Learning and Health Care Disparities in Dermatology", JAMA Dematology, 154(11) (2018)
- Agirre, Eneko, "On Robustness and Domain Adaptation using SVD for Word Sense Disambiguation", Computational. Linguistics (2018)
- Apley, Daniel W., "Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models", Journal of the Royal Statistical Society, Series B (Methodology), 82(4) (2020)
- Aristotle, "Nicomachean Ethics, bk. V" (1999), Batoche Books
- Article 29 Data Protection Working Party, "Guidelines on Automated Individual Decision-Making and Profiling for the Purposes of Regulation 2016", 679, 2018
- Barocas, Solon, "Big Data’s Disparate Impact", California Law Review, 104 (2016)
- "Fairness and Machine Learning: Limitations and Opportunities" (2023)[URL]
- Bambauer, Jane, "The Algorithm Game", Notre Dame Law Review, 94 (2018)
- Blodgett, Su Lin, "Language (Technology) is Power: A Critical Survey of ‘Bias’ in NLP", Annual Meeting of the Association for Computational Linguistics (ACL) (2020)
- Castelvecchi, Davide, "Can We Open the Black Box of AI?", Nature, 538 (2016)
- Cofone, Ignacio N., "Strategic Games and Algorithmic Secrecy", McGill Law Journal, 64(4) (2019)
- Crawford, Kate, "The Trouble with Bias", a keynote at Conference on Neural Information Processing Systems (NIPS) (2017)
- D’Amour, A., "Underspecification Presents Challenges for Credibility in Modern Machine Learning", Journal of Machine Learning Research, 23(1) (2022)
- Dandekar, Pranav, "Biased Assimilation, Homophily, and the Dynamics of Polarization", Proceedings of the National Academy of Sciences (PNAS), 110(15) (2013)
- Danziger, Shai, "Extraneous Factors in Judicial Decisions", Proceedings of the National Academy of Sciences (PNAS), 108(17) (2011)
- Dietvorst, Berkeley J., "Algorithm Aversion : People Erroneously Avoid Algorithms After Seeing Them Err", Journal of Experimental Psychology: General, 144(1) (2015)
- Dwork, Cynthia, "Fairness Through Awareness", Proceedings of the 3rd Innovations in Theoretical Computer Science Conference (ITCS) (2012)
- Dwork, Cynthia, "The Algorithmic Foundation of Differential Privacy", Foundations and Trends® in Theoretical Computer Science, 9(3–4) (2014), Now Publishers
- Dworkin, Gerald, "Autonomy and Behavior Control", The Hastings Center Report, 6(1) (1976)
- Epstein, Robert, "The Search Engine Manipulation Effect (SEME) and its Possible Impact on the Outcomes of Elections", Psychological and Cognitive Sciences, 112(33) (2015)
- Fagan, Frank, "The Impact of Artificial Intelligence on Rules, Standards, and Judicial Discretion", Southern California Law Review, 93(1) (2019)
- Friedman, Jerome H., "Greedy Function Approximation : A Gradient Boosting Machine", The Annals of Statistics, 29(5) (2001)
- Guo, L.N., "Bias In, Bias Out: Underreporting and Underrepresentation of Diverse Skin Types in Machine Learning Research for Skin Cancer Detection－A Scoping Review", Journal of American Academic Dermatology, 87(1) (2022)
- Guyon, Isabelle, "Design of the 2015 ChaLearn AutoML Challenge", International Joint Conference on Neural Networks (IJCNN) (2015)
- Jussupow, Ekaterina, "Why Are We Averse Towards Algorithms? A Comprehensive Literature Review on Algorithm Aversion", European Conference on Information Systems (ECIS2020) (2020)
- Keijzer, Marijn A., "The Complex Link between Filter Bubbles and Opinion Polarization", Data Science, 5(2) (2022)
- Kim, Been, "Examples are not Enough, Learn to Criticize! Criticism for Interpretability", Conference on Neural Information Processing Systems (NIPS) (2016)
