PDF(1664 KB)
PDF(1664 KB)
PDF(1664 KB)
点击搜索之前：针对搜索引擎自动补全算法偏见的实证研究
Before Clicking Enter: An Empirical Study of Search Engine Autocomplete Algorithmic Bias
人工智能的广泛应用引发了学术界关于算法价值偏见的担忧和探讨。以搜索引擎自动补全算法为研究对象，本研究运用算法审计方法，在模拟创建大规模用户搜索词的基础上，对收集自主流搜索引擎的47011条自动补全预测词条所反映出的价值偏见进行了检验。研究发现，在性别、年龄、户籍三类社会群体属性上，自动补全算法偏见与长期存在的社会歧视一致，反映出女性相较于男性、中老年群体相较于青年、农村相较于城市的不利社会地位；特征话题与搜索平台均显著调节这三类属性与算法负面偏见之间的关系。研究认为，在用户交互实践中社会偏见通过自动补全算法强化了其媒介可见性，而平台则受到流量商品逻辑的影响成为不平等再现的媒介载体。研究进一步探讨了算法偏见与数字不平等的交互机制、以及在对抗算法偏见过程中不同利益相关方的作用。
The widespread application of artificial intelligence has triggered concerns and discussions about algorithmic bias in academia. Targeting search engine autocomplete algorithms the approach of algorithm auditing, this study created large-scale user search queries and examined the algorithmic bias reflected by 47,011 autocomplete predictions collected from three leading search engines. It is found that autocomplete algorithm bias is consistent with the long-standing social discrimination in three social group attributes: gender, age, and region, reflecting the disadvantaged social status of women compared to men, middle-aged and elderly groups compared to the youth, and rural areas compared to urban areas. Both topics and search platforms significantly moderate the relationship between these three attributes and algorithmic bias. This study argues that social bias is reinforced by autocomplete algorithms through users’ interaction practices, while platforms act as a media vehicle for the reproduction of inequality under the influence of traffic commodity logic. Further discussion revolves around the mechanisms by which algorithmic bias interacts with digital inequality, and the role of different stakeholders in combating algorithmic bias.
搜索引擎 / 自动补全算法 / 算法偏见 / 数字不平等 / 算法审计
Search engines / autocomplete algorithm / algorithmic bias / digital inequality / algorithm auditing
| [1] |
陈昌凤, 师文(2022). 人脸分析算法审美观的规训与偏向:基于计算机视觉技术的智能价值观实证研究. 《国际新闻界》,(3),6-33.
|
| [2] |
陈杰, 郭晓欣, 钟世虎(2021). 户籍歧视对农村流动人口城市定居意愿的影响研究. 《社会科学战线》,(2),89-96.
|
| [3] |
陈力丹(2019). 人的记忆、搜索引擎与新闻传播学研究——搜索引擎批判. 《新闻界》,(1),52-59.
|
| [4] |
郭小平, 秦艺轩(2019). 解构智能传播的数据神话:算法偏见的成因与风险治理路径. 《现代传播(中国传媒大学学报)》,(9),19-24.
|
| [5] |
贺光烨, 吴晓刚(2015). 市场化、经济发展与中国城市中的性别收入不平等. 《社会学研究》,(1),140-165+245.
|
| [6] |
胡鞍钢, 周绍杰(2002). 新的全球贫富差距:日益扩大的“数字鸿沟”. 《中国社会科学》,(3),34-48+205.
|
| [7] |
雷霞(2021). 搜索引擎智能推荐的权力控制与人的能动性. 《现代传播(中国传媒大学学报)》,(5),145-151.
|
| [8] |
林曦, 郭苏建(2020). 算法不正义与大数据伦理. 《社会科学》,(8),3-22.
|
| [9] |
邱林川(2013). 《信息时代的世界工厂》. 桂林: 广西师范大学出版社.
|
| [10] |
冉晓醒, 胡宏伟(2022). 城乡差异、数字鸿沟与老年健康不平等. 《人口学刊》,(3),46-58.
|
| [11] |
师文, 陈昌凤(2022). 国内主流搜索引擎的算法审计研究. 《新闻大学》,(10),84-100+122-123.
|
| [12] |
吴帆(2008). 认知、态度和社会环境:老年歧视的多维解构. 《人口研究》,(4),57-65.
|
| [13] |
许向东, 王怡溪(2020). 智能传播中算法偏见的成因、影响与对策. 《国际新闻界》,(10),69-85.
|
| [14] |
张玉宏, 秦志光, 肖乐(2017). 大数据算法的歧视本质. 《自然辩证法研究》,(5),81-86.
|
| [15] |
赵万里, 谢榕(2020). 数字不平等与社会分层:信息沟通技术的社会不平等效应探析. 《科学与社会》,(1),32-45.
|
| [16] |
郑莉, 曾旭晖(2016). 社会分层与健康不平等的性别差异基于生命历程的纵向分析. 《社会》,(6),209-237.
|
| [17] |
张婍, 冯江平, 王二平(2009). 群际威胁的分类及其对群体偏见的影响. 《心理科学进展》,(2),473-480.
