Adolescents' Algorithmic Resistance to Short Video APP's Recommendation: The Dual Mediating Role of Resistance Willingness and Resistance Intention.

Adolescents have gradually become a vital group of interacting with social media recommendation algorithms. Although numerous studies have been conducted to investigate negative reactions (both psychological and behavioral reactance) that the dark side of recommendation algorithms brings to social media users, little is known about the resistance intention and behavior based on their agency in the daily process of encountering algorithms. Focusing on the concept of algorithm resistance, this study used a two-path model (distinguishing resistance willingness and resistance intention) to investigate the algorithmic resistance of rural Chinese adolescents (N = 905) in their daily use of short video apps. The findings revealed that the perceived threat to freedom, algorithmic literacy, and peer influence were positively associated with the resistance willingness and intention; while the independent psychology on algorithmic recommendations significantly weakened resistance willingness and intention. Furthermore, this study verified the mediating role of resistance willingness and intention between the above independent variables and resistance behavior. Additionally, the positive impact of resistance willingness on resistance intention was confirmed. In conclusion, this study offers a comprehensive approach to further understanding adolescents' algorithmic resistance awareness and behavior by combining psychological factors, personal competency, and interpersonal influences, as well as two types of resistance reactions (rational and irrational).

Introduction

With the advent of the platform society (Van Dijck et al., 2018), algorithms are increasingly becoming an indispensable infrastructure for social media platforms to interact with their users (Arriagadaand Ibáñez, 2020). Compared with the active but time-consuming video searching behavior, recommendation algorithms make it more efficient for social media users to obtain the diverse and abundant content they are interested in through accurate recommendations based on big data, completing the transformation from “I look for information” to “information finds me” (Park and Kaye, 2020). Despite the potential advantages that recommendation algorithms provide to users, they also have the “dark side” (Salmela-Aro et al., 2017; Springer and Whittaker, 2018; Ma et al., 2021). Previous research has shown that as a structural power that constrains users’ agency (Reviglio and Agosti, 2020; Schwartz and Mahnke, 2021), recommendation algorithms pose a range of problems such as visibility hegemony (Swart, 2021), operational opacity (Kulshrestha et al., 2017), bias and discrimination (Kulshrestha et al., 2017), information overload (Lin et al., 2020), disinformation and misinformation (Clark, 2020) and privacy breaches (Lam and Hsu, 2006), which may lead to negative psychological and behavioral responses from platform users, such as social media fatigue (Bright et al., 2015; Dhir et al., 2019; Whelan et al., 2020; Fan et al., 2021; Pang, 2021), fear of missing out (Roberts and David, 2020; Tandon et al., 2021), and forced disconnection (Nguyen, 2021; Vanden Abeele et al., 2022), platform migration (Maier et al., 2015; Luqman et al., 2017), etc. Although such researches have explored the operation mechanisms of algorithm from an ontological perspective, focusing on its structuring power (i.e., the structural limitations of data-tracking apps on user information access and platform use) (Beer, 2017; Ford, 2021; Morrison, 2021; Welch, 2021), little attention has been paid to the dynamic processes of how users encounter algorithm and exploit their agency (DeVito et al., 2017; Ettlinger, 2018; Klinger and Svensson, 2018; Rubel et al., 2020; Karizat et al., 2021; Velkova and Kaun, 2021). It means that the existing research situation makes the user, who is the subject of the algorithmic experience, overlooked in the algorithmic culture (Velkova and Kaun, 2021).

However, besides the platform and algorithm designers, users of algorithms are also key actors in constructing algorithmic outputs (Ettlinger, 2018; Mircicǎ, 2020), whose agency in the process of interacting with algorithms always presents two sides. One side is reflected in the conscious and active cooperation with the structural power of algorithms (Schwartz and Mahnke, 2021), such as actively following the rules set by the algorithm in order to gain visibility “rewards” (Bucher, 2012), The other one is presented as a more radical way to “rewriting” and “subversion” the logic and output of algorithm (Bucher, 2012; DeVito et al., 2017; Siles et al., 2020). From these two perspectives, it is of great importance to re-investigate the interactive practices that users’ engagement during their encounters with algorithms in order to gain a deeper understanding of the dynamics of their everyday use. Drawing on Giddens’ structuration theory (1984), which inspired a “dialectic of control” in the digital age (Schwartz and Mahnke, 2021), some researchers have emphasized the significance of the user as an actual actor influencing the recommendation algorithm: while the algorithm “structures” users, their feedback to the algorithm by utilize the algorithm’s own rules is also becoming an important power for reshaping the algorithm’s rules, for example, using the personalized recommendations to “mislead” the algorithm (Leong, 2020). In summary, based on the rethinking of user subjectivity and agency in the “algorithm-user” relationship, Couldry and Powell (2014) argued that the current focus of algorithm research should shift from “algorithm-centric” to “algorithm-user interaction,” in other words, from exploring the one-way structural shaping power hegemony of algorithms to the two-way interactive practices between users and algorithms in their everyday social media use. Especially, it is crucial to Concentrate on users who give positive feedback after encountering algorithms is key to re-establishing user subjectivity and promoting continuous adjustment and optimization of the platform algorithms themselves (Velkova and Kaun, 2021). In the context of the turn in the overall perspective of algorithm research, more and more studies have begun to focus on users’ agency in recent years, examining how they understand, feel and engage with algorithm interaction practice (Bucher, 2017; Fletcher and Nielsen, 2019; Schwartz and Mahnke, 2021; Swart, 2021). In turn, such as “decoding algorithms” (Lomborg and Kapsch, 2020), “algorithmic imagination” (Bucher, 2017), “algorithmic folk theory” (Karizat et al., 2021) and a series of related concepts have been proposed to describe users’ subjectivity and agency when encountering algorithms.

