Understanding User Beliefs About Algorithmic Curation in the Facebook News Feed

“As Rader and Gray (2015) point out, the feedback-loop characteristics of these systems make user beliefs an important component in shaping the overall system behaviour.”

“First, our research contributes to emerging studies that investigate people’s mental models and folk theories of algorithms (Bucher, 2016; Eslami et al., 2015; Rader and Gray, 2015), and social studies on automation, artificial intelligence and machine learning more broadly.”

“In order to find out how an intelligent system should explain its behavior for non-experts, research on machine learning explanations has studied how users interpret intelligent agents [47, 164] and algorithms [171] to find out what users expect from machine explanations.”

“…influence every facet of our lives, including the quality of healthcare and education we receive [16, 31, 32, 49], which news or social media posts we see [4, 18, 84], who receives a job [46, 82], who is released from jail [8, 23], and who is subjected to increased policing [65, 71, 94].”

“ent [33], judicial sentencing and bail [3, 14, 35, 37], and hiring [43]. In addition, these models influence decisions about what news people see on social media platforms (e.g., Twitter and Facebook, [2, 10, 48, 61]), provide recommendations for what movies to watch (e.g., Netflix, [7]) or what music to listen to on on…”

“Yet many users are unaware of this algorithm [7] or develop personal beliefs about it [22].”

“These theories are often incomplete, are generated out of each user’s personal experiences, and serve to help them both to filter information and to make decisions about how they interact with the world (Gelman and Legare, 2011; Jones et al., 2011; Rader and Gray, 2015).”

“%) in their sample believe they do not see every post their friends create, but also that ‘ users vary widely in the degree to which they perceive and understand the behaviour of content fi ltering algorithms ’ (Rader & Gray, 2015, p. 181).”

“Users may become aware of an algorithm through repeated experiences with a system [46].”

“Previous work has used various methods to study users’ algorithmic literacy, from survey questionnaires (Cotter & Reisdorf, 2020; Gran et al., 2020; Rader & Gray, 2015) and experiments (Eslami et al., 2015) to focus groups (Siles et al., 2019) and in-depth interviews, often aided by card sorts…”

“For a work dealing with user beliefs considering how the algorithm applied behind the ‘Newsfeed’ works, please see Rader & Gray [72].”

“Previous work investigated users’ awareness of curation algorithms, paths toward awareness and effects of awareness [9, 29], yet much remains unknown about which folk theories exist to make sense of these algorithms, the connection between design and the development of folk theory, and the utility…”

“Rader and Gray argue that “because user behavior is both input for algorithms and constrained by them, these patterns of belief may have tangible consequences for the system as a whole” [64].”

“For an overview of philosophical theories of intention, see Setiya (2018). periods of time (Lanier, 2018; Rader & Gray, 2015).11 The algorithms do this by monitoring the amount of time particular kinds of content keep the particular user engaged with the platform, and they use that data to…”

“, [5, 10, 11, 29]) have focused on operational theories, defining algorithmic awareness as starting at what we call operational folk theories.”

“Recent studies indicate that most people are unaware of filters in social media (Eslami et al. 2015; Rader and Gray 2015).”

“For example, studies have found that some participants are not aware of algorithmic curation in the Facebook News Feed [31,69] or the gathering of online behavioral data and its use for inferencing [86], or underestimate the prevalence and scale of data gathering and its use in practical…”

“People adjust their behavior once they are aware an algorithm is in play, including closely comparing who is shown in their feeds to their prior expectations, among other tactics [7, 11, 12, 42].”

“, investigated the users’ interpretations of the AI/ML systems and their algorithms [82] [83].”

“Other studies have focused on individual experiences with algorithmic systems through examining the information sharing practices around visibility on Instagram [29], individual’s awareness of news feed personalization [47, 106, 107], and the ways algorithms can influence self-presentation on various dating applicatio…”

“As mentioned, HCI studies have fruitfully applied folk theory frameworks to understand perceptions of algorithms (Eslami et al., 2016; Rader and Gray 2015, also French and Hancock 2017), underling the value of user perspectives and producing useful typologies.”

“…the possibility of context collapse [8, 41, 58] and makes it hard for users to understand which audiences they are “performing” their self for [18, 48], exacerbating the existing problem of inaccurately imagined audiences, where the user’s own idea of who will likely see content diverges from the…”

“Research on user theories about recommendation algorithms covers, for example, Facebook news feed curation [21, 22, 44], or Spotify music suggestions [49], and how user assumptions influence user engagement [15].”

“Academic work has often con fl ated AI knowledge with opinions about how algorithms operate in speci fi c settings like Facebook News Feed and TikTok (e.g., Rader and Gray 2015; Siles, Valerio-Alfaro, and Meléndez-Moran 2022), or measured knowledge in terms of self-reported assessments of competency.”