- Larrazabal, A.J., "Gender Imbalance in Medical Imaging Datasets Produces Biased Classifiers for Computer-aided Diagnosis", Proceedings of the National Academy of Sciences (PNAS), 117(23) (2020)
- Lee M. S., "Towards Gender Equity in Artificial Intelligence and Machine Learning Applications in Dermatology", Journal of the American Medical Informatics Association, 29(2) (2022)
- Lundberg, Scott M., "A Unified Approach to Interpreting Model Predictions", Conference on Neural Information Processing Systems (NIPS) (2017)
- Logg, Jennifer M., "Algorithm Appreciation : People Prefer Algorithmic to Human Judgment", Organizational Behavior and Human Decision Processes, 151 (2019)
- Madhavan, P., "Similarities and Differences Between Human-Human and Human-Automation Trust : An Integrative Review", Theoretical Issues in Ergonomics Science, 8(4) (2007)
- Malgieri, Gianclaudio, "Why a Right to Legibility of Automated Decision-Making Exists in the General Data Protection Regulation", International Data Privacy Law, 7(4) (2017)
- Martini, Mario, "Subsumtionsautomaten ante portas? – Zu den Grenzen der Automatisierung in verwaltungsrechtlichen(Rechtsbehelfs-)Verfahren", Deutsches Verwaltungsblatt, 133(17) (2018)
- Ng, Andrew, "CS229 Lecture Notes", ~ (2020)
- Park, Sangchul, "Bridging the Global Divide in AI Regulation: A Proposal for Contextual, Coherent, and Commensurable Framework", Washington International Law Journal, 32(2) (2024)
- Pasquale, Frank, "The Black Box Society: The Secret Algorithms That Control Money and Information" (2016), Harvard University Press
- Ramji-Nogales, Jaya, "Refugee Roulette : Disparities in Asylum Adjudication", Stanford Law Review, 60 (2007)
- Schmitz, Heribert, "Neues zum E-Government: Rechtsstaatliche Standards für E-Verwaltungsakt und E-Bekanntgabe im VwVfG", Neue Zeitschrift für Verwaltungsrecht (NVwZ) (2016), C.H. Beck
- Selbst, Andrew, "Meaningful Information and the Right to Explanation", International Data Privacy Law, 7(4) (2017)
- Shapley, Lloyd S., "Notes on the N-Person Game－II: The Value of an N-Person Game", ~ (1951)
- Soares, Nate, "The Technological Singularity" (2017), MIT Press
- Tramèr, Florian, "Stealing Machine Learning Models via Prediction APIs", USENIX Security Symposium (SEC) (2016)
- Turing, Alan M., "On Computable Numbers, with an Application to the Neue Zeitschrift für VerwaltungsrechtEntscheidungsproblem", Proceedings of the London Mathematical Society, Series 2, 42(1) (1937)
- Tversky, Amos, "Extensional versus Intuitive Reasoning : The Conjunction Fallacy in Probability Judgement", Psychological Review, 90(4) (1983)
- Veale, Michael, "Clarity, Surprises, and Further Questions in the Article 29 Working Party Draft Guidance on Automated Decision-making and Profiling", Computer Law & Security Review, 34(2) (2018)
- Verma, Sahil, "Fairness Definitions Explained", Proceedings of the International Workshop on Software Fairness (2018)
- Wachter, Sandra, "Counterfactual Explanations without Opening the Black Box : Automated Decisions and the GDPR", Harvard Journal of Law & Technology, 31 (2017)
- Wachter, Sandra, "Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation", International Data Privacy Law, 7(2) (2017)
- Weinshall-Margel, Keren, "Overlooked Factors in the Analysis of Parole Decisions", Proceedings of the National Academy of Sciences (PNAS), 108(42) (2011)
- Windoffer, Alexander, "Verwaltungsverfahrensgesetz: Großkommentar" (2019)