Being one of these concepts, “algorithmic resistance” has been developed to characterize an interventional and energetic practice of social media users toward recommendation algorithms in the digital age (DeVito et al., 2017; Ettlinger, 2018; Karizat et al., 2021; Velkova and Kaun, 2021). Combining De Certeau’s theory of everyday resistance practice (De Certeau, 1984), Feinberg’s critical theory of techno politics (Feenberg, 2002), Foucault’s theory of productive micropower (Foucault, 1980), and digital activism theory (Klinger and Svensson, 2018), algorithmic resistance focuses on social media platform users’ active resistance practices against the platform’s algorithms using their cognitive and practical abilities when encountering them. This type of resistance differs from negative usage behaviors (e.g., stopping use due to social media fatigue),which is a combination of positive awareness and practice against the cognitive and behavioral threat recommendation algorithms bring. Through active participatory use and negotiation to exploit the existing rules of the algorithm, users as actors try to change the original logic of algorithmic output (Clark, 2020),violate the “original” meanings set by the algorithm and finally create diverse meanings to meet their usage purposes (Eslami et al., 2016; Ettlinger, 2018; Velkova and Kaun, 2021). Driven by this concept, a number of studies have explored diverse forms of algorithmic resistance among social media users, such as attempting to domesticate algorithms using social media content recommendation rules (Sujon et al., 2018; Leong, 2020; Siles et al., 2020); utilizing algorithmic ranking rules to make the content one wants to display weight up to become “visible” (Velkova and Kaun, 2021); “teasing” and “confusing” the TikTok algorithm by changing personal preference settings to achieve “unexpected” recommendation visibility effects (Cotter, 2019), fighting against recommendation algorithms’ discrimination toward marginalized groups by frequently adjusting identity privilege settings (Mittmann et al., 2021), and so on. Nevertheless, the above-mentioned studies on algorithmic resistance still have some shortcomings, which leave some space for expansion in this study. First, some studies are still limited to exploring the level of resistance awareness (Sujon et al., 2018; Siles et al., 2019a,2020), lacking an examination of the link between algorithmic awareness and further resistance behaviors, which this study seeks to explore. Second, existing research on users’ negative reactions to platform recommendation algorithms has focused more on the influence of users’ psychological factors on resistance (Velkova and Kaun, 2021). For example, most studies have used the S-O-R framework (Lin et al., 2020; Ma et al., 2021) while lacking consideration of individual ability factors as well as interpersonal factors. Therefore, this study attempts to propose an explicit research model that can integrate the above-mentioned factors that affect algorithmic resistance. Third, some studies have made an ambiguous distinction between algorithmic recommendation content, a particular feature of the algorithm (i.e., greedy), and the recommendation system itself (Ma et al., 2021), which makes it impossible to distinguish what users’ resistance intentions and behaviors are targeted, and this study points to the recommendation system of short video platforms. Furthermore, although several prior studies have attempted to “decode algorithms” from a user perspective (Lomborg and Kapsch, 2020), most of them have been conducted through qualitative methods such as semi-structured interviews and self-reports (Koenig, 2020; Lomborg and Kapsch, 2020; Swart, 2021). Whereas such methods are relatively trivial for generalizing factors that influence users’ algorithmic resistance and may lack general validity, this study attempts to construct generalizable quantitative models. Finally, the factors that lead to the implementation of algorithmic resistance behaviors are not necessarily rational, that is, users are likely to follow other social media users in the same resistance behaviors without a clear sense of resistance (Ettlinger, 2018), especially since users themselves are more likely to be influenced by others in the process of behavior (Geber et al., 2021), implying their behavior may not exactly follow rational paths, but that irrational behavioral triggers are also present. However, many studies do not distinguish these two paths carefully. Consequently, the study tries to integrate two models, the theory of planned behavior (TPB) and the theory of prototypical willingness, in an attempt to establish a dual mediation model of rationality and irrationality by separating the willingness to resist (irrational path) from the intention to resist (rational path).

What also should be mentioned is that adolescents are now a hugely vital group for using social media, particularly short video apps like TikTok and Kuaishou (Bossen and Kottasz, 2020; Basch et al., 2021; Mittmann et al., 2021), and the recommendation algorithms of these apps are becoming digital infrastructure that shapes their everyday practices (Shin et al., 2021), making it a common situation that adolescents are encountering algorithms. Although several studies have examined various limitations of algorithmic hegemony on the cognitive and usage dimensions of adolescent social media users from a critical perspective (Bucher, 2012), for example, critically discussing various restrictions and manipulations of platform algorithms on the content visibility and access of them (Rieder et al., 2018; Petre et al., 2019; Bandy and Diakopoulos, 2021), they did not give enough attention to the demonstration of agency in adolescent-algorithm interactions, implying that ‘agency is often neglected in the emerging discussion of the consequences of algorithmic culture’ (Velkova and Kaun, 2021: 524), and that adolescents’ active interactions with algorithms in their perception and everyday use are somewhat overlooked (Shin et al., 2020). In other words, how adolescents employ their subjectivity and agency to understand, engage with, and experience algorithms are not fully understood (Shin, 2020; Swart, 2021), and even less is known about how adolescents participate in the resistance activities based on their views and usage of algorithms (Karizat et al., 2021).

To fill the above gaps, this study focuses on the concept of “algorithmic resistance” and its empirical evidence among adolescents. By examining the interaction practices of Chinese rural adolescents with recommendation algorithms during their daily use of short video apps, we constructed a dual-mediated path model of algorithmic resistance that integrates psychological factors, interpersonal influences, and personal capabilities, aiming to investigate the relationship between adolescents’ active resistance behavior to recommendation algorithms during their daily use of short video platforms and their resistance intention and willingness. The research questions of this study are as follows:

RQ1: What specific factors influence the algorithmic resistance of Chinese rural adolescents to the recommendation algorithms of short video platforms?

RQ2: Are there two mutually distinguished mediating paths between influencing factors and resistance behaviors? What are their differences and connections respectively?

Literature Review

Algorithm Literacy, Algorithmic Resistance Intention, and Willingness

Algorithmic literacy (AL) refers to the ability of users to perceive and use algorithms in their daily encounters with social media (Hamilton et al., 2014; Shin, 2020; Dogruel et al., 2021; Shin et al., 2021). Compared to algorithmic awareness, algorithmic literacy covers among users’ basic awareness of the algorithm, critical consciousness, and their ability to use it (Hamilton et al., 2014). According to Dogruel (2021), algorithmic literacy has two main components: algorithmic knowledge and algorithmic usage awareness, which in turn can be divided into four dimensions: (1) awareness and knowledge, (2) critical evaluation, (3) coping strategies, and (4) creation and design skills. While being a user of digital media naturally implies interacting with the algorithmic infrastructure that provides the shape for personalized interfaces and media content, we cannot assume that every media user is fully aware of this (Eslami et al., 2016). A standard public will not have the necessary time or know-how to judiciously determine the consequences of that algorithmic pattern (Kemper and Kolkman, 2019), suggesting that there is variability in users’ algorithmic literacy both within and across the above four main dimensions.

In particular, social media users’ algorithmic literacy has a positive impact both on their intention and willingness to resist algorithms at the same time. According to the distinction between behavior willingness and behavior intention by Gibbons et al. (1998a), resistance willingness (RW) refers to users’ irrational behavioral choice tendency which is not through deeply thought out and is based more on others’ influence and their negative emotion toward algorithms, while resistance intention (RI) can be defined as a more rational behavioral tendency after users have evaluated their ability, resistance results, and algorithmic influence during their daily usage. However, although resistance willingness and resistance intention are two variables with essential differences, separately representing the psychology of irrational and rational influences on behavior, few studies have distinguished this in their exploration. This study is conducted with adolescents in their growing years, whose behavior is characterized by both irrational and rational characteristics. Therefore, the distinction between these two variables is considered necessary.

