Resilience and disempowerment in algorithmic systems

Abstract

Many social media platforms now employ adaptive recommendation algorithms to present content to users, raising concerns about how these systems reduce user agency. Accordingly, scholars have begun investigating how users develop awareness of these algorithms and how they use their understanding to engage with these systems effectively. Taking a focused experimental, mixed-methods approach, this study investigates the choices users (N = 263) make when interacting with algorithmically mediated feeds of social media content and their perceptions of these systems. We found that people made more homogeneous selections when interacting with an adaptive algorithm compared to an algorithm that maintained content diversity. In addition, we found themes of resilience, disempowerment, and distress in participants’ experiences with our algorithmically mediated feeds. Findings call attention to the complex interplay between individual-level differences and algorithmic influences on decision-making when engaging with social media content.

Present across a variety of digital platforms, recommendation algorithms work by mining data from user behaviours including views, likes, comments, and shares, and then presenting new content likely to be of interest. Recommendation algorithms were first developed to help reduce the cognitive load required to explore the vast array of information online (Pariser, 2011). Now, personalized recommendation algorithms enable the passive viewing of online content across a variety of digital platforms with little user effort. To recommend and present content effectively, both the explicit choices people make (e.g. likes and shares) as well as their more implicit, automatic decisions (e.g. time spent viewing a post) inform the algorithm (Zimmer et al., 2019). This means that without requiring explicit decision-making, algorithms can exploit users’ innate biases towards certain content (e.g. emotion-evoking, divisive, and/or status-related posts), creating a feed built on engagement patterns, rather than users’ explicitly indicated preferences (Brady et al., 2023). The presentation of this enticing content keeps users on a platform longer, thereby generating profit through paid advertisements and promotions. Thus, although algorithms enhance company profit, they may do so by creating environments driven by users’ implicit choices.

One outcome of recommendation algorithms may be the homogenization of content for users (e.g. Nguyen et al., 2014), although such claims have been disputed in recent years (Jones-Jang and Chung, 2024). Given the dynamic interplay between algorithms, recommendations, and user behaviour, a key component in the debate about the consequences of algorithms on users is investigating how algorithmic systems shape behaviour and decentre choice (Schwarz, 2025; Ytre-Arne and Moe, 2021). Despite rising unease surrounding the algorithmic influence on user experience, the degree to which these systems shape the moment-to-moment decisions people make online remains unclear.

The unfortunate downside to this algorithmic content curation process, is that users might find themselves trapped in a cycle of viewing and engaging with false, unsafe, politically polarizing, or emotionally distressing content (Cho et al., 2020; Harriger et al., 2022; McLoughlin and Brady, 2024). Algorithms and the content they recommend shape the way that people interact with and understand digital platforms (Siles et al., 2019), and impact the way people construct images of themselves and others (Joseph, 2025; Lee et al., 2022; Mcdonald et al., 2024). Algorithms, and the digital environments they create, therefore have the potential to impact people’s thoughts, feelings, and behaviours.

Importantly, the degree to which algorithmic systems shape users’ experiences largely depends on their awareness and understanding of the complex interaction between their actions online and their algorithm-mediated digital environment. Algorithmic awareness is the degree to which users are aware of how algorithms shape their online experiences (Hargittai et al., 2020; Zarouali et al., 2021). Studies examining algorithmic awareness have found mixed results. For example, some work suggests that users’ awareness of algorithms is generally quite low (Eg et al., 2023; Eslami et al., 2015; Zarouali et al., 2021) while several studies suggest that most users have at least a basic awareness of algorithms and the ways in which they curate social media content (Dogruel, 2021; Gran et al., 2021; Koenig, 2020; Peterson-Salahuddin and Diakopoulos, 2020). One explanation for this discrepancy is that algorithmic awareness appears to be context or platform dependent (Swart, 2021). In addition, differences in algorithmic awareness may relate to demographic differences across samples. Specifically, users with higher education, young adults, and those with more experience online tend to have a greater awareness of algorithms on social media platforms (Oeldorf-Hirsch and Neubaum, 2023). Together, these findings suggest that algorithmic awareness develops from experience both on- and offline and varies across platforms and individuals. However, the way in which algorithmic awareness first develops in a novel situation and how initial sensemaking processes unfold remain unclear.

The way that people make sense of and understand algorithms shapes how they feel and what they do online (Bucher, 2017). Users often report feeling a connection with the algorithms employed on digital platforms in that these systems understand the kind of content to present that aligns with their interests and identity (Taylor and Chen, 2024). While this personalization is viewed as a benefit of algorithmic systems, many also note their pitfalls. Everyday users report distress and dissatisfaction when algorithms present unwanted content and express concern over algorithm-mediated censorship (Swart, 2021). Interestingly, users also report unease with personalized algorithms when these systems appear to “know too much” about their behaviour (De Groot et al., 2023). However, users report that the benefits of personalized recommendation algorithms outweigh their risks and challenges (Park et al., 2025). One way that users attempt to maximize the benefits of algorithmic systems is through building and enacting algorithmic literacy skills.

