Introduction

1. A critical look at the concept of filter bubbles

2The growing use of online personalisation algorithms raises many concerns about the ability of users to maintain a diversified information system. The notion of the “filter bubble” [Pariser, 2012] is often used to describe a state of intellectual isolation resulting from algorithmic mechanisms which remove from our information horizon any content which runs counter to our ideas or preferences. In reaction to this premise, it is often advised to adopt random behaviour which would disrupt the algorithm, or even to avoid platforms which offer such services [Bozdag and van den Hoven, 2015]. While this theory has the merit of drawing widespread attention to the issues and effects of personalisation algorithms, it is nevertheless misleading in several respects. It could lead us to believe that automated recommendations have a uniform effect on information practices. However, the results of academic research on the subject are sometimes contradictory, and paint a complex picture which depends on both the platform studied and the field of application of the algorithm.

3While it has long been thought that online search results are closely adapted to a user's profile, recent work suggests that there is a significant overlap of results between multiple searches from different users [Haim et al., 2018; Krafft et al., 2018; Puschmann, 2019]. As regards news recommendations, for example, according to a study led by Facebook [Bakshy et al., 2015], while the platform's algorithm does indeed cause a decrease in exposure to contrary opinions, a user's choice of social connections has a greater influence on the personalisation of the news feed. More experimental studies have also found that recommendation algorithms have a negative effect on the diversity of opinions encountered. Claussen et al. [2019] observe that personalising the homepage of a news website decreases the diversity of content accessed, while Dylko et al. [2017] find that news personalisation limits exposure to dissonant information. However, studies in other fields of application qualify these observations. Aiello and Barbieri [2017] examined contact recommendations on Flickr and Tumblr. They observed that automated recommendations increase the diversity of the social networks of users who receive them. According to the authors of the study, the system exposes users to a greater variety of contacts than if they had explored blogs or searched directly on the website.

4In the case of streaming services, the study by Roth et al. [2020] shows that video recommendations on YouTube steer users towards uniform video groups which favour “popular” videos, i.e. those viewed by thousands of users. When they studied the MovieLens website, which recommends films based on scores awarded by the user community, Nguyen et al. [2014] found that users who are exposed to a recommendation algorithm access less and less diverse content over time. However, this decline is less significant than it is for other users of the platform who are not exposed to recommendations. As for Spotify, music consumption guided by the use of recommendation systems is associated with a decrease in the diversity of songs listened to individually in the long term [Anderson et al., 2020]. However, Holtz et al. [2020] and Datta et al. [2017] observe a positive effect of automated recommendations on the diversity of listening practices in general. Holtz et al. [2020] note that even if a wider variety of content is accessed on a platform-wide scale, each individual tends to focus on more specific interests. Finally, the work of Beuscart et al. [2019] on Spotify indicates that automatic recommendations help to guide users towards less popular songs and are part of the exploratory approach of enlightened users.

5It should be pointed out that the studies cited here are very unequal in the scope of their results, the nature of their methodology and the size of their samples, which makes it difficult to compare their results. However, these works tend to indicate that the effect of recommendation algorithms on information usage cannot be reduced to a gradual isolation in a uniform space. While the integration of recommendation systems has a significant influence on the information horizon of individuals, it is not as uniform as the filter bubble theory might suggest. The nature of the recommended elements, the technology used, the interface proposed to the user and the context in which the mechanism is used are all likely to influence viewing or listening practices [Dalhgren, 2021]. Furthermore, beyond these technical considerations, online content consumption practices are also more varied than expected. Fletcher and Nielsen [2017] observe that in the context of an information ecosystem with a diversified media offering, most (adult) users combine content from niche media with content from media aimed at a wider population. These results corroborate the conclusions of Dubois and Blank [2018], who point out that users of social networks usually combine several sources of information (online searches, access to so-called “legacy” media, social media, etc.). Garrett et al. [2013] and Newman et al. [2020] also find that generally, even those who use partisan news websites also receive their news from legacy media. It is therefore quite rare for users of online platforms to find themselves in a perfectly closed digital environment.

6In reality, theories such as that of the filter bubble tend to mask the complexity of the uses which develop in relation to these mechanisms. The fact that this complexity exists does not mean that the use of recommendation algorithms is without consequence: instead, the analysis of their effects must be made with the contexts of use borne in mind, in order to go beyond a deterministic representation of users who are left without agency. While an abundant literature is beginning to emerge around these multiple contexts of use, several individual markers are still overlooked and deserve our attention in order to further our understanding of media practices. A relevant line of research to meet this objective would be to differentiate between these practices according to the age of users. The results presented below examine the specific practices of young users in Brussels.