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
The movement toward what is often described as Web 2.0 is usually understood as a large-scale shift toward a participatory and collaborative version of the web, where users are able to get involved and create content. As things stand we have so far had little opportunity to explore how new forms of power play out in this context of apparent ‘empowerment’ and ‘democratization’. This article suggests that this is a pressing issue that requires urgent attention. To begin to open up this topic this article situates Web 2.0 in the context of the broader transformations that are occurring in new media by drawing on the work of a number of leading writers who, in various ways, consider the implications of software ‘sinking’ into and ‘sorting’ aspects of our everyday lives. The article begins with this broader literature before exploring in detail Scott Lash’s notion of ‘post-hegemonic power’ and more specifically his concept of ‘power through the algorithm’. The piece concludes by discussing how this relates to work on Web 2.0 and how this work might be developed in the future.
|
| [24] |
Recent studies have enhanced our understanding of digital divides by investigating outcomes of Internet use. We extend this research to analyse positive and negative outcomes of Internet use in the United Kingdom. We apply structural equation modelling to data from a large Internet survey to compare the social structuration of Internet benefits with harms. We find that highly educated users benefit most from using the web. Elderly individuals benefit more than younger ones. Next to demographic characteristics, technology attitudes are the strongest predictors of online benefits. The harms from using the Internet are structured differently, with educated users and those with high levels of privacy concerns being most susceptible to harm. This runs counter to intuitions based on prior digital divide research, where those at the margins should be most at risk. While previous research on digital inequality has only looked at benefits, the inclusion of harms draws a more differentiated picture.
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
|
| [34] |
|
| [35] |
|
| [36] |
|
| [37] |
|
| [38] |
|
| [39] |
|
| [40] |
|
| [41] |
|
| [42] |
|
| [43] |
|
| [44] |
|
| [45] |
|
| [46] |
Google. (2019). Autocomplete policies. Retrieved from https://support.google.com/websearch/answer/7368877?hl=en.
|
| [47] |
|
| [48] |
|
| [49] |
Recently, there has been an upsurge of attention focused on bias and its impact on specialized artificial intelligence (AI) applications. Allegations of racism and sexism have permeated the conversation as stories surface about search engines delivering job postings for well-paying technical jobs to men and not women, or providing arrest mugshots when keywords such as "black teenagers" are entered. Learning algorithms are evolving; they are often created from parsing through large datasets of online information while having truth labels bestowed on them by crowd-sourced masses. These specialized AI algorithms have been liberated from the minds of researchers and startups, and released onto the public. Yet intelligent though they may be, these algorithms maintain some of the same biases that permeate society. They find patterns within datasets that reflect implicit biases and, in so doing, emphasize and reinforce these biases as global truth. This paper describes specific examples of how bias has infused itself into current AI and robotic systems, and how it may affect the future design of such systems. More specifically, we draw attention to how bias may affect the functioning of (1) a robot peacekeeper, (2) a self-driving car, and (3) a medical robot. We conclude with an overview of measures that could be taken to mitigate or halt bias from permeating robotic technology.
|
| [50] |
|
| [51] |
|
| [52] |
|
| [53] |
As firms are moving towards data-driven decision making, they are facing an emerging problem, namely, algorithmic bias. Accordingly, algorithmic systems can yield socially-biased outcomes, thereby compounding inequalities in the workplace and in society. This paper reviews, summarises, and synthesises the current literature related to algorithmic bias and makes recommendations for future information systems research. Our literature analysis shows that most studies have conceptually discussed the ethical, legal, and design implications of algorithmic bias, whereas only a limited number have empirically examined them. Moreover, the mechanisms through which technology-driven biases translate into decisions and behaviours have been largely overlooked. Based on the reviewed papers and drawing on theories such as the stimulus-organism-response theory and organisational justice theory, we identify and explicate eight important theoretical concepts and develop a research model depicting the relations between those concepts. The model proposes that algorithmic bias can affect fairness perceptions and technology-related behaviours such as machine-generated recommendation acceptance, algorithm appreciation, and system adoption. The model also proposes that contextual dimensions (i.e., individual, task, technology, organisational, and environmental) can influence the perceptual and behavioural manifestations of algorithmic bias. These propositions highlight the significant gap in the literature and provide a roadmap for future studies.