In terms of the impact of algorithmic literacy on resistance intention and willingness, some studies have shown that social media users are not passive bystanders but active participants when encountering algorithms, and they gradually become aware of the vital role they play in shaping algorithms and intervening in algorithmic connections, which further increases the likelihood that they will actively implement algorithmic resistance behaviors by using various tactics (van der Nagel, 2018; Velkova and Kaun, 2021). Research by Lomborg and Kapsch (2020) also confirmed that increased awareness of users’ algorithmic critique helps them to make more sensible decisions during their interactions with algorithms. Meanwhile, a study of Instagram users’ interactions with recommendation algorithms showed that users’ perceived awareness of algorithmic visibility limitations significantly influenced their possibilities of utilizing algorithmic rules to intervene in the output to ultimately serve their interests (Cotter, 2019). Bucher’s (2017) study about Facebook users’ experience of algorithms also implied that when users realize that the results of Facebook’s algorithm operation can be changed by personal intervention, they would positively develop strategies to trick the algorithm into operating toward unexpected goals. While the degrees and types of algorithmic literacy also differ in their impact on the psychology to resist, Ettlinger (2018) distinguished between the impact of different critical awareness of algorithms on their resistance intention and resistance willingness based on users’ motivation, where individuals who are fiercely critical of algorithms are more likely to resist them than those who lack critical awareness. This finding shows that users’ basic, critical, and practical perceptions of algorithms positively contribute to their resistance psychological states (including both resistance willingness and intention). Grounded on the above literature, this study proposes the following hypotheses:

H1: Users’ algorithmic literacy positively predicts their algorithmic resistance intention.

H2: Users’ algorithmic literacy positively predicts their algorithmic resistance willingness.

Perceived Threat to Freedom, Algorithmic Resistance Intention, and Willingness

The perceived threat to freedom can be defined as a psychological situation when an individual’s freedom is disrupted by an external force (Brehm and Brehm, 1981). When individuals feel that their liberty is threatened, they are likely to develop resistance psychology and may, in turn, consciously take some positive actions to regain their liberty based on their resistance psychology. Hence, it is clear that the trigger for this psychology is the restriction and threat to individual freedom by external forces. In the specific field of media use, previous studies have examined the negative psychological and behavioral responses of social media users (i.e., information overload, narrowed vision, restricted access, and forced connection) such as social media fatigue (Bright et al., 2015; Whelan et al., 2020; Pang, 2021), fear of missing out (FOMO) (Tandon et al., 2021), forced disconnection (Nguyen, 2021; Vanden Abeele et al., 2022), and platform transfer (Maier et al., 2015; Luqman et al., 2017), and so on. However, less research has been conducted concerning the psychological condition of users who actively resist after feeling the sense of unfreedom brought by algorithms. The fact is that as a broadly dominant media power (Beer, 2017; Krasmann, 2020; Cotter, 2021), algorithmic restrictions on information visibility and access to platform functions (Bucher, 2012; Bandy and Diakopoulos, 2021) are also likely to create a certain sense of unfreedom for users, which in turn may lead to resistance to the outputs of recommendation algorithms. For illustration, Lomborg and Kapsch (2020) discovered through qualitative interviews that some users’ strong negative emotional reactions (i.e., anger, upset, hatred, discomfort, and so on) to the tags imposed on them by the “personalized recommendations” can lead to active resistance to algorithms.

Moreover, Karizat et al. (2021) revealed from a study about algorithmic resistance in American TikTok users that adolescents’ perceptions of identity discrimination from recommendation algorithms that restrict their access to and use of content lead to a sense of disillusionment with the algorithm, which further enhances their potential to resist implicit algorithmic discrimination. A study of Spotify users’ feedback shows that adolescents define the platform as “a very annoying dude” or “the most intense of your friends” when they perceive Spotify’s recommendation algorithm as a threat without their permission. Users prefer that algorithms hide their “face” rather than draw attention to them, which is also a form of representation of resistance psychology (Siles et al., 2020). Therefore, based on the relationship between the algorithmic threat to freedom and resistance intention, the study proposed the following hypotheses:

H3: Users’ perceived threat to freedom positively predicts their algorithmic resistance intention.

H4: Users’ perceived threat to freedom positively predicts their algorithmic resistance willingness.

Peer Influence and Algorithmic Resistance Willingness

Peer perceptions and behavioral practices are significant factors influencing adolescents’ perceptions and behaviors, and numerous prior works have tested the role of peer persuasion and behavioral modeling in promoting adolescents’ willingness to engage in risky behaviors (Bauman and Ennett, 1994; Werner-Wilson and Arbel, 2000; Scull et al., 2010; Geber et al., 2021). It is noteworthy that in recent years, a growing group of researchers has extended the validation of this impact to the digital medium aspect, such as exploring the role of peer influence in adolescent online privacy disclosure (Mullen and Fox Hamilton, 2016; Fan et al., 2021), and its facilitating effect on willingness to “Like” on Instagram (Sherman et al., 2016).

Indeed, as an intervening “risky behavior of abnormal use,” adolescents’ behavioral willingness to resist algorithms is also affected by peers’ algorithmic knowledge sharing, showing that third groups (especially interpersonal channels) are an important source of personal algorithmic knowledge for adolescent users. In turn, adolescents acquire and refine their knowledge about algorithms through conversations with acquaintances and friends (Lomborg and Kapsch, 2020), which can then have an impact on their basic cognition of algorithms (i.e., how algorithms make some content invisible while giving strong exposure to some content) and thus shape their willingness to resist algorithms. This means that if individual users have more information from others about the negative problems caused by algorithms (i.e., surveillance, echo chambers, filter bubbles), their opinions about algorithms may also be relatively negative (i.e., “disturbing,” “harmful”), which subsequently motivates users to resist. By the same token, if individuals are in a group that has an overall negative attitude toward algorithms and positively develops algorithmic resistance tactics, despite their low critical awareness of algorithms, they are also likely to choose to comply with others and reinforce their willingness to resist due to the convenience of usage brought by resistance behavior (Lomborg and Kapsch, 2020). Previous researches on adolescent risk behavior also suggest that peer prototype is a specific factor that more possibly leads to the generation of irrational resistance willingness rather than rational resistance intentions (Gibbons et al., 1998b; Pomery et al., 2009). Hence, consistent with the arguments presented above, the following hypothesis was formulated:

H5: Peer influence is positively associated with adolescents’ resistance willingness to the recommendation algorithms of short video apps.