2. How young people in Brussels express their algorithmic imaginaries

7We postulate that the way in which young people in Brussels perceive recommendation mechanisms and their effects plays a part in how they use recommendation systems, obtain information and form their opinions. In order for this perception and what it conveys to be examined, comments on the states of mind and feelings which algorithms help to generate need to be gathered [Bucher, 2017: 32]. This approach through experience of the effects of algorithms may be supplemented by a study of the practices and tactics which young people put in place in order to function in relation to the algorithms which exist on the platforms they use [Wiard, Lits and Dufrasne, 2022].

8The first stage of this exploration of the perception of algorithms by young people in Brussels involved 27 exploratory interviews with secondary school teachers (n=9), young people in Brussels aged 15 to 24 (n=9) and experts in the world of media and media education (n=9) in the Brussels-Capital Region. As young people's information practices cannot be dissociated from the social networks which enable them to make sense of information [Claes et al., 2021], these interviews, first of all, enabled us to identify adults' ideas of the relations which young people have with algorithms. The material gathered from teachers revealed that they have a very pessimistic view, with the majority of them describing students as not giving much thought to the presence of algorithms in their daily information practices. Teachers conveyed the image of a harmful and dangerous digital world, where the allegedly proven effects of filter bubbles, cyberbullying, disinformation and conspiracy theories are intertwined.

“They don't know or they haven't understood (...) one of them explained to me [based on disinformation videos viewed on YouTube]: 'The Egyptians already had electricity, otherwise how would they have built the pyramids? Because it's dark inside the tombs, and if they had had torches, they would have been full of smoke. (...)'.” (Teacher 1)

9Paradoxically, this understanding of the digital world is based on a limited and partial use of social networks by the teachers interviewed, who are sometimes unfamiliar with the specific features of each network. On the other hand, among the young people we interviewed in Brussels, we identified an initial level of awareness of the existence of recommendation mechanisms on the social networks they used, beginning with the fact that they mentioned algorithms in connection with suggested content. Utopian and techno-deterministic ideas of the omnipotence of these technologies and their ability to transform the world were also very prevalent. Finally, the comments by experts contained more subtleties, underlining the fact that young people are more informed than they appear, referring to their socio-technical context to explain their relationship to information and algorithms. These initial results are an invitation to extend algorithm education to teachers and educational stakeholders, and to equip them with the tools they need to have a better understanding of the digital world we live in.

10The nature of the relationship which young people in Brussels have to recommendation algorithms was then investigated via 16 focus groups in 2019, each lasting approximately 1,5 hours and involving between 4 and 7 adolescents aged 15 to 18 (a total of 79 people). These focus groups were organised in schools during school hours, without the intervention of teachers. As the Brussels school system is marked by a phenomenon of ethnic and social segregation [Sacco et al., 2016], the focus groups were carried out in five schools in the Brussels-Capital Region selected in order to ensure a certain diversity in terms of the socio-economic conditions of the adolescents in the study. To facilitate the activity, the research team used visual tools such as social network pictograms to be positioned on graphic representations, according to different scenarios (figures 1 to 2). The activities led young people to take individual positions and to discuss very generic themes as a group (types of social network and their uses), before gradually moving on to more concrete information practices (their opinions and ways of finding out about a current issue, for example).

11The discussions and debates which emerged among the adolescents allowed us to further our understanding of their relationship with information. For these young people, content encountered online will achieve the status of information when they are able to find an interest in it and/or when the content gives rise to an emotion. Content which generates joy, well-being or satisfaction (humorous, inspiring, surprising content, etc.), or, on the contrary, induces anger, disgust or disdain (a video of a shooting or an assault, for example) influences its qualification as information. For example, online content related to an attack abroad, the private life of an influencer or football player, homework or family events is generally considered by young people to be information [Mercenier, Wiard and Dufrasne, 2021]. Young people access this content qualified as information via various platforms (Instagram, YouTube, Google (news), WhatsApp and Snapchat). The emergence of interest and/or emotion with respect to content must be understood in the light of the socio-technical network which enables young people to give meaning to information. For example, they explain that they use different devices (smartphone, computer, TV, etc.) to access content considered to be information, and to assess whether it is of interest. These uses are supplemented by the decisive intervention of those close to them and of authority figures (parents, grandparents, teachers, influencers on social networks, etc.), who play a crucial role in distinguishing and producing meaning with respect to the content in circulation.