Algorithmic Literacy refers to a diverse set of skills aimed at directly shaping digital experiences through intentional interactions with an algorithm (Oeldorf-Hirsch and Neubaum, 2025). The extent of users’ algorithmic literacy skills may therefore play a critical role in whether users can reap the benefits of algorithmic systems (e.g. finding community; Erickson, 2024) or get trapped in cycles of unwanted or harmful content (Andalibi and Garcia, 2021). With strong algorithmic literacy, users can employ strategies to inform the algorithm of their preferences thereby avoiding unwanted content and enhancing access to enjoyable content. For example, a user who believes that engagement metrics (e.g. viewing, liking, and commenting) are influential may deliberately reduce their interactions with unwanted content types and simultaneously enhance engagement with more desirable posts (Chen, 2024; Lin, 2025). Accordingly, algorithmic awareness and knowledge are critical antecedents of algorithmic literacy (Alvarado and Waern, 2018). Only recently has research started to assess how technological cues on digital platforms shape users’ experiences, suggesting that a lack of algorithmic transparency creates unease (Oeldorf-Hirsch et al., 2025). Here, we aim to understand how users develop algorithmic awareness in opaque algorithmic systems and under what conditions this awareness translates to proactive strategies. This research therefore provides important insight into how users might optimize their choices when engaging with algorithms.

Current study

This study aimed to develop a deeper understanding of how users comprehend and interact with algorithmically mediated content feeds. First, we aimed to address growing concerns that algorithms reduce content diversity and decentre user choice by examining engagement patterns across two distinct algorithms (one that maintains content diversity and one that adapts to a user’s previous choices). Second, much of the research assessing algorithmic awareness has heavily relied users’ past experiences with familiar digital platforms, and thus the processes by which awareness and literacy skills first develop under opaque algorithms is largely unknown. In addition, most quantitative studies tend to employ survey-based designs, which have been criticized for oversimplifying the unique experiences of users (Oeldorf-Hirsch and Neubaum, 2025).

To remedy these issues, we employ an experimental design. First, we quantitatively address trends in behaviour using an experimental design that examines participants’ engagement with content across two distinct algorithms. Next, to capture participants’ unique experiences and gain insight into the thoughts that guided their behaviour during the experiment, we qualitatively analyse participants’ responses to open-ended, post-experimental questions. This study therefore provides valuable insight about the development of algorithmic awareness in a novel context. While recommendation algorithms themselves are not new, the millions of available applications on digital platforms (e.g. Allix et al., 2016) in tandem with increasing sophistication of algorithmic systems over the past decade (Rainie and Anderson, 2017), suggest that users have an abundance of opportunity to engage in novel algorithm-mediated environments. For example, whereas Facebook and Instagram have long been used globally, in the last 5 years short-form video platforms such as TikTok, which feature highly algorithmically personalized feeds, have quickly grown in popularity (Eddy, 2024). This study will therefore provide foundational insight into how people interact with and discern algorithms in an unfamiliar environment, which is becoming increasingly common in today’s rapidly evolving digital world. We examine the following research questions and hypothesis within a controlled, algorithm-mediated environment across a focussed sample of women and feminine-identifying/presenting people:

• Research Question 1: Do recommendation algorithms, compared to those which maintain content diversity, lead users to select less diverse content?

• Hypothesis 1: We expect that users will select less diverse content when interacting with a recommendation algorithm, compared to one that maintains content diversity.

• Research Question 2: Are participants able to develop an awareness of algorithms in a novel context?

• Research Question 3: If participants develop an awareness of our experiment’s recommendation algorithm, can they then use algorithmic literacy skills when interacting with an algorithmically mediated feed?

Methods

Procedure and design

This study received approval from our university’s research ethics board prior to data collection (protocol #: 124324, 6 February 2024) and took place in a lab setting. Participants provided consent and completed several pre-task questionnaires (e.g. demographics and mood), before beginning the experiment. During the experiment, participants navigated through social media content under two distinct algorithms: one that maintained content diversity within a feed over time (steady-state algorithm); and one that adapted feed content based on a participant’s previous choices (recommendation algorithm). Block order was randomized across participants. In each block, participants completed 10 trials in which they selected a short-form video to view from a 4 × 5 grid of thumbnail images. Each block began with an equal distribution of content (four unique videos from each of five different pre-tested content categories). The videos available for selection refreshed on each trial depending on which algorithm was employed within the block and the choices participants made. Regardless of their selections, participants always had at least one video available from each of the five categories available on their feed. Members of the research team created thumbnails for each video to ensure that they accurately reflected the content in each video (see Jones et al., under review for a sample). In addition, all videos within a category were framed with a coloured border to further assist participants in discriminating among the video types (e.g. blue, green, and orange). Frame colours were randomly assigned to video categories on a participant-by-participant basis. Once a participant viewed a video, that video was removed from the content list, so no video was available for selection more than once. After each video, participants rated their positive and negative emotions, and how much they liked the video. We randomized block order across participants. Finally, participants completed a follow-up questionnaire assessing their perceptions of the task and their decision-making strategies.