“…point of view, most of these studies were qualitative analyses of user perceptions of algorithmic curation embedded in these particular services, partly through probing perceptions of algorithms at work (e.g., Bucher, 2017; Eslami et al., 2015; Hamilton et al., 2014; Rader & Gray, 2015).”

“Consistent with previous studies (Eslami et al., 2015; Rader & Gray, 2015), many participants did not understand that Facebook’s information display is highly dictated by opaque algorithms that are not well understood.”

“Most online platforms today curate content generated by their users in different ways—presenting content in an ordered form to increase user engagement (e.g., Facebook NewsFeed [56]), promoting sponsored content (e.g., ads [44]), and moderating (or screening) undesirable content [33, 41, 63].”

“For example, recent work studied users’ of awareness of opaque social media feed curation algorithms [14] and users’ folk theories on how algorithms work [8, 12, 16, 36].”

“As work on more general populations [25, 26, 31, 32, 88] focuses only on initial folk theory formation, we do not yet have a detailed understanding of the everyday process of repeated folk theorization and adaptation that casual, non-professional users employ to maintain their ability to…”

“…and Gray discuss how the algorithms employed in e.g. Facebook create a feedback loop, in which the user’s behavior (shares or likes) define what he/she also consumes, creating a loop in which the algorithm always presents items that the user in turn will almost certainly interact with [23:173].”

“Algorithms display subsets of a corpus of information for consumption by prioritizing, classifying, associating, and filtering information (Eslami et al., 2015; Rader & Gray, 2015).”

“We note that interpretation of the term is further confounded by the “AI effect” (the phenomenon that once AI successfully solves a problem and the solution becomes commonplace, it is no longer considered to be AI) [40], as well as lack of awareness of algorithmic processing in common systems [22, 49, 58].”

“Further, because algorithms change as a result of the data input they receive in a kind of feedback loop [50, 55], Bucher [9] even argues that these perceptions can then influence the nature of the algorithms themselves.”

“Early work on folk theorization suggested that baseline algorithmic awareness was low, and that user folk theories were mostly abstract, focused on what a system does and why, not how [22, 23, 43] — what I have previously referred to as functional folk theorization [13].”

“Opaque algorithmic curation that takes humans out of the loop aims to maximize consumption and interactions, possibly causing viral appeal to take precedence over truthfulness (Rader & Gray, 2015).”

“Even when users are intimately familiar with a system, they are often unaware that algorithms filter their access to information [14] and users hold beliefs about algorithms, which, true or not, influence how they use systems [39].”

“…obscured as trade secrets, computational imaginaries are necessary to interpret algorithmically based systems ranging from the Facebook news feed (Eslami et al., 2015; Rader and Gray, 2015), to Google Play’s app recommender system (Ananny, 2011), to Twitter’s trending topics (Gillespie, 2011).”

“In contrast, explanation generation methods–used in this paper– aim to explain models that are not directly human-understandable (e.g., deep neural networks).”

“As Rader and Gray assert, mismatch between user’s and algorithm’s goals can create tensions among users [35].”

“Awareness has also been addressed through somewhat more focused questions, leading users to speculate about why content is presented to them, but still usually without explicit reference to algorithms (e.g. Koenig, 2020; Powers, 2017; Rader and Gray, 2015).”

“Rader and Gray [28] investigated user understanding of algorithmic curation in Facebook’s News Feed, and whether they believed that all their friends’ posts appeared on their feed.”

“Previous research (DeVito et al., 2017; Eslami et al., 2015; Rader & Gray, 2015) discovered that individuals’ understanding of AI technologies or systems often contains misperceptions and assumptions because general end users do not have inner knowledge about how algorithms work backstage.”

“…studies have investigated users’ awareness of and interaction with algorithms — including news feed curation algorithms (Eslami et al. 2015; 2016; Rader and Gray 2015) and ridesharing management algorithms (Lee et al. 2015) — but how users perceive and manage the bias that an algorithm brings to…”

“Rader and Gray [102] studied user beliefs about news feed curation on Facebook, the algorithmic behaviors that users notice and the behaviors to which they respond.”

“Confronted with an opaque system, humans cannot build appropriate trust; they often either blindly follow the system’s decisions and recommendations or do not use the sys-tem (Herse et al., 2018; Rader & Gray, 2015).”

“Folk theories can help us understand how users perceive the systems that they are interacting with [12, 20, 39], including uncovering inaccurate knowledge that users might have about a system in order to intervene onmisguided or even harmful behavior on the platform [13, 20, 39].”

“The processes through which users become aware of and react to algorithms has been a topic of recent research interest [33–35, 37, 106].”

“In attempting to chart the sociotechnical gap in XAI, this paper thus builds on and extends the CSCW community’s long-standing tradition of designing and studying explainable computing systems [52, 69, 83, 93, 101, 122, 135].”