Algorithmic Dependent Psychology, Resistance Intention, and Willingness

Dependency is defined as a continued relationship in which the satisfaction of needs or the attainment of goals by one party is contingent upon the resources of another party (Ball-Rokeach and DeFleur, 1976; Baier, 1986). Media dependency theory (MDT) proposed by Ball-Rokeach and DeFleur (1976) focuses on this ongoing relationship between mass media and individual users, stating that individuals become dependent on mass media use for motivations such as entertainment, information acquisition, and maintaining connections. In the digital age, with the research on media dependence expanding into social media, several studies have conceptualized and empirically explored the psychology of users’ dependence on SNSs (Yang et al., 2015; Kim and Jung, 2017; Lee and Choi, 2018; Han et al., 2019), especially emphasizing the importance of trust (Pop et al., 2021), interactivity (Kaye and Johnson, 2017), and personal psychological characteristics (i.e., anxiety, depression) (Lǎzǎroiu et al., 2020; Porter et al, 2020) in the generation of users’ social media dependence.

Similarly, the psychology of user dependence on algorithms can be understood as the expectation that users maintain an ongoing steady relationship with the recommendation algorithm during their daily interactions with the algorithm, and this dependence is largely based on the fact that recommendation algorithm can provide users with stable, continuous, and interesting content. Studies on Facebook Page Ranking algorithms showed that the convenience, engagement, and immersive mind-flow experience that algorithms bring to adolescents (Salmela-Aro et al., 2017) reinforces their dependence on the algorithm and blinds them to its dark side (lack of critical awareness) (Schwartz and Mahnke, 2021). At the same time, recommendation algorithms may also maintain adolescents’ persistent use by motivating them to actively interact with the algorithm by using content visibility as a “reward” (Bucher, 2012; Eslami et al., 2016), cultivating dependent psychology. This dependency on recommendation algorithms is likely to lead adolescent users to view them as a “necessity” when using social media, ultimately reducing their fatigue and resistance to recommendation algorithms. Some studies also have found that immersion and mind flow intensify adolescents’ negative reactions to recommendation algorithms (Lin et al., 2020; Ma et al., 2021). In summary, the following two hypotheses are considered reasonable:

H6: Dependent psychology is negatively associated with adolescents’ resistance intention to the recommendation algorithms of short video apps.

H7: Dependent psychology is negatively associated with adolescents’ resistance willingness to the recommendation algorithms of short video apps.

The Dual Mediating Role of Algorithmic Resistance Intention and Willingness

Social media users’ affective attitudes and willingness to act toward algorithms influence their eventual behavior of interacting with (resisting) algorithms. Previous studies concerning user responses to social media and recommendation algorithms have been conducted under the S-O-R framework (Lin et al., 2020; Ma et al., 2021), which regard users as an object that receive a stimulation, generate psychological reactions, and makes corresponding behavioral responses, without making a clear distinction between users’ behavioral awareness and behavioral willingness. However, not all social media behaviors are logical or rational (Li and Ye, 2022). As some researchers have argued in recent years, media practices are undergoing an “emotional turn” (Giaxoglou and Döveling, 2018; Orgeret, 2020), which suggests that users’ social media use is often both rational and irrational (Branley and Covey, 2018; Oh et al., 2021). Their behavioral intentions (BI) and behavioral willingness (BW) jointly shape their interactive practices with social media. In particular, since the subject of this study is a group of adolescents who are still in their cognitive maturity and whose behavioral choices are inherently characterized by the combination of rationality and irrationality, it is considered necessary to distinguish between willingness to resist and resistance, based on both the Theory of Planned Behavior (TPB) Model and the Prototype Willingness Model (PWM) and their integrated model (Rivis et al., 2006).

The mediating variable of the rational path is the resistance intention (RW) formed by adequate rational thinking, evaluation and choice. This concept relies on the Theory of Planned Behavior (TPB) proposed by Ajzen and Madden (1986). The premise of the theory of planned behavior proposed by Ajzen and Madden (1986) assumes that people’s behavioral choices are rational and based on a series of comprehensive judgments about themselves, which are centrally reflected in behavioral intentions (BI), which are generally rational and are also based on a series of comprehensive judgments about themselves. It is impacted by social norms, behavioral attitudes, and self-efficacy, which influence users’ behavioral choices (Rivis et al., 2006). Similarly, algorithmic resistance intention refers to the users’ resistance psychological reaction to the recommendation algorithm after rationally weighing their usage situation (i.e., the threat to freedom of use, dependence on the algorithm, and interpersonal communication), usage ability (i.e., algorithmic literacy) and the possible resistance outcomes, which may eventually be transformed into the practice of resistance behavior. This study assumes that resistance intention may play a mediating role in the relationship between perceived threat to freedom, algorithmic literacy, dependency psychology, and resistance behavior.

However, the mediating variable of resistance intention does not incorporate social media users’ emotional, irrational willingness to act. Hence, we used the irrational mediating variable of resistance willingness (RW) to fill this gap. Following the Prototype Willingness Model (PWM) proposed by Gibbons et al. (1998a), resistance willingness (RW) points to a relative lack of planning or premeditation and self-focused resistance psychology of users to recommendation algorithms. Although algorithms in social media interact with users all the time, it cannot be assumed that every user has a clear sense of critical reflection on them. Ettlinger (2018) also emphasizes that individual algorithmic resistance is not always based on an explicit sense of resistance after sufficient rational thought, but may also be a reflection under a herd mentality or an irrational, emotional state of mind, which is equally lead to resistance behaviors, implying that “online practices how from the unconscious desire referenced implicitly or explicitly in bleak conceptualizations of digital governance” (Ettlinger, 2018:4). In differentiation from RI, the factors influencing RW to highlight the reinforcing effect of adolescent users’ irrational imitation to peer prototypical behaviors on their willingness to resist algorithms (Gibbons et al., 1998a), Consequently, the study assumes that resistance willingness plays a mediating role in the relationship between the perceived threat to freedom, algorithmic literacy, dependent psychology, and peer influence and resistance behavior.

Notably, an integrated model based on the two forementioned theories also discussed the relationship between adolescents’ behavior willingness and behavior intentions, confirming the positive contribution of willingness to intention (Rivis et al., 2006; Frater et al., 2017). The findings have shown that the stronger the desire of adolescents to perform the behavior, the more likely they are to be stimulated to incorporate the performance of the behavior into their rational plans (van Lettow et al., 2016). Therefore, the present study similarly included the relationship between resistance willingness and resistance intention in the hypotheses. Based on the above discussion, this study constructs a conceptual model of algorithmic resistance that integrates “influencing factors-dual mediators-behavioral responses” and proposes the following three sets of hypotheses (the conceptual model is presented in Figure 1):

FIGURE 1

Conceptual model.

