Most developed societies in the 21st century are largely dependent on social media. From political campaign management to everyday communication, these technologies revolutionized the way in which almost any aspect of a modern society operates. People feel more interconnected, impactful or simply view social media as a handy technology that simplifies life. Having more than 2 million users, Facebook has become the paragon of all social media and what the term embodies. Amidst all the practical capabilities of social media such as Facebook, the underlying drawbacks of their natural architectures is often overlooked. In a nutshell, Facebook and other similar social media use detailed algorithms to properly manage the arrangement of content on each individual profile. These algorithms make sure that the news feed on each Facebook profile is comprised of posts that support the viewpoints of that particular profile. Consequently, each Facebook user lives in a filter bubble of self-tailored content. On one hand, these algorithms exist to create highly personalized spaces where individuals can view various posts on topics that are of interest to them. This creates an efficient medium for entertainment, where users have the ability to constantly view new content without much effort. On the other hand, the existence of filter bubbles in spheres, such as politics, has detrimental effects for a society. A viewpoint cannot be objectively justified without disproving certain contradictions proposed by an opposition. Within a domain where such contradictions are virtually non-existent, similar perspectives have no choice but to promote each other. This is unfavourable as it undermines the whole concept of objectivity in argumentation. Filter bubbles create a realm of communication in which opinions are corroborated, and contradictions cease to materialize.
The fundamental algorithmic architecture of modern social media is bound to affect all users. The aim of this research is to divulge the extent to which each user is subject to a self-preserved sphere of opinions and interests. For this purpose, a fake Facebook profile of Mark Hooker, a 32-year-old male from Massachusetts USA, was created on the 7th of March 2019. Mark was set to become a testing dummy, with the sole purpose of discovering whether filter bubbles are prominent even in fresh accounts with no prior influences. Mark received fictitious personal attributes and interests, ranging from loving dogs to supporting anti-abortion, along with details regarding his education and various other background information. Since filter bubbles and echo chambers are self-made and preserved environments, Mark began sharing his interests with Facebook’s algorithms. From sharing dog images to commenting on anti-abortion and pro-Trump posts, Mark shaped his virtual persona. The aim of creating a specific set of interests and opinions for Mark was to see whether this influenced the content Mark received on his news feed. If highly controversial topics, such as abortion, were presented to Mark only from a single perspective, one that supports his current opinion, the existence of filter bubbles could be affirmed. Moreover, due to the vast amount of content even new accounts are subject to daily changes. The types of posts on Mark’s feed had to be analysed through an external program called FbTrex. This enabled precise observations regarding the types and opinions of most posts that Mark was subject to. With this data, the existence of filter bubbles and echo chambers could be researched more elaborately.
Over the course of three weeks, we liked and shared content that was appropriate to Mark’s predetermined characteristics. What we discovered after the analysis of his collected data was that, regarding the form of the posts (fig. 1), most of it (43,89% in total) was from groups; the rest concerned 29,95% of posts, 17,38% of videos, 7,49% of photo and 0,92% of events. Many posts which appeared on Mark’s profile (fig. 2) were dog-related, which seems like a logical consequence since Mark is a member of many dog-related public Facebook groups. Another significant category on Mark’s timeline was ‘pro-Trump’ posts, being 17,28% of all posts. This could also be expected, since Mark liked news pages, including Breitbart and Politico, which support Trump. Furthermore, as a construction engineer, Mark was very interested in technology and new inventions, and therefore, 11,64% of the posts were tech-related. Moreover, Mark further received posts were connected to anti-abortion, Christianity, funny memes and sports, which were conform his own interests. However, oddly enough, Mark also received posts that were not according to his own interests, including ‘beauty’ and ‘DIY’. These interests are more conform the creators’ interests. We believe that the appearance of these odd interests is due to the fact that Facebook has either caught our IP addresses or knows our personal profiles because of Facebook’s default account switcher, and therefore, sees Mark’s internet body as part of ours. Nevertheless, we have discovered that Mark did not even once receive liberal news or a viewpoint that fully contradicts his own; thus, very much seeming to live in a filter bubble.
Additionally, we managed to shift the content in Mark’s Facebook feed. For instance, at the start of Mark’s profile, the majority of the posts on his feed were connected to ‘anti-abortion’, ‘pro-Trump’ and ‘dogs’. His account was still very limited and Facebook’s personalisation algorithm could not “guess” much about Mark’s data body yet. However, over the next few days, we began to like different posts, such as technology. This resulted into Mark’s timeline being mainly tech-related; thus, his personalised feed had to be changed again and we decided to like and share various different content, including funny memes, cooking and sports. This indeed led to a news feed with noticeably more distinct topics. We were thus able to control Mark’s own data body in the first few weeks of the existence of his account. This could be explained by the fact that Mark’s account was still in a developing phase, and Facebook’s algorithm did not fully know his data body yet. His personalisation could therefore be easily changed in the beginning and new information was added whenever he left another digital footprint, considering that Facebook simply did not have enough data to “guess” Mark’s interests yet.