“” These stories are then characterized into dense feature vectors by feature extractors and learned models [17,26–28].”

“…influence every facet of our lives, including the quality of healthcare and education we receive [16, 31, 32, 49], which news or social media posts we see [4, 18, 84], who receives a job [46, 82], who is released from jail [8, 23], and who is subjected to increased policing [65, 71, 94].”

“Feed2 aggregates various leaf microservices’ responses into discrete “stories” that are then characterized into dense feature vectors by feature extractors and learned models [26, 51, 101, 132].”

“, 2017) and ‘intuitive causal explanatory theories’ (Rader and Gray, 2015) about how they work.”

“…and social networking sites; users are often unaware that news media and informative content consumed on social media news feeds or timelines are uniquely customized to each user’s preferences (Flynn, Nyhan, & Reifler, 2017; Qiu, Oliveira, Shirazi, Flammini, & Menczer, 2017; Rader & Gray, 2015).”

“For example, both Rader and Gray (2015) and Eslami et al. (2016) investigated whether or not users knew that Facebook did not display all available posts in their news feeds.”

“A similar study was conducted by Rader and Gray [37], who classified a wide range of user beliefs about Facebook’s news feed algorithm.”

“Many researchers have explored sensemaking around social media and news feeds [7, 12, 18, 20, 48].”

“In the current era of big data (including big social network data), various social networking sites or services—such as Facebook, Google+, LinkedIn, Twitter, and Weibo [16,17]—are commonly in use.”

“A small number of studies have investigated users’ awareness of algorithms in general, in domains other than brand communication (Bucher 2017; eslami et al. 2015; gran, Booth, and Bucher 2021; Klawitter and Hargittai 2018; Rader and gray 2015).”

“Previous research has examined expectations for particular features, including the Facebook Like button (Scissors et al., 2016), algorithmic curation (Bucher, 2017; Eslami et al., 2015; Rader and Gray, 2015), and design changes (Eslami et al., 2016).”

“An algorithm is even deciding on the posts and news people will see on social media [40, 44].”

“…algorithms for bias, but what is missing from these investigations is frequently the human component: how these algorithms manifest themselves in people’s experiences of their devices is often assumed rather than demonstrated (Eslami et al. 2015; Hamilton et al. 2014; Rader and Gray 2015).”

“When algorithmic systems are opaque, end-users may develop folk theories to explain how the underlying algorithms operate and the outcomes and consequences of the systems [18,26,79].”

“…with some pointing to low awareness of the existence of algorithms and limited understanding of how they filter information (Eslami et al., 2015; Gran et al., 2021; Powers, 2017; Swart, 2021; Zarouali et al., 2021), while others detect a high awareness thereof (Proferes, 2017; Rader & Gray, 2015).”

“People do not always understand the output of algorithms [29, 72] despite having many opinions about what should be done if algorithms are biased [31], imperfect [30], or discriminatory [5, 70].”

“Inspired by a prior study [47], before our respondents entered the survey we also emphasized that detailed responses would support our analysis and improve MTurk workers’ future experience.”

“In all cases, researchers have emphasized that users’ intuitive theories about how social media feeds are composed affect their behaviors on those platforms [25, 30, 69].”

“For instance, Rader and Gray (2015) conducted a qualitative inquiry of people’s beliefs surrounding Facebook’s algorithm, which uncovered various ways they adapted their behavior in response to perceptions of algorithmic curation in their newsfeeds.”

“media content filtering, feedback can undermine the diversity of users’ interaction scopes [55].”

“In addition to critical scholarship and commentary on extractive AI and its hidden labour, we orient this article towards recent research that examines the sensemaking that occurs when humans encounter algorithms within the mundane interactions on digital platforms (Rader and Gray, 2015).”

“Social media algorithms rely on user behavior to curate platform feeds, but often there can be mismatch between user goals and the algorithm’s functioning [108].”

“In particular, participation in platforms that employ designs that affect information visibility (e.g., Facebook’s News Feed) have consequences for people based on how much they understand these systems to work [18,44].”

“However, while regular users invent folk stories as a way of explaining Facebook’s mechanism (Rader and Gray, 2015), editors – who also devise theories – can put these theories to the test, much like a ‘naïve scientist’.”

“…survey methods, information flow solipsism Critical algorithm studies scholars have argued that the algorithms of spaces such as Twitter and Facebook are opaque, often misunderstood by users and thus carry the potential to negatively impact users’ experiences (Bucher, 2016; Rader & Gray, 2015).”

“For one, users may variously seek diversity [12], be variously responsive to recommendation [1], use it for various purposes [13] or have various expectations about it [14] — in these respects, the “average user” does not really exist.”

“While the concept has significant overlap with “folk theories” and has received considerable attention, in this study, we adopt the terminology of folk theories, developed and adopted in particular within CSCW and CHI [45,46,55,89,116,142].”