Steady-state and recommendation algorithms

In the block containing the steady-state algorithm, the proportion of videos in the participant’s feed (i.e. 20% per content category) remained the same regardless of a participant’s previous selection, thus maintaining content diversity. At the end of each trial, the list of videos available for selection was shuffled, and new thumbnails populated the grid, to signify the selection of a novel video set (Figure 1(a)). Thus, although the exact videos available refreshed on each trial, the number of choices within a content category remained constant.

Figure 1.

Steady-state versus adaptive algorithms.

In the recommendation algorithm block, the available content was modified based on a participant’s previous selections (Figure 1(b)), mimicking a simple recommendation algorithm. Specifically, the recommendation algorithm shifted the distribution of videos available to include more videos of the content type(s) they chose, reducing content in non-selected domains. For example, if a participant selected a video from content-category 1, on the next round the number of videos within that category increased by one, while the number of videos within a different, randomly selected category decreased by one. If the participant made a second consecutive selection of a video from content-category 1, the algorithm repeated this rule. A third and fourth consecutive selection from this category each caused the number of videos from that category to increase by two per trial. Additional consecutive selections from the category increased category offerings by three videos, until grid saturation (16 videos from the same category). To ensure that participants always had the option to choose a video from another content category, at least one video from each content category remained in the grid, regardless of choice behaviour.

Video content and categories

This study is part of a larger project investigating how people navigate through mood-inducing social media content and the emotion regulation strategies they employ in algorithmic systems. As such, we selected content that would elicit both positive and negative emotions as well as “neutral” content that would not necessarily result in mood changes. Although several validated mood-inducing video sets exist, we wanted participants to experience videos that would be like those found on social media. In addition, we wanted to ensure that the available content was not so unpleasant (e.g. graphic, violent, or disgust/fear inducing) that participants would completely avoid those categories. Given these constraints we chose content categories that were not inherently unpleasant but might elicit negative emotions via construal processes such as social comparison (e.g. Fardouly et al., 2017).

Literature in the field of digital media suggests that appearance-focussed content can elicit feelings negative emotion. While evidence suggests these forms of content consistently elicit negative effects in women (Cohen et al., 2019; Gurtala and Fardouly, 2023; Xu, 2024), this association is weaker and inconsistent across men (Brasil et al., 2024). Moreover, men’s exposure to these forms of content is infrequent compared to that of women, with women reporting exposure to body-focussed content online at double the rates of men (Griffiths and Stefanovski, 2019). As this study aims to be a first step in understanding how algorithms with known sets of rules (as opposed to social media algorithms which are opaque) impact people’s behaviours and perceptions, we chose to take a focussed approach to this experiment by selecting appearance-focussed content for our negative mood eliciting stimuli. Consequently, we limited our sample to women and feminine-identifying/presenting folks. Although these categories only represent a fraction of the content participants might see on social media, this focussed approach allows us to explore how people navigate through algorithmically mediated feeds that include emotion-inducing content, without the noise associated with a wider variety of available content.

To create balance in the experiment, we sought to include an additional category of videos aimed at eliciting positive emotions (e.g. cute animals and inspiring scenery) and a non-emotion-eliciting, neutral category (demonstrations of interior painting, cement mixing, cleaning, etc.). The first author (SMJ) and two trained research assistants selected videos aligned with these desired content categories from the social media platform TikTok. We chose TikTok because it is an extremely popular short-form video platform among young adults (Eddy, 2024) and its algorithm has been the topic of significant public discussion (e.g. Rawlinson, 2023; Smith, 2021). Videos were pre-tested to confirm that they had the desired effect on mood and were associated with our intended categories (Jones et al., under review). The five video categories we included were (1) Thinspiration, (2) Fitspiration, (3) Body Positivity, (4) Non-body-focussed Positive (e.g. cute animals), and (5) Neutral (e.g. construction and cleaning).

