How Google Autocomplete Algorithms about Conspiracy Theorists Mislead the Public

Introduction: Google Autocomplete Algorithms Despite recent attention to the impact of social media platforms on political discourse and public opinion, most people locate their news on search engines (Robertson et al.). When a user conducts a search, millions of outputs, in the form of videos, images, articles, and Websites are sorted to present the most relevant search predictions. Google, the most dominant search engine in the world, expanded its search index in 2009 to include the autocomplete function, which provides suggestions for query inputs (Dörr and Stephan). Google’s autocomplete function also allows users to “search smarter” by reducing typing time by 25 percent (Baker and Potts 189). Google’s complex algorithm is impacted upon by factors like search history, location, and keyword searches (Karapapa and Borghi), and there are policies to ensure the autocomplete function does not contain harmful content. In 2017, Google implemented a feedback tool to allow human evaluators to assess the quality of search results; however, the algorithm still provides misleading results that frame far-right actors as neutral. In this article, we use reverse engineering to understand the nature of these algorithms in relation to the descriptive outcome, to illustrate how autocomplete subtitles label conspiracists in three countries. According to Google, these “subtitles are generated automatically”, further stating that the “systems might determine that someone could be called an actor, director, or writer. Only one of these can appear as the subtitle” and that Google “cannot accept or create custom subtitles” (Google). We focused our attention on well-known conspiracy theorists because of their influence and audience outreach. In this article we argue that these subtitles are problematic because they can mislead the public and amplify extremist views. Google’s autocomplete feature is misleading because it does not highlight what is publicly known about these actors. The labels are neutral or positive but never negative, reflecting primary jobs and/or the actor’s preferred descriptions. This is harmful to the public because Google’s search rankings can influence a user’s knowledge and information preferences through the search engine manipulation effect (Epstein and Robertson). Users’ preferences and understanding of information can be manipulated based upon their trust in Google search results, thus allowing these labels to be widely accepted instead of providing a full picture of the harm their ideologies and belief cause. Algorithms That Mainstream Conspiracies Search engines establish order and visibility to Web pages that operationalise and stabilise meaning to particular queries (Gillespie). Google’s subtitles and blackbox operate as a complex algorithm for its search index and offer a mediated visibility to aspects of social and political life (Gillespie). Algorithms are designed to perform computational tasks through an operational sequence that computer systems must follow (Broussard), but they are also “invisible infrastructures” that Internet users consciously or unconsciously follow (Gran et al. 1779). The way algorithms rank, classify, sort, predict, and process data is political because it presents the world through a predetermined lens (Bucher 3) decided by proprietary knowledge – a “secret sauce” (O’Neil 29) – that is not disclosed to the general public (Christin). Technology titans, like Google, Facebook, and Amazon (Webb), rigorously protect and defend intellectual property for these algorithms, which are worth billions of dollars (O’Neil). As a result, algorithms are commonly defined as opaque, secret “black boxes” that conceal the decisions that are already made “behind corporate walls and layers of code” (Pasquale 899). The opacity of algorithms is related to layers of intentional secrecy, technical illiteracy, the size of algorithmic systems, and the ability of machine learning algorithms to evolve and become unintelligible to humans, even to those trained in programming languages (Christin 898-899). The opaque nature of algorithms alongside the perceived neutrality of algorithmic systems is problematic. Search engines are increasingly normalised and this leads to a socialisation where suppositions are made that “these artifacts are credible and provide accurate information that is fundamentally depoliticized and neutral” (Noble 25). Google’s autocomplete and PageRank algorithms exist outside of the veil of neutrality. In 2015, Google’s photos app, which uses machine learning techniques to help users collect, search, and categorise images, labelled two black people as ‘gorillas’ (O’Neil). Safiya Noble illustrates how media and technology are rooted in systems of white supremacy, and how these long-standing social biases surface in algorithms, illustrating how racial and gendered inequities embed into algorithmic systems. Google actively fixes algorithmic biases with band-aid-like