|
| [54] |
|
| [55] |
We explore data from a field test of how an algorithm delivered ads promoting job opportunities in the science, technology, engineering and math fields. This ad was explicitly intended to be gender neutral in its delivery. Empirically, however, fewer women saw the ad than men. This happened because younger women are a prized demographic and are more expensive to show ads to. An algorithm that simply optimizes cost-effectiveness in ad delivery will deliver ads that were intended to be gender neutral in an apparently discriminatory way, because of crowding out. We show that this empirical regularity extends to other major digital platforms.
|
| [56] |
|
| [57] |
|
| [58] |
Software applications (apps) are now prevalent in the digital media environment. They are the site of significant sociocultural and economic transformations across many domains, from health and relationships to entertainment and everyday finance. As relatively closed technical systems, apps pose new methodological challenges for sociocultural digital media research. This article describes a method, grounded in a combination of science and technology studies with cultural studies, through which researchers can perform a critical analysis of a given app. The method involves establishing an app’s environment of expected use by identifying and describing its vision, operating model and modes of governance. It then deploys a walkthrough technique to systematically and forensically step through the various stages of app registration and entry, everyday use and discontinuation of use. The walkthrough method establishes a foundational corpus of data upon which can be built a more detailed analysis of an app’s intended purpose, embedded cultural meanings and implied ideal users and uses. The walkthrough also serves as a foundation for further user-centred research that can identify how users resist these arrangements and appropriate app technology for their own purposes.
|
| [59] |
|
| [60] |
|
| [61] |
|
| [62] |
|
| [63] |
|
| [64] |
|
| [65] |
|
| [66] |
|
| [67] |
|
| [68] |
|
| [69] |
Information providing and gathering increasingly involve technologies like search engines, which actively shape their epistemic surroundings. Yet, a satisfying account of the epistemic responsibilities associated with them does not exist. We analyze automatically generated search suggestions from the perspective of social epistemology to illustrate how epistemic responsibilities associated with a technology can be derived and assigned. Drawing on our previously developed theoretical framework that connects responsible epistemic behavior to practicability, we address two questions: first, given the different technological possibilities available to searchers, the search technology, and search providers, who should bear which responsibilities? Second, given the technology’s epistemically relevant features and potential harms, how should search terms be autocompleted? Our analysis reveals that epistemic responsibility lies mostly with search providers, which should eliminate three categories of autosuggestions: those that result from organized attacks, those that perpetuate damaging stereotypes, and those that associate negative characteristics with specific individuals.
|
| [70] |
|
| [71] |
|
| [72] |
Health systems rely on commercial prediction algorithms to identify and help patients with complex health needs. We show that a widely used algorithm, typical of this industry-wide approach and affecting millions of patients, exhibits significant racial bias: At a given risk score, Black patients are considerably sicker than White patients, as evidenced by signs of uncontrolled illnesses. Remedying this disparity would increase the percentage of Black patients receiving additional help from 17.7 to 46.5%. The bias arises because the algorithm predicts health care costs rather than illness, but unequal access to care means that we spend less money caring for Black patients than for White patients. Thus, despite health care cost appearing to be an effective proxy for health by some measures of predictive accuracy, large racial biases arise. We suggest that the choice of convenient, seemingly effective proxies for ground truth can be an important source of algorithmic bias in many contexts.Copyright © 2019 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works.
|
| [73] |
|
| [74] |
With the rise of computer algorithms in administrative decision-making, concerns are voiced about their lack of transparency and discretionary space for human decision-makers. However, calls to ‘keep humans in the loop’ may be moot points if we fail to understand how algorithms impact human decision-making and how algorithmic design impacts the practical possibilities for transparency and human discretion. Through a review of recent academic literature, three algorithmic design variables that determine the preconditions for human transparency and discretion and four main sources of variation in ‘human-algorithm interaction’ are identified. The article makes two contributions. First, the existing evidence is analysed and organized to demonstrate that, by working upon behavioural mechanisms of decision-making, the agency of algorithms extends beyond their computer code and can profoundly impact human behaviour and decision-making. Second, a research agenda for studying how computer algorithms affect administrative decision-making is proposed.