Post-experimental questionnaire

To assess participants’ perceptions of the experiment and determine whether they were aware of the algorithms in each block we asked several follow-up questions. Using a visual-analog scale ranging from “Not at all (0)” to “Extremely (100),” participants responded to the following statements: (1) “The task was fun,” (2) “The task was engaging,” (3) “I actively chose the content I watched,” (4) “The task seemed like it wanted me to watch certain videos more than others”. Participants then answered an open-ended question to determine whether they were able to detect differences between the two algorithms: “Did you notice anything different in the two parts of the task (the first set of selections you made versus the second)? If so, what?”

Participants

A total of 275 undergraduate women and feminine presenting/identifying students enrolled in an elective Introductory Psychology course participated in this study. As the appearance-focussed videos depicted bodies of women/feminine-presenting people, we limited the sample to those who might directly relate to feminine body ideals or experience the standards of beauty promoted (e.g. thin-ideal) and challenged (e.g. body positivity) within the presented content (e.g. Monteiro and Poulakis, 2019). Accordingly, we removed three participants who did not provide their gender identity. We also removed five participants who completed the survey portion of the study in less than 5 minutes, as this indicated inattentive responding. Thus, our final sample included 263 women, 18 of whom additionally identified as feminine identifying/presenting. Participants’ ages ranged from 18 to 28 (M = 18.67, SD = 1.17). Half (50%) of participants reported White or European as their racial or ethnic identity while the remaining participants identified as South Asian (14%), East Asian (11%), multiple races/ethnicities (9%) or another identity (16%; e.g. South-East Asian, West Asian, and Hispanic).

Analyses

The goals of this study were to determine the degree to which algorithms impact the diversity of the choices people make under different algorithms and how people develop algorithmic awareness in a novel context. We therefore took a mixed-methods analytic approach. First, we assessed whether content diversity differed across the two algorithms quantitatively with a purpose-written script in R 4.5.1 (R Core Team, 2025). Anonymized, aggregated data and analysis scripts are located on the study’s OSF page. Second, we conducted a thematic analysis (Braun and Clarke, 2006, 2021; Fereday and Muir-Cochrane, 2006) on responses to our post-task open-ended question to evaluate participants’ algorithmic awareness and use of algorithmic literacy skills.

Quantitative analyses

We utilized the Shannon Diversity Index (SDI) to calculate the diversity of participants’ selections in each block. Commonly used in environmental studies and biology, the SDI calculates the diversity of species in a given environment (Shannon, 1948). For the present study, we looked at the diversity of content categories participants selected in each block using the diversity function from the vegan package (Oksanen et al., 2025) in RStudio 4.5.1 (R Core Team, 2025).

To assess whether participants’ choices differed in diversity across the two algorithms, depending on algorithm order, we constructed a two-way mixed effects analysis of variance (ANOVA) with selection diversity as our outcome variable, algorithm type as the within-subjects factor, and algorithm order as the between-subjects factor. After running our initial model, we found substantial heteroskedasticity and violations of normality, rendering our p-values unsuitable for interpretation. As a result, we used Nadine Spychala’s (2020) bootstrap_2way_rm_anova function to obtain bootstrapped estimates of the two-way mixed effects.

Qualitative analyses

To assess algorithmic awareness and literacy in the context of our experiment, we used a hybrid inductive and deductive approach to thematically analyse the responses on our open-ended post-experimental question. Specifically, we assessed how participants perceived the two distinct algorithmically mediated feeds in relation to dimensions of algorithmic literacy (Oeldorf-Hirsch and Neubaum, 2025). We followed the approaches outlined by Fereday and Muir-Cochrane (2006) and Braun and Clarke (2006, 2021) for a hybrid thematic analysis.

The first author acted as a primary coder for this study. At each step, the second author of this article, provided feedback. Before initial coding, we developed a preliminary codebook derived from Oeldorf-Hirsch and Neubaum’s (2025) Dimensions of Algorithmic Literacy. We tested the reliability of our preliminary codebook by assessing how well our preliminary codes fit the text and revised when necessary. We then identified preliminary themes within the text, applied our revised codebook to the data and inductively identified additional codes (e.g. mood changes and loss of control) that represented potential new themes within the text. From our inductive and deductive codes, we generated six initial themes. After reassessment of the text in relation to our initial themes, we condensed these themes into two.

Results

Content diversity and quantitative perceptions

On average, participants found the task fun (M = 71.25, SD = 19.38) and engaging (M = 76.40, SD = 17.53). Interestingly, despite participants reporting that they actively chose the content the viewed (M = 80.04, SD = 20.31), they also felt that the task forced them to watch certain types of content over others (M = 76.66, SD = 24.07). Participants’ responses to the items measuring active choice and task control were not correlated (r = −0.01, p = 0.92).