solutions, which means the errors remain inevitable constituents within the algorithms. Rising levels of automation correspond to a rising level of errors, which can lead to confusion and misdirection of the algorithms that people use to manage their lives (O’Neil). As a result, software, code, machine learning algorithms, and facial/voice recognition technologies are scrutinised for producing and reproducing prejudices (Gray) and promoting conspiracies – often described as algorithmic bias (Bucher). Algorithmic bias occurs because algorithms are trained by historical data already embedded with social biases (O’Neil), and if that is not problematic enough, algorithms like Google’s search engine also learn and replicate the behaviours of Internet users (Benjamin 93), including conspiracy theorists and their followers. Technological errors, algorithmic bias, and increasing automation are further complicated by the fact that Google’s Internet service uses “2 billion lines of code” – a magnitude that is difficult to keep track of, including for “the programmers who designed the algorithm” (Christin 899). Understanding this level of code is not critical to understanding algorithmic logics, but we must be aware of the inscriptions such algorithms afford (Krasmann). As algorithms become more ubiquitous it is urgent to “demand that systems that hold algorithms accountable become ubiquitous as well” (O’Neil 231). This is particularly important because algorithms play a critical role in “providing the conditions for participation in public life”; however, the majority of the public has a modest to nonexistent awareness of algorithms (Gran et al. 1791). Given the heavy reliance of Internet users on Google’s search engine, it is necessary for research to provide a glimpse into the black boxes that people use to extract information especially when it comes to searching for information about conspiracy theorists. Our study fills a major gap in research as it examines a sub-category of Google’s autocomplete algorithm that has not been empirically explored before. Unlike the standard autocomplete feature that is primarily programmed according to popular searches, we examine the subtitle feature that operates as a fixed label for popular conspiracists within Google’s algorithm. Our initial foray into our research revealed that this is not only an issue with conspiracists, but also occurs with terrorists, extremists, and mass murderers. Method Using a reverse engineering approach (Bucher) from September to October 2021, we explored how Google’s autocomplete feature assigns subtitles to widely known conspiracists. The conspiracists were not geographically limited, and we searched for those who reside in the United States, Canada, United Kingdom, and various countries in Europe. Reverse engineering stems from Ashby’s canonical text on cybernetics, in which he argues that black boxes are not a problem; the problem or challenge is related to the way one can discern their contents. As Google’s algorithms are not disclosed to the general public (Christin), we use this method as an extraction tool to understand the nature of how these algorithms (Eilam) apply subtitles. To systematically document the search results, we took screenshots for every conspiracist we searched in an attempt to archive the Google autocomplete algorithm. By relying on previous literature, reports, and the figures’ public statements, we identified and searched Google for 37 Western-based and influencial conspiracy theorists. We initially experimented with other problematic figures, including terrorists, extremists, and mass murderers to see whether Google applied a subtitle or not. Additionally, we examined whether subtitles were positive, neutral, or negative, and compared this valence to personality descriptions for each figure. Using the standard procedures of content analysis (Krippendorff), we focus on the manifest or explicit meaning of text to inform subtitle valence in terms of their positive, negative, or neutral connotations. These manifest features refer to the “elements that are physically present and countable” (Gray and Densten 420) or what is known as the dictionary definitions of items. Using a manual query, we searched Google for subtitles ascribed to conspiracy theorists, and found the results were consistent across different countries. Searches were conducted on Firefox and Chrome and tested on an Android phone. Regardless of language input or the country location established by a Virtual Private Network (VPN), the search terms remained stable, regardless of who conducted the search. The conspiracy theorists in our dataset cover a wide range of conspiracies, including historical figures like Nesta Webster and John Robison, who were foundational in Illuminati lore, as well as contemporary conspiracists such as Marjorie Taylor Greene and Alex Jones. Each individual’s name was searched on Google with a VPN set to three countries. Results and D

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.