However, we need to note that the results presented above may not be completely accurate due to certain limitations which we encountered. First, we need to acknowledge that his profile is completely new, therefore, this data is significantly smaller than years old Facebook profile. Furthermore, the research conducted on his profile was too short – only 3 weeks – to fully analyse his account, which is not enough to effectively prove the existence of a filter bubble. Second, the data collected with the FbTrex tool is only public, meaning that all the private posts from his feed were not collected and analysed. This is important to note, because Mark’s results significantly differ from the reality. For example, the themes of ‘anti-abortion’ and ‘christian’ were the most frequent on his feed, due to his active participation in private Christian groups. However, the graphs show that only a relatively small amount of posts and sources were related to this, and in contrast, even tell us that ‘dog’ and ‘pro-Trump’ posts were most apparent on his feed (fig. 2; fig. 3). FbTrex is therefore not a sufficient enough tool to present the reality of Mark’s data body.
In order to paint a completed picture of Facebook’s personalisation algorithm, comparison is inevitable. Since Mark’s profile was created only for the purpose of this project, it didn’t quite show a reflection of the algorithm over time. Therefore, the profile of Lucia Holaskova was chosen to demonstrate the long term consequences of the filter bubble. However, it failed to help as there were more negatives encountered rather than positives. Firstly, 86,03% of the data was error, therefore leaving only 13,97% for observation. Regarding data on the form of the posts, it was made up of 9,56% groups, 2,21% videos and 2,21% photos (fig. 4). Concerning content categories, the most prevalent was ‘friends’, which came up to 23,53%, followed by ‘funny’ coming in at 21,32%, with the third most prevalent category being ‘groups’ at 15,44% (fig. 5). However, these categories are very vague, due to the large quantity of limitations and errors which happened during the data extraction phase. Yet, even though the data was ineffective and hazy, it did confirm the presence of a filter bubble on Lucia’s profile. Judging by the percentages of the content categories, it goes perfectly in line with her interests. Lucia uses her Facebook mainly to communicate with her friends, which is the most prevalent category. Moreover, she has many meme pages in her likes (fig. 6), which also goes in line with the fact that ‘funny’ is the second most prevalent category. The only discrepancy is the percentage of politics related content, which showed up at 1,47%. Even though Lucia is not very active in showcasing her political values and opinions, she has a lot of liberal news sources in her interests, and checks them regularly.
Many limitations were encountered, most notably errors in data – there were no links to posts in the .cvc file which was generated from FbTrex, and the form of said posts was unclear. Therefore, there was no way of finding out what exactly the posts were about, thus creating vague categories in data such as ‘friends’, with unknown content (fig. 5). Another limitation that was encountered in the data was the presence of unfamiliar people which were not in Lucia’s friends, nor were they people that could be considered famous (fig. 6). Accordingly, it was concluded that these people were probably from groups that Lucia is in, thus appearing on her profile. Yet, this fact is still more or less debatable as the source of these displaced people cannot be traced. This resulted in the formation of another vague category – ‘groups’. Moreover, due to error the ‘groups’ category and the ‘events’ category are not present in the forms graph at all (fig. 4). These were the most dominant limitations that could be found, resulting in a lot of individual sources that could not be placed in any category or form, since they had no origin, so technically they were completely useless for this research.
Comparison between the two
When looking at the two separate accounts side-by-side there are a few points of interest to acknowledge. For one, generally speaking, the varied categories of posts found on Mark’s and Lucia’s feeds were very representative of their interests. Mark’s feed was particularly representative of our group’s decisions when it came to creating his profile, and there was only a small amount of discrepancy between the interests and attributes given to him versus what was on his feed. Another important thing to address is the fact that Mark lacked a vital component of what makes social networks function; that is, Mark lacked sociality. Roughly 34% of the posts on Lucia’s feed consisted of events and/or from Lucia’s friends. Without any friends, or specifically, friends with varied interests, beliefs, and values, it is hard to truly say whether or not there was a true filter bubble present on Mark’s profile. Since our group only liked and shared content we decided on when we created the profile there was no possibility of an opposition regarding his own beliefs as there would be on most people’s Facebook accounts. Due to this, it is hard to make any concrete statements regarding Mark’s feed and a filter bubble.
Furthermore, FbTrex proved to be the source of a few different limitations. First and foremost, due to the limited amount of time we actually worked with FbTrex, it is difficult to make sense of a relatively short lived dataset. Data takes time to curate and explore fully, so with a limited time frame we were sceptical to jump to conclusions. Most notably, however, and as previously stated earlier, when viewing the data results from the real profile of Lucia, roughly 86% of it was unusable due to FbTrex not being able to generate a huge quantity of links to their corresponding posts, thus making true comparison difficult and rather unattainable. When working with data, especially in a project such as this, precision and clarity are vital for creating a successful and viable project that has a lasting impact.