|
| [75] |
|
| [76] |
|
| [77] |
Media scholars have studied and critiqued search engines – and in particular the dominant commercial actor, Google – for over a decade. Several conceptual and methodological problems, such as a lack of technological transparency, have made a detailed analysis of concrete power relations and their effects difficult. This paper argues that a microeconomic approach can aid media scholars in examining the complex interactions that underpin the dynamics of information visibility unfolding around the Google search engine. Using the concept of a ‘three-sided market’, we characterize the business model built around google.com as the foundation of the company’s success. We then argue that the combination of search and advertising services, and in particular advertising network services, creates powerful incentives to orient the results page in self-serving ways, leading to fundamental conflicts of interest exacerbated by Google’s dominant position in both markets. Based on search engines’ mass media-like capacity to shape public discourse, we consider the identification of economic forces both as a prerequisite for a robust critique of the current situation and as a starting point for thinking about regulatory measures.
|
| [78] |
Algorithms, as constitutive elements of online platforms, are increasingly shaping everyday sociability. Developing suitable empirical approaches to render them accountable and to study their social power has become a prominent scholarly concern. This article proposes an approach to examine what an algorithm does, not only to move closer to understanding how it works, but also to investigate broader forms of agency involved. To do this, we examine YouTube’s search results ranking over time in the context of seven sociocultural issues. Through a combination of rank visualizations, computational change metrics and qualitative analysis, we study search ranking as the distributed accomplishment of ‘ranking cultures’. First, we identify three forms of ordering over time – stable, ‘newsy’ and mixed rank morphologies. Second, we observe that rankings cannot be easily linked back to popularity metrics, which highlights the role of platform features such as channel subscriptions in processes of visibility distribution. Third, we find that the contents appearing in the top 20 results are heavily influenced by both issue and platform vernaculars. YouTube-native content, which often thrives on controversy and dissent, systematically beats out mainstream actors in terms of exposure. We close by arguing that ranking cultures are embedded in the meshes of mutually constitutive agencies that frustrate our attempts at causal explanation and are better served by strategies of ‘descriptive assemblage’.
|
| [79] |
|
| [80] |
|
| [81] |
|
| [82] |
|
| [83] |
Google’s autocomplete function provides “predictors” to enable quick completion of intended search terms. The predictors reflect the search trends of a population; they capture societal beliefs and perceptions about a variety of subjects. This study explores the predictors provided by Google United States when searching for information about older men and women.
|
| [84] |
|
| [85] |
|
| [86] |
|
| [87] |
|
| [88] |
United Nations Europe and Central Asia(2018). United Nations Development Programme focus gender equality. Retrieved from http://www.eurasia.undp.org/content/rbec/en/home/gender-equality.html.
|
| [89] |
|
| [90] |
|
| [91] |
This article examines the extent to which economic, cultural, social, and personal types of engagement with the Internet result in a variety of economic, cultural, social, and personal outcomes. Data from a representative survey of the Dutch population are analyzed to test whether engagement with a certain type of activity is related to “collateral” benefits in different domains of activities, independent from the socioeconomic or sociocultural characteristics of the person. The results show that what people do online and the skills they have affect outcomes in other domains and that this is independent of the characteristics of the person. This means that policy and interventions could potentially overcome digital inequalities in outcomes through skills training and providing opportunities to engage online in a broad variety of ways. A semiologic rather than an economistic approach is more likely to be effective in thinking about and tackling digital inequalities.
|
| [92] |
Organizations often employ data-driven models to inform decisions that can have a significant impact on people's lives (e.g., university admissions, hiring). In order to protect people's privacy and prevent discrimination, these decision-makers may choose to delete or avoid collecting social category data, like sex and race. In this article, we argue that such censoring can exacerbate discrimination by making biases more difficult to detect. We begin by detailing how computerized decisions can lead to biases in the absence of social category data and in some contexts, may even sustain biases that arise by random chance. We then show how proactively using social category data can help illuminate and combat discriminatory practices, using cases from education and employment that lead to strategies for detecting and preventing discrimination. We conclude that discrimination can occur in any sociotechnical system in which someone decides to use an algorithmic process to inform decision-making, and we offer a set of broader implications for researchers and policymakers.
|
/
| 〈 |
|
〉 |
