Přehled o publikaci

The paper presents the findings and recommended remedies based on an interdisciplinary research done by international lawyers and AI experts which proves indispensable to tackle such questions, within the framework of a grant research project „Artificial Intelligence and Human Rights: Risks, Opportunities and Regulation“, supported by the Technological Agency of the Czech Republic. The relation between the AI and human rights is for now double-edged. AI brings great benefits to all sectors of society and strengthens the progress, social well-being and economic competitiveness. At the same time, however, it poses risks to a variety of human rights and fundamental freedoms, be it due to the intrinsic technological processes, human input or its abusive or malicious use in practice. Prohibition of discrimination is one of the outstanding rules being at stake across all types of AI technologies. Predictive justice biased against an ethnical minority? Grant of mortgage discriminating one gender? Denial of access at the airport via facial recognition? These current cases raise the topicality of such an urgent task to be overcome. Ensuring a nonbiased AI technology is an overarching challenge for all regulators and actors of the whole AI life-cycle, in order to offer the society an effective and lawful tool. In order to assure the effective implementation of the human rights norm of the prohibition of discrimination in the AI domain, it is first necessary to identify the root problem of the bias in order to formulate a remedy. The root problem can reside in all phases of the AI-life cycle, starting with the incomplete input data, over biased transfer learning up to a malicious application. Adequately, the remedy is diverse: the problem might be solved via a technical adjustment of, for example, the machine learning procedure, or rules on the processing of data, or a more robust legal interference restricting or banning the development and use of certain AI technologies if the technical remedy is not sufficient. Another challenge of the transposition of the prohibition of discrimination into the AI domain is the quantification of the protected value as well as its (dis)regard: Can AI-based facial recognition system be constructed without taking into account certain external human features? And how to avoid a bias caused by incomplete data? In addition, indirect discrimination multiplies the challenges in view of the difficult technical preliminary detectability of relevant data and links.