Our bootstrapped two-way mixed effects ANOVA (Figure 2) revealed several significant findings. First, we found a significant main effect for algorithm type, F(1, 261) = 49.31, p < 0.0001. Supporting our hypothesis (H1) that users will select less diverse content when engaging with our recommendation algorithm compared to our steady-state algorithm, we found that when people engaged with the recommendation algorithm, they selected a narrower range of content types (M = 1.09, SD = 0.33) relative to the steady-state algorithm (M = 1.23, SD = 0.28). Block order did not have a significant main effect on selection diversity, F(1, 261) = 1.74, p = 0.19. However, a significant interaction emerged between algorithm type and block order, F(1, 261) = 17.84, p < 0.0001. While both groups made narrower choices when interacting with the recommendation algorithm, this effect was stronger for those who received the steady-state algorithm first (M = 1.03, SD = 0.38), compared to those who received the recommendation algorithm first (M = 1.15, SD = 0.28).

Figure 2.

Selection diversity across algorithm and block order.

We investigated this interaction further by running a Welch’s two-sample t-test to compare the homogeneity of each group’s (recommendation vs steady-state algorithm first) feed after fully interacting with the recommendation algorithm (i.e. trial 10). We found that those who received the steady-state algorithm first had more homogeneous feeds when subsequently interacting with the recommendation algorithm (M = 1.17, SD = 0.28), relative to those who received the recommendation algorithm first (M = 1.25, SD = 0.24). These data suggest that choice diversity reduced more strongly when the initial experience was algorithmically unconstrained, t(242.88) = 2.53, p = 0.01, 95% CI [0.02, 0.15], d = 0.32.

Thematic findings from participant responses

The purpose of our qualitative analyses was to explore whether participants formed an awareness of the two distinct algorithms employed in our task, and whether this awareness led to the effective use of algorithmic literacy skills. Although many participants noticed a difference between the two algorithmically mediated feeds, explicit discussion of algorithms was scarce (5% of participants). While it appears that participants did not develop a specific awareness of the algorithms in our task and thus demonstrate subsequent literacy skills, we did find that participants showed at least some lower-level awareness that an external factor (i.e. our algorithm) may have shaped their digital experience. Instead, we found that participants often discussed changes in content presentation, their own behaviour, and their mood throughout the task – reflecting experiential differences. Accordingly, we identified two major themes across participants’ responses: Inherent Resilience in Algorithmic Systems and Disempowerment and Distress in Algorithmic Systems (Figure 3). These themes highlight the ways in which users may unintentionally inform algorithms and how algorithms subsequently shape user behaviour. To maintain anonymity, quoted participants are identified by an anonymous code.

Figure 3.

Thematic findings of interactions in algorithmic systems.

Inherent resilience in algorithmic systems

Although most participants in our sample did not directly indicate awareness of the algorithm/algorithmic literacy, some of them reported using strategies that ultimately enhanced their experience. That is, they reported using the skills necessary to influence our algorithms but often did so without explicitly articulating key aspects of algorithmic literacy – awareness and understanding.

Participants reported shaping their feeds via positive strategies including emotion regulation, exploration, and flexibility. Those engaging in such strategies identified their emotional states along with aspects of the task that contributed to those emotions. These participants reported subsequently modifying their behaviour to either maintain their current emotional state, alleviate negative emotions, or foster positive emotions:

In the second half [recommendation algorithm], when I became aware of how negatively certain videos impacted me, I started choosing animal videos and confidence boosting videos much more than the first half [steady-state algorithm]. (P1820)

I only chose the videos that I had an interest in watching and that I knew would make me feel good. (P1645)

In the second set [recommendation algorithm], I watched more animal videos than I did in the first block [steady-state algorithm] because it brought me more positive feelings than the more body-oriented videos. (P2136)

In the second part [recommendation algorithm] I actively chose more tiktoks that weren’t related to body image as I realized in the first part that it was bringing me negative emotions when I did watch that sort of content. (P1962)

In the second section [steady-state algorithm] I decided to watch more nature and DIY videos to avoid watching videos that made me feel bad about myself. (P1539)

Comments like these illustrate an active process of stimulus selection for the purpose of affect regulation, along with awareness of the emotional effects of different content types. Participants identified content categories through exploration. For example, P2234 discusses learning about the task and types of content available, before selecting videos that she knew would be of interest and increase her mood:

In the first set [recommendation algorithm] I was mostly just picking random videos as I did not really know what they were about and was unfamiliar with the task. In the second set [steady-state algorithm] I chose videos that would interest me and increase my mood. (P2234)

Participants reported that once they identified the emotional impact of certain content, they were able to avoid videos that they anticipated would negatively impact their mood and select more positive videos. In most cases, participants indicated exploring content early in the task, but some also reported retroactive exploration when the content they were viewing no longer suited their entertainment and emotional needs. For example, P1941 discusses exploration as a reactive emotion regulatory tool:

I selected so many body image videos which made me get bored, so I decided to choose other types of videos which I had not yet explored in case they were entertaining or interesting. (P1941)

P2234’s and P1941’s comments illustrate the role of exploration for entertainment and emotion regulation purposes when engaging with algorithmically mediated feeds. One drawback of such exploration is that participants may encounter undesirable content or negative emotion-eliciting content. In this sense, engagement must be iterative, with continual reflection and exploration, as noted by P1989:

I noticed that in the first part [recommendation algorithm], I intentionally chose more cute videos so that I would feel happy. In the second section, I tried to watch other videos and found they made me uncomfortable, so I switched back to the animal videos. (P1989)

In this example, P1989 appears to retain awareness of her emotional state and the strategies available for regulating that state throughout the task. By responding to changes in her environment and emotions, P1989 refocused engagement towards content that benefitted her well-being.

Disempowerment and distress in algorithmic systems

A concerning number of participants reported a perceived inability to escape loops of harmful or unwanted content. Surprisingly, participants reported these negative cycles across both algorithms, highlighting the critical role that users play in their own social media experiences. Across both algorithms, participants who entered negative cycles reported an inability to recognize alternative options and felt trapped in cycles of harmful content. Notably, participants’ experiences with these problematic cycles appeared more distressing when they engaged with our recommendation algorithm, in which some participants felt manipulated or disempowered. These harmful cycles were reinforced by repeated oversight of available content, perceived loss of autonomy, and perceived algorithmic manipulation.

In our task, regardless of choice behaviour, we ensured that at least one video from each content category was always available for selection, allowing participants a way to change their feed content in the recommendation algorithm. Nonetheless, a number of participants reported an inability to find these alternatives during the task, particularly when engaging with the recommendation algorithm:

In the second part [steady-state algorithm] of the task, the feed included videos that were encouraging self-love and reducing insecurity whilst in the first part of the task [recommendation algorithm], there were none. (P1444)

In the first part of the task [recommendation algorithm], the videos all seemed to be body-related content. The second part [steady-state algorithm] of the task, body-related content was there but there were other things to choose from as well. (P1885)

In these examples, participants’ perceptions of their feed are misaligned with the recommendation algorithm’s behaviour. That is, participants falsely reported that certain content categories were unavailable to them when engaging with the recommendation algorithm. Thus, it appears that as some participants engaged with this algorithm, which selectively amplified content, they were less able to attend to other material.

Participants also reported feeling a loss of control and greater passive use as they continued to engage with the task. This gradual shift to more passive use occurred across both algorithms, though was more frequently described when the recommendation algorithm was second, suggesting an effect of both order and algorithm type on participants’ experience. This progressive perceived loss of autonomy is illustrated in the following quotations:

The second time [steady-state algorithm], I was less consciously choosing the videos and found myself gravitating towards more of the body checking content. (P1570).

In the second half [recommendation algorithm] of the task, it was harder to resist watching videos I knew would make me feel self-conscious. (P1810).

For the first part [steady-state algorithm], I felt that I was more open to watching a variety of different videos, but in the second portion [recommendation algorithm] I was only compelled to watch the videos with body-checking/pro-ana content. (P2004).

Interestingly, the content some participants felt compelled to engage with over time and in the recommendation algorithm was content that they self-classified as harmful. Indeed, participants felt unable to resist selecting content they described as eliciting negative thoughts (“self-conscious”) or promoting eating disorders (“pro-ana,” a colloquial term for content that promotes restrictive eating). These findings suggest that even though users may be able to recognize the emotional impact of certain types of videos, some users have difficulty understanding how to avoid such content, especially if that content continues to dominate their feed.

Importantly, although the quotations above speak to an internal loss of control over viewed content, an additional subset of participants felt instead that the recommendation algorithm actively reduced their autonomy. That is, participants blamed the algorithm for pushing this harmful content:

It seemed like the system wanted me to watch more of videos of females and their bodies even though I had negative emotions associated with it. (P1662).

There were more body-oriented videos in the second part [recommendation algorithm] compared to the first part [steady-state algorithm]. Therefore, I did not have the chance to look at relaxing content that was unrelated to feelings of comparison, or satisfying, relaxing content that was present in the first part. I believe the second part was more aimed at getting me to watch videos about the body. (P1849)

Interestingly, some participants felt similarly in the steady-state algorithm. Specifically, they felt that task encouraged them to watch certain content over others:

The second part [recommendation algorithm] allowed me to choose happier videos. The first part [steady-state algorithm] I believe really encouraged me to watch workout videos whereas in the second part [recommendation algorithm] if I watched an animal video, I would have the opportunity to watch more animal videos. (P1836)

The second set [steady-state algorithm] offered a lot more body tiktoks, so I had to choose those. (P1737)

In the examples above, participants felt pushed towards harmful content despite equal category representation within the steady-state algorithm. This suggests that certain content may capture attention to a greater degree when it appears within a feed, making it seem as though the algorithm controlled participants’ choices, when in fact it did not.