H8: Resistance intention mediates the relationship between (a) algorithmic literacy (b) perceived threat to freedom (c) dependent psychology and resistance behavior.

H9: Resistance willingness mediates the relationship between (a) algorithmic literacy (b) perceived threat to freedom (c) peer influence (d) dependent psychology and resistance behavior.

H10: Resistance willingness mediates the relationship between (a) algorithmic literacy (b) perceived threat to freedom (c) peer influence (d) dependent psychology and resistance behavior by facilitating resistance intention.

Research Method

Participants and Data Collections

Participants were randomly recruited from seventh to ninth grades in a local middle school, and eleventh grade in a high school, which is in the middle area of China. In order to ensuring randomness of sampling while keeping the number of randomly sampled participants in each grade to match the overall sample, stratified cluster sampling method was adopted to randomly select classes proportionally in each grade of each school to issue questionnaires. Variables such as gender and grade of the participants were considered as well. The questionnaires were uniformly distributed by the researchers to the participants to complete, with a 10–20 min filling time for each one. Two major sections are covered in the questionnaire: the personal information of the participants and the relevant measurement questions of algorithmic resistance. To ensure the validity of questionnaire completion, discriminative questions like “have you ever used short video apps like TikTok or Kuaishou?” were designed to filter invalid questionnaires. Also, the researchers supervised the whole survey process to answer any questions the participants might have. It should be noted that written informed consents were obtained from school leaders, teachers, and guardians respectively before conducting the survey. The questionnaires were completed under the principle of voluntary and anonymous, which were collected, numbered, and data entered together by after completion. In final, a total of 905 responses (response rate was 93.33%, 65 participants were excluded for the missing data on the main variables) were obtained, with an average age of 14.3 years old (121 participants in seventh grade, 137 participants in eighth grade, 145 participants in ninth grade, and 502 participants in eleventh grade).

It should be particularly explained is that all of the respondents were studying in a boarding school and were only allowed to go home for one day every three weeks during the semester, suggesting that they spent more time with peers than parents or teachers, and correspondingly, their cognition and behavior were more likely to be affected by their friends. Likewise, this kind of environment may also make their smartphones become a major media channel for them to learn and understand the outside world. According to a survey conducted by China Youth and Children Research Center on the characteristics of adolescents’ short video usage (China Youth and Children Research Center, 2021), 65.6% of adolescents regularly use short video APPs, with middle and high school students having the highest percentage of usage at 70.3%,implying that watching short videos is one of their main purposes for after-school use of smartphones. Meanwhile, compared to the recommendation algorithms of other types of social media platforms, the ones of short video APPs are unique for the flexibility to connect to known or unknown ties as well as present oneself free and creatively change usage settings (Siles et al., 2020), and is more likely to create an immersive experience due to its hyper-visual format (Salmela-Aro et al., 2017). All of these features mean the algorithmic resistance of short videos is complex and makes the study meaningful. It is also notable that while several studies in recent years have addressed the issue of adolescents’ digital media literacy (Kim and Yang, 2016; Turner et al., 2017) and the digital use divide (Peter and Valkenburg, 2006; Zhong, 2011), few studies have paid attention specifically to algorithmic literacy and algorithmic interaction practices among adolescents’ digital technology use in disadvantaged areas. Given that respondents in this study were from schools in rural China, their lower economic capital is probably impacting algorithmic cognitive and usage literacy as part of digital capital, which in turn shapes their distinctive algorithmic resistance practices. Thus, this study selected adolescent short video app users in rural areas as the research participants of algorithmic resistance to filling the above gaps.

Table 1 shows the descriptive findings. To be specific, the study conducted preliminary statistics on the basic demographic variables of the participants, mainly including their gender, age, the commonly used short video app, and the corresponding time of use. The statistical results showed that out of 905 participants, 420 (46.4%) were male and 485 were female (53.6%). The mean age of the participants was 15.53 years old, with the majority concentrated in the 12–18 age group (98.79%) and a small proportion of participants under 12 years old (0.55%) and over 18 years old (0.66%). Participants’ average weekly use time of short video apps was 13.15 h, with most of them concentrated in the range of 6–19 h (70.28%), while 9.39% less than 6 h and 20.33% more than 19 h. In addition, Bilibili, TikTok, Kuaishou, and Pear Video accounted for 96.57% of the participants’ commonly used short video APPs.

TABLE 1

Survey participants′ descriptive characteristics.

Characteristics	Category	Number	Percentage
Gender	Male	420	46.4
	Female	485	53.6
Age	<12	5	0.55
	12–15	398	43.98
	16–18	496	54.81
	>18	6	0.66
Average use time weekly	<6 h	85	9.39
	6–12 h	399	44.09
	13–19 h	237	26.19
	>19 h	184	20.33
Commonly used types	TikTok	215	23.76
	Kuaishou	210	23.2
	Bilibili	432	47.73
	Pear Video	17	1.88
	Other	31	3.43

Measurement

If not specifically stated, this study uniformly used a five-point Likert scale for the measurement of user responses, in which 1 represents strongly disagree and 5 represents strongly agree. The measured independent variables include perceived free threat, algorithmic literacy, peer influence, and dependence on algorithms.

Perceived Threat to Freedom

In particular, the scale measuring perceived threat to freedom was adapted from the psychological reactance scale developed by Hong and Page (1989). This scale was developed based on the psychological response theory proposed by Brehm (1966) and has been proved to have good reliability and explanatory power. The final scale used for measurement contains three questions (i.e., “The recommendation algorithm makes me have a sense of unfreedom when using a short video app.”).

Algorithmic Literacy

For the measurement of algorithmic literacy, the scale developed by Dogruel et al. (2021) was used, covering 11 items in two interrelated dimensions: the knowledge about algorithms and the awareness of algorithms use (i.e., “The recommendation algorithm affects the content I see,” “I can use the recommendation algorithm well to find short videos that interest me.”).

Peer Influence

The measurement of peer influence was based on the scale of peer influence on adolescents developed by Werner-Wilson and Arbel (2000), which has been shown to have good reliability. The final scale used contained a total of three questions (i.e., “My friends sometimes complain about the lack of freedom that the recommendation algorithm brings to them.”).

Dependent Psychology

Finally, the measurement of dependent psychology integrates the unified theory of acceptance and use of technology (UTAUT) (Williams et al., 2015) and the flow scale (Chang, 2013; Wang et al., 2017), including four items totally (i.e., “The recommendation algorithm has made it very convenient for me to use short video apps.”).

