How can we create a more just society with A.I.?

正义可以被视为“客观的”或通过权力进行调解[乔姆斯基和福柯，1971；科斯坦萨-乔克，2018]。在不同的法律和道德框架中寻找共同点[Floridi&Cowls，2019；Jobin等人，2019年]就是前者的一个例子。在后一种情况下，正义是对非公平社会的“要求”，确保对受伤害最严重的人提供保护[Cugueró-EScott fet&Fortin，2014]。通过人工智能实现这种类型的司法的困难在于，人工智能主要用于分类和预测[Vinuesa等人，2020]。越来越多的证据表明，人工智能加速并加剧了社会偏见，助长了权力的不平等分配[O‘Neil，2016，第3页，Noble，2018；Benjamin]。在提供准确和公平的预测方面的“取舍”也不成比例地影响亚群体[Yu等人。2020]，这意味着具有多种形式的边缘化的人比具有规范特征的人更容易被人工智能误解[Costanza-Chock，2018]。虽然应该有法律和道德框架来管理我们使用人工智能的方式，但少数群体的声音仍然代表不足[Buolamwini，J.和Gebru，T.，2018，Costanza-Chock，2018；Magalhães&Coulry，2020]，而且几乎没有执法和问责的结构[Mittelstadt，2019]。我们需要重新思考人工智能是如何将正义作为一个关系概念来贡献的，其中包括权力和边缘化的维度。我的建议汇集了必要的文化、技术和社会技术专业知识，以扩展我们目前在社会公益人工智能(AI4SG)实证研究中的正义概念。首先，核心团队将开发一个人工智能和“正义”的概念模型，其中包括a)用来框定人工智能任务并评估其有效性的不同正义定义，b)根据该定义可以回答的问题，以及c)在这一过程中被确定为可以接受的权衡。研究小组将把AI4SG的学术文献与支撑这项研究的伦理、法律或政治框架联系起来，找出与其他社会正义模式相比，AI4SG内部的正义如何运作的差距或冲突。特别是，我们将探讨这样的问题：对正义的不同立场是否与人工智能不相容？我们能找到正义出现的新途径吗？为了扩展我们的概念模型，我们将进行三个案例研究，在这些案例中，少数群体的利益在特定的人工智能任务中被忽视：1)基于性别的性别歧视分析中的非二元人群；2)通过内容审查歧视性地去掉性工作者或艺术家的平台；3)作为反恐方法的一部分，禁止影子活动人士。案例研究将探索这些社区的正义概念与人工智能任务之间的冲突，以及存在哪些替代解决方案。他们还将为应对网络危害和使用人工智能技术帮助识别和分类相关案例的全球问题做出贡献。最后，为了测试替代解决方案，一个由人工智能组成的多部门咨询委员会和社区专家将聚集在一起，为人工智能研究人员创造一个设计挑战。通过两个顶级人工智能会议的研讨会发布，挑战将是优先考虑边缘化的观点。挑战的结果及其评估将为处理AI4SG中的错误和权衡提供一套指导方针。我们的贡献是a)揭示人工智能研究人员如何定义正义与我们在AI4SG中关注哪些正义问题之间的联系；b)反思人工智能对哪些社会的好处；以及c)影响并激励研究人员质疑人工智能研究围绕可接受的权衡和错误的假设。这项研究将把社会科学家、社区专家和人工智能研究人员聚集在一起，探索通过专注于最大化被边缘化群体的人工智能的好处，可以开辟哪些新的调查路线

Justice can be viewed as "objective" or mediated through power [Chomsky & Foucault, 1971; Costanza-Chock, 2018]. Finding commonalities across different legal and ethical frameworks [Floridi & Cowls, 2019; Jobin et al., 2019] is an example of the former. In the latter, justice is a "requirement" for non-equitable societies, ensuring protection for the most harmed [Cugueró-Escofet & Fortin, 2014]. The difficulty in achieving this type of justice through A.I. is that A.I. is used primarily for classification and prediction [Vinuesa et al., 2020]. Growing evidence indicates that A.I. accelerates and compounds social bias, contributing to unequal distributions of power [O'Neil, 2016, p. 3, Noble, 2018; Benjamin]. "Trade-offs" in providing accurate and fair predictions also impact sub-populations disproportionately [Yu et al. 2020], meaning that people with multiple forms of marginalisation are more likely to be misunderstood by A.I. than those with normative characteristics [Costanza-Chock, 2018]. While there are legal and ethical frameworks that should govern the way we use A.I., minority voices are still under-represented [Buolamwini, J. and Gebru, T., 2018, Costanza-Chock, 2018; Magalhães & Couldry, 2020] and there are few structures for enforcement and accountability [Mittelstadt, 2019]. We need to rethink how A.I. is contributing to justice as a relational concept, which includes dimensions of power and marginalisation. My proposal draws together the cultural, technical, and socio-technical expertise necessary to extend our current notions of justice in empirical research for A.I. for social good (AI4SG). To start with, the core team will develop a conceptual model of A.I. and "justice" that includes a) different definitions of justice used to frame the tasks of A.I. and evaluate their efficacy, b) the questions that can be answered under that definition and c) the trade-offs that are determined to be acceptable in the process. The research team will map scholarly literature from AI4SG to the ethical, legal or political frameworks that underpin the research, identifying gaps or conflicts in how justice is operationalised within AI4SG in comparison with other social justice models. In particular, we will explore the questions: are different positions on justice incompatible with A.I.? Can we identify new pathways for justice to emerge? To extend our conceptual model, we will conduct 3 case studies in which minority interests are ignored within specific A.I. tasks: 1) non-binary people in gender-based analysis of sexism 2) discriminatory deplatforming of sex workers or artists through content moderation and 3) shadow-banning activists as part of a counter-terrorism approach. The case studies will explore conflicts between these communities' concept of justice and the A.I. task, and which alternative solutions exist. They will also contribute to the global problem of tackling online harm and using A.I. techniques to help identify and classify relevant cases.Finally, to test alternative solutions, a multi-sectoral Advisory Board of A.I. and community experts will be brought together to create a design challenge for A.I. researchers. Issued through 2 workshops at top-level A.I. conferences, the challenge will be to prioritise marginalised perspectives. The outputs of the challenge and their evaluation will inform a set of guidelines for dealing with errors and trade-offs in AI4SG. Our contribution is to a) expose connections between how A.I. researchers define justice and which justice questions we attend to in AI4SG; b) reflect on the benefits of A.I. for which societies; and c) influence and inspire researchers to question assumptions of A.I. research around acceptable trade-offs and errors. This research will bring together social scientists, community experts and A.I. researchers to explore what new lines of inquiry can be opened by focusing on maximising the benefits in A.I. for marginalised groups