构建合乎道德的人工智能：AI数据标注圆桌会议成果回顾

ROUNDTABLEREADBACK This readback is based on a roundtable on AI Data Labelling, held on the 16th of June 2020,hosted by Aapti Institute and Dr. Alex Taylor, City, University of London. This is part of a largerresearch collaboration that intends to shed light on labelling and annotation processes, andcollaboratively envision best practices for fair and equitable Artificial Intelligence. This exploratorydiscussion brought together startups, researchers and civil society leaders to unpack the processesand structures that are involved in data labelling or annotation work. - B U I L D I N G E T H I C A L A R T I F I C I A L I N T E L L I G E N C E T O G E T H E R - A I D A T A L A B E L L I N G The COVID-19 pandemic and lockdown are shifting the structure of markets and nature of availablejobs, well-beyond the present moment. With these immediate changes and longer-term trends, itwould seem timely to revisit and explore the state of technologically-mediated labour, where work isboth found and delivered online. Also presenting pressing questions is the increasing promotion of AIin emerging and evolving forms of contemporary work. AI and machine learning (ML) are beingtouted by multinational technology firms as the backbone to future visions of work, particularlyremote work. Yet many questions remain about the specific role of computational systems in thesenew forms of work and the shape such AI and ML systems should take. Furthermore, we are havingto confront longstanding and in some cases new questions about automation in the workplace andwhat we imagine to be not just productive but fair and just ideas of labour. It was with this as a backdrop that the roundtable we report on below was held. Rather than attendto the oft-hyped, idealised visions of work, however, we opted to pay attention to the work ‘behindthe scenes’, conducted to make AI/ML work. Specifically, we wanted the primary focus of theroundtable to be on a labour that is rarely discussed and often invisible to end users, data labelling orannotation. Our interest was (and remains) in how data labelling or annotation plays a critical role inthe success of the future visions of work, and what practices and structures are necessary to achievesuch a success. We thus wanted to promote a discussion to better understand the conditions of thislabour and both the social and technical issues that arise in enabling it. Annotation and labelling for Artificial Intelligence (AI) offer employment opportunities for many.Across multiple sectors including e-commerce, autonomous vehicles and healthcare, AI is beingpromotedas an innovative solution to provide analytics, automate processes and evolve businessmodels.However, AI is dependent on the availability of labelled and classified datasets. Withoutnecessary meanings attached to data, machine learning models cannot be trained or function in thereal-world. This critical role is played by humans who have the skills and know-how to classify newdata sets, and approach problems of uncertainty and ambiguity in data and context in ways thatcomputer models find notoriously hard to replicate. To carry out these tasks, crowdsourcing platforms and startups often employ thousands of workersto label the datasets of text, images, video, etc. Previous work, including from one of our roundtableparticipants, Mary Gray, has shown the difficult labour conditions that these workers can face,surfacing not only the hidden nature of this work, but its potential for low wages, exploitation andabuse. It is clear that pathways for improving labour conditions and regulation need to be set out andwhere possible used to help inform national public and organisational policy programmes. In order forus to do this, we need a fuller understanding and more systematic representation of the currentbusiness models, the nature of work, and concerns of workers. Building and developing regulation formore equitable AI depends on the recognition and deeper understanding of this labor. As the futureof work becomes more digital, it is imperative to understand how this work is done, who does it, andhow it can be structured and performed to enhance dignity and protect rights, while enablinginnovation. M E T H O D O L O G Y The Roundtable was divided into two sections, the first featured presentations from three startups:Playment, TaskMonk and iMerit. As limited information exists on the state of AI labelling, startupsprovided detailed insight into the industry and current processes. The second half provided time forresearchers to discuss their work and explore areas for future research collaboration. This alsoenabled startups to engage in discussion around research, policy and upcoming trends. SPEAKERS MODERATORS PART 1:Connecting with the Startup Ecosystem Startups like Playment have emerged in response to the inadequacies such as the lack of consistency inquality from legacy platforms likea Amazon Mechanical Turk. Playment seeks to democratize theav