Resistance Willingness and Resistance Intention

The measurement of mediating variables highlighted the distinction between resistance willingness and resistance intention, and the measure of both was based on the question design of the integrated model of adolescent risk behavior by Rivis et al. (2006). Of those, three items were used to measure resistance willingness (i.e., “I sometimes urge to resist the short video recommendation algorithm”), while the measurement of resistance intention contained four items(i.e., “After careful consideration, I have the intention to turn off the recommendation algorithm notifications”).

Resistance Behavior

Since the dependent variable algorithmic resistance encompasses multiple types of resistance behavioral practices, and there is no established available scale, this study integrated both qualitative research on algorithmic folk theory (Karizat et al., 2021), research on domesticating algorithms (Siles et al., 2019b; Leong, 2020) and research on discontinues usage of social media (Luqman et al., 2017) for scale design based on a clear conceptualization of algorithmic resistance (Velkova and Kaun, 2021). The final measurement of resistance behavior was developed as a 10-item scale (i.e., “I will actively search and watch content other than short video app algorithmic recommendations.”).

Statistical Analyses and Common Method Bias Test

Considering the need to validate the complex relationships of multiple sets of variables in the study, structural equation modeling (SEM) was utilized to examine the relationships between constructs. SEM is a methodology for representing, estimating, and testing a network of relationships between variables (measured variables and latent constructs). Further model construction and data analysis were conducted using Smartpls 3.0 (Hair et al., 2019). Smart PLS 3.0 is an advanced data analysis tool to measure and assess models that can run partial least square SEM analysis (Sarstedt and Cheah, 2019). Compared to the covariance-based SEM (CB-SEM), the partial least squares SEM (PLS-SEM) provided by SmartPLS3.0 is a causal modeling approach aimed at maximizing the explained variance of the dependent latent constructs (Hair et al., 2011).

The Kaiser–Meyer–Olkin (KMO) test was used in this study to conduct tests of sample adequacy and data appropriateness. The result was 0.889, which is higher than the accepted threshold of 0.5 (Hair et al., 2019). Hence, it was considered that exploratory factor analysis could be performed on the question items and Bartlett’s test results demonstrated significant (p < 0.01). For the test of commonality method bias, Harman’s single factor test was used and the results showed that a single factor explained only 21.8% of the total variance (below the 50.0% acceptable threshold), which indicates that the data was free from CMB.

Results

Construct Reliability, Validity, and Multicollinearity Test

Table 2 shows the constructs’ loading factors, Cronbach’s alpha scores, combined reliability (CR), and average variance extracted (AVE). Among them, reliability tests include both the α coefficient test of Cronbach and the combined reliability (CR) test. According to the recommendation of Hair et al. (2019), it can be seen that the Cronbach’s α in this article are all >0.7, and the CR is <0.8, indicating that the whole measuring scale has good reliability.

TABLE 2

Sample items, means, Cronbach′s alpha scores, factor loadings, CR, AVE for each construct.

Constructs	Sample Items	Item means	factor loadings	CR	AVE	Cronbach′s α	VIF
Algorithmic Literacy	AL1	4.74	0.733	0.861	0.671	0.825	1.012
	AL2	3.45	0.741				1.547
	AL3	3.66	0.749				2.158
	AL4	3.80	0.785				2.162
	AL5	3.92	0.744				1.802
	AL6	3.90	0.758				1.711
	AL7	3.69	0.778				1.582
	AL8	3.41	0.775				1.454
	AL9	3.53	0.795				1.439
	AL10	3.68	0.703				1.419
	AL11	3.91	0.771				1.452
Perceived Threat to Freedom	PTF1	2.99	0.847	0.827	0.615	0.798	1.323
	PTF2	3.57	0.710				1.369
	PTF3	3.50	0.789				1.398
Peer Influence	PI1	2.93	0.725	0.83	0.62	0.772	1.348
	PI2	2.83	0.837				1.545
	PI3	2.82	0.796				1.305
Dependent Psychology	DP1	3.29	0.711	0.865	0.616	0.791	1.388
	DP2	3.20	0.819				1.758
	DP3	2.79	0.818				1.972
	DP4	2.87	0.786				1.737
Resistance Intention	RI1	3.02	0.777	0.896	0.684	0.845	1.706
	RI2	3.00	0.864				2.250
	RI3	3.35	0.826				1.908
	RI4	3.11	0.838				2.039
Resistance Willingness	RW1	2.94	0.816	0.869	0.689	0.774	1.573
	RW2	3.02	0.855				1.796
	RW3	3.08	0.818				1.513
Resistance Behavior	RB1	3.33	0.753	0.879	0.638	0.843	1.082
	RB2	3.07	0.712				1.383
	RB3	3.31	0.748				1.342
	RB4	2.83	0.774				2.385
	RB5	2.94	0.805				2.837
	RB6	3.05	0.827				2.951
	RB7	3.20	0.782				2.422
	RB8	3.43	0.763				1.804
	RB9	3.34	0.765				2.241
	RB10	3.33	0.774				2.020

The validity of the study was tested separately by convergence validity and differentiation validity. Among them, the average variance (AVE) of each latent variable was shown in Table 2. As we can see, each of them is >0.5, representing that each variable has a higher convergence validity. At the same time, the Fornell–Larcker criterion proposed by Fornell and Larcker (1981) and the heterotrait monotrait (HTMT) ratio of the correlations proposed by Henseler et al. (2015) were used to validate the differentiation validity of scales. The square root of the AVE value of all latent variables in Table 3 is greater than the correlation coefficient between latent variables of the diagonal, and the HTMT values shown in Table 4 range from 0.103 to 0.729 (below the threshold value 0.90), indicating that the distinguishing validity of the measurement model can be accepted. In addition, there were no items that needed to be removed as the external factor loading values of all question items were greater than 0.7 which can be accepted.

TABLE 3

Fornell–Larcker criterion.

	AL	DP	PI	PTF	RB	RI	RW
AL	0.609
DP	–0.004	0.785
PI	0.183	–0.161	0.787
PTF	0.230	–0.293	0.313	0.784
RB	0.234	–0.316	0.293	0.415	0.662
RI	0.229	–0.304	0.29	0.411	0.625	0.827
RW	0.222	–0.238	0.491	0.408	0.388	0.467	0.830

TABLE 4

Heterotrait–monotrait ratio.

	AL	DP	PI	PTF	RB	RI
AL
DP	0.103
PI	0.226	0.215
PTF	0.294	0.370	0.434
RB	0.285	0.386	0.379	0.535
RI	0.259	0.369	0.381	0.508	0.729
RW	0.253	0.304	0.661	0.525	0.478	0.578

The VIF values of the items were also examined to avoid the appearance of multicollinearity among the variables which may affect the quality of the model. The results were shown in Table 2. All VIF values ranged from 1.012 to 2.951, which is less than the threshold value of 3 suggested by Hair et al. (2019), so there is no multicollinearity.