Algorithms have become ubiquitous on the Internet, especially on social media. Every major platform relies on its data analytics in order to provide the most accurate recommendation system to their audience. Moreover, these recommendation systems have more influence on users’ choice than anything nowadays. However, all of the positive outcomes from algorithms are collected in one negative bubble: the personalized filter bubble. Due to these personalised information bubbles, a user’s ability to critically evaluate information is damaged, which is a major threat to the political sphere as people generally do not seek information once they have already obtained it. Therefore, if users receive their daily news from Facebook, which shows them information catered to their viewpoint, they will not go looking for it elsewhere, thus leading to a gap in information. Yet, most people view their Facebook feeds as a window to the world, painting a bigger picture of what is happening around them. However, a “bigger” picture does not mean a full one. Many do not realize that they have a very limited set of information sources due to filter bubbles.
Mark’s account, even only three weeks old, with no friends and connections, and only groups and official pages, still managed to get trapped inside a filter bubble. However, during the three weeks we managed to tweak the algorithms a little, by adding to his interests step-by-step, day-by-day. This would not have been possible if he already had a digital footprint, because there is no way out of the echo chambers. Furthermore, Facebook’s algorithm seems to be smarter than anticipated, because it found a way around the fake persona Mark. From giving friend suggestions of the creators themselves to posts on Mark’s timeline that go in line with the creators’ interests. Facebook knows everything and can easily identify if a profile is indeed real or fake.
Additionally, we were able to appreciate the importance of Blue Feed, Red Feed [read our blog about it here], because through Mark’s profile we received one-sided news. There was not a single controversial opinion or article. Facebook blinds people entirely, making their feed like a tunnel with no other view but the road ahead. In contrast, Blue Feed, Red Feed provides both possible political perspectives, giving the audience the chance to actually think for themselves and choose what to agree and disagree with.
Overall, we reached the conclusion that such research should be conducted on multiple accounts who have established their online engagement and have completely different interests. Only then will it be possible to conduct a completely objective experiment and analyse real results. Fake accounts have nothing else for the algorithms but the IP addresses of the creators, giving subjective results and influencing the filter bubble. By being created anew their data bodies are easily manipulated and changed. Therefore, we did experience a small but adjustable filter bubble on Mark’s profile, however, we cannot make concrete statements about this filter bubble due to the many limitations of a fake profile. Nevertheless, we do believe that filter bubbles do not only shape the digital world; they shape society’s opinions. There is no escaping the echo chamber, the only option is for all of us to stop and think for a minute before sharing content online.
Written by: Amber Kouwen (11674105), Nanda Mohamed (11845910), Lucia Holaskova (11742321), Ivana Sramkova (11826711), Desislava Slavova (11832517) and Aidan Fahle (11788178)
Word count: 2590
Virtualisation of data and managing datajlab as an end-product: Amber
Explanation Mark’s graphs and limitations: Desislava and Amber
Explanation Lucia’s graphs and limitations: Lucia
Comparison Mark’s and Lucia’s results and limitations: Aidan
Bakshy, Eytan, Solomon Messing, and Lada Adamic. 2015. ‘Replication Data for: Exposure to Ideologically Diverse News and Opinion on Facebook’. https://doi.org/10.7910/DVN/LDJ7MS.
Bilton, Ricardo. n.d. “The Wall Street Journal’s New Tool Gives a Side-by-Side Look at the Facebook Political News Filter Bubble.” Nieman Lab (blog). Accessed February 20, 2019. http://www.niemanlab.org/2016/05/the-wall-street-journals-new-tool-gives-a-side-by-side-look-at-the-facebook-political-news-filter-bubble/.
Graff, Ryan. 2016. ‘How WSJ Used Data and Design to Show Americans Their Polarized Politics and Media’. Northwestern University Knight Lab. 21 June 2016. https://knightlab.northwestern.edu/2016/06/21/how-wsj-used-data-and-design-to-show-americans-their-polarized-politics-and-media/.
Keegan, Jon. 2016. “Blue feed, red feed.” The Wall Street Journal 18 (2016).
Nagulendra, Sayooran, and Julita Vassileva. 2014. “Understanding and controlling the filter bubble through interactive visualization: a user study.” In Proceedings of the 25th ACM conference on Hypertext and social media, pp. 107-115. ACM.
Nguyen, Tien T., Pik-Mai Hui, F. Maxwell Harper, Loren Terveen, and Joseph A. Konstan. 2014. “Exploring the filter bubble: the effect of using recommender systems on content diversity.” In Proceedings of the 23rd international conference on World wide web, pp. 677-686. ACM.
Resnick, Paul, R. Kelly Garrett, Travis Kriplean, Sean A. Munson, and Natalie Jomini Stroud. 2013. “Bursting your (filter) bubble: strategies for promoting diverse exposure.” In Proceedings of the 2013 conference on Computer supported cooperative work companion, pp. 95-100. ACM.
Tufekci, Zeynep. 2015. “Algorithmic harms beyond Facebook and Google: Emergent challenges of computational agency.” Colo. Tech. LJ 13 (2015): 203.