数据标注治理:可信人工智能的后台风险与治理转向  

Data Annotation Governance:Backstage Risks and Governance Shifts for Trustworthy AI

作　　者：胡泳[1] 张文杰 HU Yong;ZHANG Wen-jie(School of Journalism and Communication,Peking University,Beijing,100871)

机构地区：[1]北京大学新闻与传播学院

出　　处：《云南社会科学》2024年第6期29-36,共8页Social Sciences in Yunnan

摘　　要：在人工智能模型训练前,须先对数据进行人工识别与标注,因此处于“后台”的数据标注成为引致人工智能“前台”幻觉与偏见风险的重要环节。近年来,隐蔽于人工智能后台的数据标注劳动者随着政策文件、媒体报道和调查报告的公布而逐渐浮现,进一步推动学界警惕与反思技术革新的神秘化倾向。然而,从风险治理的角度来看,参与数据标注实践的多元主体仍然处于权责模糊的灰色地带,阻碍了可信人工智能目标的实现。全球主要经济体的数据标注治理路线显示,当前针对数据标注的治理对象以“人工智能服务提供者”为主,且倾向于将数据置于私人个体秩序中。通往可信人工智能的风险治理亟须将治理范围从“提供者”拓展至“数据供应链”,建设一种多元主体共同参与的集体性治理制度,进而更为细致地关照人工智能生产中的相关群体利益,为不稳定的数据劳动者提供具体的社会保障。Before training artificial intelligence(AI)models,it is necessary to manually identify and annotate the data.Therefore,the“backstage”data annotation process is one of the most important stages that can lead to the risks of illusions and bias in the“frontstage”of AI.Recent years have seen data annotators,previously hidden in AI’s background,gain visibility through policy documents and media reports,prompting the academic community to reflect on the enigmatic nature of technological innovation.However,from the perspective of risk governance,the numerous players participating in the practice of data annotation remain in a state of confusion,with ambiguous rights and obligations,impeding the goal of trustworthy AI.This paper traces and compares the data annotation governance of the leading economies in the global AI industry.It finds that the current data annotation governance targets emphasize mainly“AI service providers”,and there is a tendency to place data within the order of private individuals.The risk governance towards trustworthy artificial intelligence urgently needs to expand the scope of governance from“providers”to the“data supply chain”,and to establish a collective governance system participated by multiple stakeholders.This will allow a deeper examination of the interests of the groups involved in AI industry and provide tangible social security for the unstable data labor force.

关 键 词：数据标注 数据治理 幻觉 偏见 幽灵工作 

分 类 号：D63[政治法律—政治学]