Structural Model Test

In line with the recommendation by Preacher and Hayes (2008), 5000 bootstrap samples were generated to calculate the path coefficients and their significance. Table 5 shows the quality test results. The coefficient of determination (R2) is an important measure to assess the explanatory power of the structural model (Hair et al., 2011). The R2 value in the present study was 0.403 for resistance behavior, which means that 40.3% of changes in resistance behavior appeared due to perceived threat to freedom, algorithmic literacy, peer influence, dependent psychology, resistance willingness, and resistance intention. This value satisfies the requirement that the R2 in the field of behavior is higher than 0.2 (Hair et al., 2011). Apart from the R2 measure, this research also used q2 to evaluate the proposed model. The result showed that q2 for resistance behavior is 0.169 > 0, indicating the model has a predictive relevance (Hair et al., 2019). Meanwhile, the value is between 0.15 (medium) and 0.35 (large), which means that it has a medium predictive relevance. The other indicator to assess the goodness-of-fit indices is the standardized root mean square residual (SRMR), which can be accepted between 0 and 0.08. This study showed an adequate model fitness level with an SRMR value of 0.065.

TABLE 5

Strength of the model (Predictive relevance, coefficient of determination, and model fit indices).

	Effect size			Coefficient of determination
Construct	SSO	SSE	Q2 (=1 – SSE/SSO)	R2	Adj. R2
Resistance Behavior	9050.000	7516.927	0.169	0.403	0.402

Goodness of fifit → SRMR = 0.065; d_ULS = 3.129; d_G = 0.7282; chi-square = 3,885.264.

The measurement model and the path coefficient test results of the research model are shown in Figure 2 and Table 6. In total, the study verified 7 direct effects and 11 indirect effects. All hypotheses of the study were accepted based on the criterion (p < 0.05).

FIGURE 2

Measurement model.

TABLE 6

Summary of path coefficients and hypothesis testing.

Hypothesis	Relationship	Path coefficient	SD	t-Value	p-Value	Decision
DIRECT EFFECT
H1	AL - > RI	0.110***	0.029	3.783	0.000	Accepted
H2	AL - > RW	0.097***	0.030	3.251	0.001	Accepted
H3	PTF - > RI	0.207***	0.035	5.909	0.000	Accepted
H4	PTF - > RW	0.234***	0.035	6.718	0.000	Accepted
H5	PI - > RW	0.382***	0.035	10.934	0.000	Accepted
H6	DP - > RI	−0.167***	0.030	5.495	0.000	Accepted
H7	DP - > RW	−0.108***	0.030	3.547	0.000	Accepted
INDIRECT EFFECT
H8a	AL - > RI - > RB	0.062***	0.017	3.673	0.000	Accepted
H8b	PTF - > RI - > RB	0.117***	0.022	5.336	0.000	Accepted
H8c	DP - > RI - > RB	−0.095***	0.019	4.951	0.000	Accepted
H9a	AL - > RW - > RB	0.012*	0.005	2.401	0.017	Accepted
H9b	PTF - > RW - > RB	0.029***	0.008	3.618	0.000	Accepted
H9c	PI - > RW - > RB	0.013*	0.005	2.489	0.013	Accepted
H9d	DP - > RW - > RB	−0.048***	0.013	3.780	0.000	Accepted
H10a	AL - > RW - > RI - > RB	0.018**	0.006	3.055	0.002	Accepted
H10b	PTF - > RW - > RI - > RB	0.042***	0.008	5.381	0.000	Accepted
H10c	PI - > RW - > RI - > RB	0.069***	0.010	6.712	0.000	Accepted
H10d	DP- > RW - > RI - > RB	−0.019***	0.006	3.208	0.001	Accepted

*p-Value < 0.05, **p-Value < 0.01, ***p-Value < 0.001, t-Value > 1.96.

The direct effect validation results of the hypothesis are as follows. First, H1 was verified (β = 0.110, t = 3.783, p = 0.000), indicating that algorithmic literacy has a significant positive effect on algorithmic resistance intention. Meanwhile, the hypothesis that algorithmic literacy positively influences algorithmic resistance willingness (H2) was also verified (β = 0.097, t = 3.251, p = 0.001), and the effect size of algorithmic literacy on resistance intention is higher. Second, the study also verified the positive effect of a perceived threat to freedom on algorithmic resistance intention (β = 0.207, t = 5.909, p = 0.000), which implies that H3 was established. Likewise, the perceived threat to freedom has a significant positive effect on algorithmic resistance willingness (β = 0.234, t = 6.718, p = 0.000), and H4 was verified. Meanwhile, the empirical findings showed that the path coefficient between peer influence and resistance willingness was significant (β = 0.382, t = 10.934, p = 0.000), thus, H5 was verified. Finally, the study provided interesting findings on the negative relationship between dependent psychology and resistance intention (β = –0.167, t = 5.495, p = 0.000) and resistance willingness (β = –0.108, t = 3.547, p = 0.000) respectively, and therefore, H6 and H7 were both examined.

The mediating role of resistance willingness and intention were also confirmed, including a total of 11 hypotheses regarding indirect effects. First, the mediating role of resistance intention between algorithmic literacy (β = 0.062, t = 3.673, p = 0.000), perceived threat to freedom (β = 0.117, t = 5.336, p = 0.000), dependent psychology (β = –0.095, t = 4.951, p = 0.000) and resistance behavior were tested, therefore, H8a, H8b, and H8c can be accepted. Second, similarly, the study found the mediating role of resistance willingness between algorithmic literacy(β = 0.012, t = 2.401, p = 0.017), perceived threat to freedom(β = 0.029, t = 3.618, p = 0.000), peer influence(β = 0.013, t = 3.489, p = 0.013), dependent psychology (β = –0.048, t = 3.780, p = 0.000) and algorithmic resistance behavior, confirming H9a, H9b, H9c, H9d. Regardless of those results, due to the positive correlation between resistance willingness and resistance intention, four additional paths of influence were thus generated, supporting H10a(β = 0.018, t = 3.055, p = 0.002), H10b (β = 0.042, t = 5.381, p = 0.000), H10c (β = 0.069, t = 6.712, p = 0.000), H10d (β = –0.019, t = 3.208, p = 0.001).

Discussion

The current study reveals the relationship among the perceived threat to freedom, personal algorithmic literacy, peer influence, resistance willingness resistance intention, and algorithmic resistance practices through model construction for a sample containing 905 participants.

