공적 영역의 자동화와 인공지능, 그리고 알고리즘적 차별

On artificial intelligence (AI), speed of cultural expectation exceeds that of technological innovation. Thus, poorly shaped cultural expectations on AI can convolute the way technology develops, and even legal policies of how it should be regulated. Algorithmic discrimination, a legal question that bias of data and prejudice of developers may escalate level of discrimination in decision making processes of AI, therefore usually considering transparency of data and ethical training of developers as solutions, is also another clear example of wrongful expectations on AI muddling legal policies. So far, algorithmic discrimination has been regarded as talking point for a horror movie with plot of artificial intelligence superior to humans having secret pathological habit of committing discrimination opaquely. Then why do we end up in that prosaic horror after building such pathological artificial intelligence and putting it into workplace, fearing whether it can commit discrimination without being noticed? Because in this narrative, as AI is vaguely believed as superior to humans, algorithmic discrimination is assumed as an inevitable risk, therefore putting algorithmic efficiency and discrimination in a trade-off relationship. This kind of narrative is understandable in context of adopting AI to private sector where optimization by AI targets monosemic and easily quantifiable management goals like profit. Algorithms for profit maximization have potential to become real superior to humans, yet by pursuing such goals so ruthlessly, other goals like anti-discrimination can be pushed aside and neglected. Trade-off relationships can be proper approach in these situations, by minimizing algorithmic discrimination while respecting management goals. However, in public sector, where goals are mostly polysemic and hard to quantify, such kind of narrative has little room for understanding. AI adoption itself cannot be the goal of AI adoption. Its socio-economic goal, automation, is not just a mechanical acceleration of existing workflow. It is rather a reconstruction of existing workflow to fit quantified optimization goals. Thus, we need to look back on what kind of goal is pursued when bringing AI into public sector. Public organizations can lose their own authentic goals like procedural fairness if they keep seeking secondary goals like cost reduction. Saving money by giving jobs a lick and a promise should not be a desirable goal for the public sector. So, blaming biased data and prejudiced developers, is not a good approach for addressing algorithmic discrimination in the public sector. Automation is not a natural disaster happening suddenly. It is not developers who indiscriminately put opaque and ruthless AI algorithms into public sector, while public organizations were sitting still. When bringing AI into public sector to promote automation, public organizations like state have substantive power to oversee whole processes of adoption such as setting specific development goals, hiring certain developers, sorting out data for training, and checking achievement of development goals, because they are the contractors who ordered such AI. Since such public organizations in this context can be regarded as least cost avoider from concept of law and economics, when bringing AI into the public sector, it is a necessary duty of public organization in charge, to examine whether optimization goals and benchmark test standards are set properly. Therefore, public organizations like state, as the contractor who ordered development of that AI, should be held primarily responsible before developers, for algorithmic discriminations resulting from improper implementation of such duty.