ABSTRACT
In 2020, the saliency-based image cropping tool deployed by Twitter to generate image previews was suspected of carrying a racial bias: Twitter users complained that Black people were systematically cropped out and, thus, made invisible by the cropping tool. As a response, Twitter conducted bias analyses, concluded that the cropping tool was indeed biased, and subsequently removed it. Soon after, Twitter hosted the first "algorithmic bias bounty challenge", inviting the general public to detect algorithmic harm in the cropping tool.
Twitter’s image cropping algorithm is a fascinating case study for exploring the push-and-pull dynamics of power relations between, firstly, algorithmic knowledge production inherent in machine learning systems, secondly, the bias discourse as resistance, and, thirdly, ensuing corporate responses as stabilization measures towards said resistance. In order to account for this three-part narrative of the case study, this paper is structured along the examination of the following three questions: (1) How is algorithmic, and especially, data-based knowledge production entrenched in power relations? (2) In what way does the discourse around bias serve as a vehicle for resistance against said power? Why and in what way is it effective? (3) How did Twitter as a company stabilize its position within and in relation to the bias discourse?
This paper explores these questions along the following steps: Section 2 lays out the interdisciplinary theoretical perspective of the analysis, combining, firstly, a mathematical-epistemic perspective that examines the mathematics underlying both machine learning systems and bias analyses with, secondly, Foucauldian concepts that make it possible to view mathematical tools as articulations of power relations. The subsequent three sections engage with the three questions posed above: Section 3, Power, is concerned with the first question, and it focuses on the algorithmic knowledge production in relation to Twitter’s cropping tool and its mathematical-epistemic foundations. Section 4, Resistance, addresses the second question, and it examines three bias analyses of the cropping tool, as well as their epistemic limitations, and it continues by conceptualizing the bias discourse in academic scholarship and activism as resistance to power. Section 5, Stabilization, engages with the third question, discussing Twitter’s response to the bias accusations and the way in which the company was able to effectively stabilize its position – rendering the bias discourse a vehicle for counter-resistance, too. This paper will be published in the open access volume Algorithmic Regimes: Methods, Interactions, and Politics (Amsterdam University Press, forthcoming), as well as on SSRN as a preprint.
Index Terms
- Power and Resistance in the Twitter Bias Discourse
Recommendations
Uncovering Gender Bias within Journalist-Politician Interaction in Indian Twitter
WEBSCI '24: Proceedings of the 16th ACM Web Science ConferenceIn this study we investigate gender bias within journalist politician interactions in India—a country in global south. Previous work identified gender bias in online political discourse, but they focused on global north and politician-general public ...
A sentiment analysis of audiences on twitter: who is the positive or negative audience of popular twitterers?
ICHIT'11: Proceedings of the 5th international conference on Convergence and hybrid information technologyMicroblogging is a new informal communication medium of blogging that differs from a traditional blog in which content is much shorter. Microbloggers post about topics that describe their current status. Twitter is a popular microblogging service and ...
Information resonance on Twitter: watching Iran
SOMA '10: Proceedings of the First Workshop on Social Media AnalyticsTwitter has undoubtedly caught the attention of both the general public, and academia as a microblogging service worthy of study and attention. Twitter has several features that sets it apart from other social media/networking sites, including its 140 ...
Comments