12The analysis of focus groups also revealed that most of the adolescents interviewed are aware of the existence of content recommendation and filtering mechanisms on the social networks. Many of them described how the organisation of videos, adverts and other content on their news feed is not random, but is due to the intervention of a filtering mechanism.

“-When you click on a video, it's going to analyse what you're watching and recommend those kinds of video.
-And that bothers you? 
-Not really. I figure YouTube needs to know our tastes so it can suggest videos we like.” (student 1)

“On Facebook it's completely crazy. If I look for plane tickets on the internet, I'll get lots of special offers for the destination I've chosen and even articles talking about it. It's unbelievable. Even on Instagram.” (student 2)

13However, they acknowledge that they do not have the words or the technical knowledge to explain how this filtering is carried out. This lack of media literacy [Fastrez and Philippette, 2017] and metalanguage does not prevent them from developing contrasting ideas and certain strategies with a view to influencing recommendation mechanisms. The effects of these recommendations are questioned by some, appreciated by others, and even perceived as limited. Some young people have criticised the presence of these mechanisms, pointing out that they can trap them “in a bubble” and that they need to escape. Others find that the existence of recommendation mechanisms is useful, as they enable them to discover content which would otherwise be considered inaccessible. Finally, some of the study participants felt that the effects of recommendations on their access to information were limited, to the extent that all they had to do was change their habits (take an interest in a new subject, for example) for the recommendation system to incorporate them.

14These results show that, despite a lack of information on how recommendation algorithms work, the majority of the young people interviewed have developed practical knowledge about these algorithms and have implemented a variety of tactics intuitively in order to influence their effects.

“On YouTube, you can see if I've already seen a video or not. Sometimes I'll watch a video a 2nd time to show YouTube clearly that I've seen it and that I want them to suggest that kind of content.” (student 3)

“It makes me uncomfortable when I realise that I've been scrolling for too long. I stop and I set my phone aside.” (student 4)

“I follow the RTL news account, but I stopped because it's a bit depressing. Because most of the time it's bad news, so I unfollowed it.” (student 5)

16Today, young people in Brussels live in a varied information ecosystem, in which recommendation algorithms play a role in their access to information. The study of the perception these young people have of algorithms shows that, while they may not have expert knowledge of recommendation mechanisms, they do have contextual and practical knowledge which enables them to navigate social networking sites with a certain degree of discernment as to the effects of some of their browsing practices on the nature of the information they access. However, this knowledge remains limited and intuitive, and not all of the young people we met have the same degree of skill – which depends more on their active attitude towards tools and information than on their level of education.

3. Emotional and confrontational issues in the formation of opinions

17The study of how young people perceive algorithmic systems revealed that they share a common core of knowledge about major current affairs or issues (in 2019) such as the “Gilets jaunes”, the climate marches, the fire at Notre-Dame de Paris, etc. The existence of this common core of information, supplemented by a myriad of content considered as information at a more individual level, is the result of how young people experience a complex information ecosystem in which technical tools and interpersonal relationships (friends, parents, teachers, influencers, etc.) are intertwined. In the vein of Moeller and Helberger's [2018] findings showing the real but limited effects of recommendation systems on the opinion-forming process of individuals, the ALG-OPINION project illustrates what these effects translate into for an audience of young people in Brussels. To use Pariser's image [2012], if there are bubbles, they are first and foremost socio-technical, porous and intertwined, rather than airtight and conducive to socio-cognitive isolation.

18These results raise at least two important issues regarding the opinion-forming process of this population. Let us first mention the emotional issues related to the algorithm experience. On the one hand, recommendation systems generate states of mind and feelings in their users which contribute to their understanding of technology [Swart, 2021] and, on the other hand, they are also vehicles for users' emotions [Kramer et al., 2014]. These two dimensions were illustrated in the case of young people in Brussels, by pointing out the variety of positions these technologies generate (critical, appreciative and agentive) and the emotions such as joy, satisfaction, anger or disgust they can convey. As the role of emotions has long been identified in the formation of (political) opinions [Hilbert et al., 2018], the ALG-OPINION project paves the way for new research aimed at understanding how such emotions intervene more specifically in the information ecosystem of young people.