Discussion

These findings provide a nuanced and granular look at the complex interaction between users and recommendation algorithms. Previous work has highlighted the critical way that algorithms shape the content people see online (e.g. Nguyen et al., 2014; Pariser, 2011), raising concern over whether recommendation systems pose a barrier to user autonomy (Swart, 2021; Ytre-Arne and Moe, 2021). To assess how recommendation algorithms shape behaviour, we employed an experimental design to examine variation in user choice across two distinct algorithms – one that presented content based on a user’s previous choices and one that did not. Moreover, while research has highlighted the importance of algorithmic awareness and literacy when engaging with recommendation systems (Oeldorf-Hirsch and Neubaum, 2025), how this knowledge develops in novel situations is largely unknown. To gain further insight, we employed a qualitative analysis assessing people’s algorithmic awareness and literacy in the context of our task. Our findings therefore provide both general and person-specific accounts of the ways in which recommendation algorithms shape digital experience, advancing knowledge in human–computer interaction.

Supporting our hypothesis, we found that users selected a narrower range of content when interacting with our recommendation algorithm compared to our steady-state algorithm. These findings suggest that recommendation algorithms may limit variety-seeking behaviour, providing real-time evidence for previous work suggesting that algorithmic recommendations are linked with reduced consumption diversity (Anderson et al., 2020). Accordingly, loss of autonomy was a prominent topic in our qualitative findings, wherein participants reported entering apparently inescapable cycles of harmful content, especially when engaging with our recommendation algorithm. While participants recognized the distress certain video categories elicited, they often felt disempowered and unable to avoid that content within their feeds. Participants attributed this perceived loss of agency both to their own self-regulation skills and to a perception that the task pushed certain content over others. Recent work has called attention to the dangers of algorithms in their tendency to both push increasingly similar content over time as well as content that is more extreme and, at times, dangerous (e.g. Center for Countering Digital Hate, 2022; Regehr et al., 2024). Our findings extend this work by showcasing how these systems reduce content availability and users’ sense of autonomy, thereby supporting recent qualitative findings in the field (Verwiebe et al., 2024; Xu et al., 2025). Importantly, this study provides new evidence on how users’ perceived loss of agency contributes to the development of cycles of viewing and interacting with harmful content on algorithmically mediated social media feeds. Together, these results highlight how recommendation algorithms can exploit pre-existing vulnerabilities, emphasizing the need for interventions aimed at helping users engage with algorithms in healthy and effective ways.

Importantly, we sought to explore whether participants developed algorithmic awareness and used this to inform behaviour. Addressing our second research question (RQ2; are participants able to develop an awareness of algorithms?), we found that few participants indicated clear awareness of the relevant algorithms and even fewer reported directly engaging (RQ3; if aware of the algorithm, do participants use algorithmic literacy skills?) with these algorithms. Regardless, many participants actively controlled their digital experience by exploring available options and selecting the videos that best suited their emotional and entertainment needs.

Emotion regulation (see Gross, 1998) has been identified as a critical factor impacting how and why people use social media (Gioia et al., 2021; Wadley et al., 2020). Indeed, a subset of our participants recognized how different content categories induced emotion and actively regulated their emotions through subsequent selections, supporting recent frameworks of digital emotion regulation (Hollenstein and Faulkner, 2024). For some participants these selections involved content that directly improved their mood (e.g. “I watched more animal videos [. . .] because it brought me more positive feelings”) or changed their thought patterns (e.g. “When I became aware of how negatively certain videos impacted me [. . .] I started choosing [. . .] confidence boosting videos”). These findings support the importance of exploratory behaviour and flexibility in digital spaces for effective emotion regulation. Thus, our finding that recommendation algorithms reduce variety-seeking behaviour may reflect algorithmic influence and loss of agency as well as attempts to optimize affective experience through intentional content selection. These results highlight the importance of emotion regulation when engaging with opaque algorithms where awareness, understanding, and algorithmic literacy may be difficult to achieve.