4. Ideas for critical algorithm education

21Alveho is a website which displays a news feed made up of press articles from RTBF. This news feed is organised by a recommendation algorithm which assigns a score to each article: the higher the score, the higher the position of the article in the news feed. This score is determined by the article's similarity to the user's browsing history and by the date of publication. By default, the first articles recommended are recent and are very similar to what the user usually views. However, the user can decide to bypass these default parameters through several mechanisms at three key moments in the recommendation process: at input level (the data captured by the system to inform the algorithm), at priority level (the way the algorithm calculates the score) and at output level (the order of recommendations). For input, users can access their browsing history and filter out certain elements which should not be considered by the algorithm. For output, it is possible to hide content categories from the list of recommendations. Finally, users can also configure the importance of similarity or the date of publication of the article for the final score, which will determine the ranking of recommendations. It is therefore possible to ask the algorithm for recent articles which do not place too much importance on a user's history. It is even possible to request articles which are different from one's browsing history.

22In an initial experiment involving 23 participants (master's students whose average age was 22) over a 5-week period in 2021, we used an eye-tracking and interview methodology to assess the extent to which the existence of these configuration tools encouraged website users to question how the algorithm worked and what its effect was on the structuring of the information provided. It appeared that these configuration elements were difficult for users to understand. Although we felt that the filtering of output and the control of algorithms seemed easy (cursors to move, directly accessible in the start window), participants had great difficulty understanding them. We identified several reasons for these difficulties. Firstly, these configuration tools are quite rare on commercial platforms: as participants were not used to modifying the algorithm, they did not realise that they could interact with certain tools. Secondly, it appeared to be difficult to associate actions on the interface with their effects on recommendations. Given that the interface included interactive graphics in addition to configuration tools and the recommendation list, some participants found it difficult to pinpoint how certain actions affected the content they were viewing. In parallel, we tested the integration of Alveho in an experimental media education activity in partnership with Média Animation asbl, a recognised media education association in French-speaking Belgium. In this context, we were able to see that the use of these functionalities in the company of a trained facilitator enabled users to understand how the algorithm works, and served as a lever for the development of critical thought with respect to these technologies.

23Overall, the results associated with the use of Alveho paint a rather complex picture of the use of algorithm configuration tools. Depending on how they are integrated into the interface, these tools can contribute to reflection on the effects of these technologies. However, if these components are poorly integrated or not well explained to users, they may return to default settings which are not always to their advantage. These observations lead us to consider media education which is not only aimed at users, but also at the designers of these systems. Regardless of the level of literacy of users, their capacity for action will always be limited by the recommendation system they use. Therefore, by developing the media literacy [Fastrez and Philippette, 2017] of IT developers, they would be provided with the skills they need to assess the social effects of their design practices.

Conclusion

24The ALG-OPINION research project was aimed at characterising the way in which young people in Brussels perceive algorithms, and at identifying how this perception influenced their information practices. Most of the people interviewed had some practical knowledge of recommendation algorithms, and had developed more or less conscious strategies in an attempt to modify their effects. While it is good news that young people are finding ways to adapt their information ecosystem to the automation of recommendations, doing so also meets what they perceive to be a need. Nevertheless, there are huge differences between individuals in terms of how well they master these skills and tactics, which is mainly the result of practical adjustments whose effectiveness and relevance are not fully understood. The results of the project therefore point to the need to formalise knowledge and adjustment possibilities further with respect to recommendation tools for young audiences. It would also seem appropriate to accelerate reflection on the issues related to recommendation algorithms in the initial and ongoing training of teachers in formal education in Brussels. The development of critical algorithmic literacy among all of these audiences appears to be a future challenge for the Brussels-Capital Region. This literacy should be supported by media education activities, which are already being developed and applied in the region.

25While the Brussels regional government does not have the power to regulate online digital spaces, it can, however, implement actions through its community commissions, aimed at improving their use in the interests of participatory democracy (e.g. Alpha-TIC, BRUZZKet). The development of more educational tools which throw light on the workings of a recommendation algorithm (as in the case of the Alveho platform), as well as the creation of productive discussion forums between the designers of recommendation systems and stakeholders from the worlds of media education, compulsory and higher education, and the media themselves, seem to us to be avenues to explore in order to meet this objective.