As a part of our experiment, we randomized the order in which participants received our recommendation algorithm. Presentation order significantly moderated the recommendation algorithm’s effect on choice behaviour. Specifically, the reduction of diverse content selection when engaging with the recommendation algorithm was particularly strong when participants received this algorithm second. One possibility for this result is that as people continued to engage in the task, they shifted from intentional decision-making processes to more automatic, habitual ones. Indeed, frequent social media use is linked with more habitual engagement behaviours on digital platforms (Anderson and Wood, 2023). Accordingly, we found a pattern of responses indicating more automatic decision-making during the second portion of the task, regardless of algorithm (e.g. “less consciously choosing the videos,” feeling “compelled” to watch certain content). Thus, another explanation for the moderating effect of block order on selection diversity is that over time, participants engaged more passively in our task, allowing the narrowing nature of the adaptive algorithm to have a stronger influence on their selections. These findings provide preliminary evidence that, for some users, engagement behaviour within a social media session may become increasingly passive, though future research should test this directly.

The moderating effect of algorithm order may, alternately, reflect the benefit of receiving our steady-state algorithm first. Though our recommendation algorithm appeared to reduce variety-seeking behaviour, at times this behaviour was protective (e.g. “When I became aware of how negatively certain videos impacted me, I started choosing animal videos and confidence boosting videos much more”). When participants received the steady-state algorithm first, they were able to explore the range of available content, without the influence of an adaptive algorithm. Prior work suggests that exploration helps people gain insight about an environment and make more informed decisions (Cohen et al., 2007). Those who began in the less restrictive environment may have entered the adaptive block with more knowledge of the content categories. Thus, another possible explanation for the moderating effect of block order is that receiving the steady-state algorithm first facilitated early exploration, allowing participants to develop informed content preferences. Thus, these participants may have entered the adaptive block already knowing which content categories they wanted to select.

Typically, social media platforms present users with a large variety of content to explore before narrowing recommendations over time as algorithms learn from a user’s engagement patterns (e.g. TikTok, n.d.). Our findings therefore highlight the ways in which algorithm design interacts with people’s different response styles to shape the personalization of their content feeds. Specifically, exploration, either through the promotion of diversity maintaining algorithms or personal self-regulation, appears to be a critical factor in facilitating positive social media experiences. Moreover, while recommendation algorithms can help people continue to find content that satisfies their emotional needs, such algorithms can also contribute to harmful experiences when users feel overpowered by these systems. Our study therefore provides support for the idea of algorithmic transparency and for giving users more control over how algorithms serve them content.

Limitations & future directions

Due to the proprietary nature of social media recommendation algorithms, our study did not attempt to replicate the full extent of people’s experiences online. Instead, we opted for a simple algorithm that promoted content based on each user’s own choices, which allowed us to avoid the contamination of one person’s feed with the preferences of another. Nonetheless, our findings provide critical insight into how simple recommendation algorithms shape behaviour, thereby enhancing understanding of how people interact with the more complex algorithms with which they regularly engage. Both Meta and TikTok, two of the largest and most popular social media companies, indicate that they recommend content to users through the assessment of what else they view on the platform (Meta, n.d.; TikTok, n.d.). Our recommendation algorithm worked in a similar manner: the more a participant viewed a certain content category, the more content from that category was available in the next trial. Thus, even our simplistic algorithm provides some ecological validity. Future work should continue to investigate how the algorithms employed on social media platforms shape online behaviour, especially over time and across platforms.

The relatively short length (~15 minutes) of our experimental task may have additionally limited our findings. It is possible that if participants were given more time in our task, they would have gained more explicit knowledge of our algorithms and a greater opportunity to develop algorithmic awareness. That is, on one hand, more experience might have contributed to the development of greater algorithmic awareness and literacy, potentially impacting choice behaviour. On the other hand, over 90% of participants reported almost daily social media use and engagement with multiple platforms, meaning that ours was a sample of people with extensive experience of different social media algorithms, a fact that should have enhanced sensitivity to algorithmic manipulation.

Despite these limitations, our experiment promotes understanding of how people initially interact with opaque algorithms in novel contexts. Understanding how people first interact with unknown algorithms is important, given the rise in the development of new social media platforms and updates to existing ones. Though participants did not express explicit awareness of the algorithms in our study, future work should assess how choice behaviour, especially in its more subtle forms when choices are based on both viewing and choosing not to view content, as in many proprietary social media algorithms. In addition, algorithmic awareness and literacy skills may develop over a longer time period than the immediate experiences participants had in our task. However, our algorithm was also quite simplistic relative to most social media algorithms, as it was based solely on the explicit choices of a single participant, rather than an entire network of views, likes, and shares. Thus, we would anticipate that the fact that so few participants recognized our algorithm for what it was is an effect that would be significantly magnified in the context of real social media. In addition, our study highlights the degree to which experiences with social media vary across individuals despite similar digital environments. Future research should assess how individual-level traits such as personality, cognitive styles, and demographics (e.g. assessing these effects in men and/or non-binary folks) predict variation in patterns of engagement in digital environments. This work underscores a significant need for research assessing when and who is most vulnerable to the impact of adaptive, recommendation algorithms, and why such differences occur.