The results of H1(β = 0.110, t = 3.783, p = 0.000) and H2(β = 0.097, t = 3.251, p = 0.001) suggest that adolescents’ algorithmic literacy in social media is an essential factor influencing their resistance willingness and resistance intention. It means that increasing adolescents’ basic knowledge, critical awareness, and abilities to use algorithms can both facilitate their intention to enhance algorithmic resisting behavior after a rational assessment of their situation and prompt them to internalize algorithmic literacy as a form of digital capital (Lindell, 2020) used to interact more effectively with algorithms (Ku et al., 2019). On the one hand, this finding further clarifies the impact of users’ algorithmic literacy on social media use (SMU), particularly on the resistance to algorithms based on their agency. On the other hand, The technical designers of the platform algorithm should have a deeper understanding of how adolescent users’ feel and understand the algorithm through systematic analysis of their feedbacks, in order to continuously optimize the experience of using recommendation algorithms and better meet adolescents’ needs. Overall, based on previous research about the relationship between algorithmic literacy and adolescents’ media use behavior (Koenig, 2020; Gran et al., 2021), this finding illuminates the important influence of algorithmic literacy on algorithmic resistance as an interventional digital practice.

The confirmation of H3(β = 0.207, t = 5.909, p = 0.000) and H4(β = 0.234, t = 6.718, p = 0.000) implies that the potential threat from recommendation algorithms to adolescents’ perceived freedom when using short video apps may dramatically enhance both their intention and willingness to resist algorithms. While the limitations of recommendation algorithm on content visibility and technical availability elicit some resistance in both contexts, the perceived threat to freedom has a comparatively small impact on resistance intentions(β < β), probably since the impact of the algorithm’s threat is weakened when users weigh carefully the algorithm’s advantages and disadvantages, consistent with previous findings about algorithmic recommendation overload leading to users’ severe negative responses (Ma et al., 2021; Pang, 2021). The above findings suggest that platform algorithm designers should be fully aware of the important influence of users’ perceived freedom in using recommendation algorithms on algorithm resistance, and pay close attention to possible problems in both content visibility and technical affordance of recommendation algorithms. Also, through stimulating users’ motivation, actively adopting users’ feedback, and drawing experience from users’ resistance for reflection, the platform can continuously revise their recommendation systems to ultimately reduce the sense of unfreedom in the recommendation process. More significantly, this discovery extends the type of user-algorithm interaction by demonstrating that users do not necessarily respond adversely in the face of threats imposed by algorithmic recommendation systems, but will utilize algorithmic rules to resist them by developing tactics positively.

H5(β = –0.167, t = 5.495, p = 0.000) was influential in predicting the role of peers’ perceptions and behaviors on adolescents’ algorithmic resistance willingness. This indicates that the views and use of the algorithm by adolescents’ peers through everyday interactions (both verbal and behavioral) indirectly alter their perceptions, attitudes, and capacity to participate with the algorithm, which translates into their willingness to resist. It is also worth noting that resistance willingness is frequently grounded on unconscious imitation of others’ prototypical perceptions and behaviors (Geber et al., 2021), which is a unique feature of adolescents during their developmental process. This finding is also consistent with the study by et al. on the influence of adolescents’ social media use by their peers (Trivedi et al., 2021; Charmaraman et al., 2022). Hence, in line with the study conducted by Lomborg and Kapsch (2020), besides the necessary algorithmic literacy education, adolescents should be encouraged to appropriately share their knowledge, skills, and experiences about recommendation algorithms, which can contribute to a more comprehensive, diverse, and critical perception of algorithms and the acquisition of corresponding foundational usage skills.

It is also worth mentioning that the findings of H6 (β = –0.167, t = 5.495, p = 0.000) and H7 (β = –0.108, t = 3.547, p = 0.000) reveal that dependent psychology is a serious negative predictor of algorithmic resistance willingness and behavioral intention, which can be divided into two main dimensions of convenience and immersion in the process of use. This implies that when adolescent users completely experience the usefulness of the recommendation algorithm in their content access and are therefore engrossed in following the algorithm’s navigation for browsing, their intention to resist and willingness to resist are both slightly diminished. This conclusion supports prior research findings that adolescent social media addiction (Brailovskaia et al., 2015) as well as immersion lead to long-term usage (Brailovskaiaand Teichert, 2020), and extends these findings to investigate the detrimental impacts on algorithmic resistance. Therefore, on the one hand, platforms should do their part to avoid uncontrolled platform use caused by algorithms. On the other hand, society, schools and families should guide adolescents to have more social interactions to establish offline interpersonal relationships, so as to prevent them from wasting amount of time to online virtual world and finally “being trapped in the screen.”

H8 (β = 0.062, t = 3.673, p = 0.000, β = 0.117, t = 5.336, p = 0.000, β = –0.095, t = 4.951, p = 0.000), H9(β = 0.012, t = 2.401, p = 0.017, β = 0.029, t = 3.618, p = 0.000, β = 0.013, t = 3.489, p = 0.013, β = –0.048, t = 3.780, p = 0.000), and H10 (β = 0.018, t = 3.055, p = 0.002, β = 0.042, t = 5.381, p = 0.000, β = 0.069, t = 6.712, p = 0.000, β = –0.010, t = 3.208, p = 0.001) were all proven to be effective, implying that both resistance willingness and resistance intention act as a mediator between motivative factors and resistance behaviors, while adolescent willingness to resist also made a significant contribution to ultimate resistance behavior practice via intention to resist, a result consistent with the transmission relationship between BW and BI in the adolescent risk behavior conducted by other researchers (Irfan et al., 2020; Ye et al., 2020). These findings imply that resistance willingness and resistance intention can serve as separate mediators to represent rational and irrational algorithmic resistance paths respectively, while impulsive willingness to resist algorithms can also induce action intentions based on rational thinking, which are eventually jointly transformed into planned algorithmic resistance behavior. The discussion of the above results reveal the specificity of adolescents’ psychology and practice of algorithmic resistance, that is, adolescents’ rational and emotional thinking patterns are together reflected in their resistance practices, which makes them present a complex attitude as “love and hate” toward algorithmic resistance: on the one hand, the structural restrictions of the algorithm on their freedom, their trust in the perception and opinions of the algorithm by their peers will motivate them to follow the irrational path and resistance emotionally. On the other hand, when they have become accustomed to the convenience provided by the algorithm, their dependence on the recommended algorithm as well as their limited algorithmic literacy will force them to consider the realistic cost of resisting the algorithm, and finally return to the rational path of resistance intention for planned algorithmic resistance under the trade-off.

Limitations and Further Research

This study also has certain limitations in the following aspects. First, the sample of the study was obtained through an offline survey of adolescent short video app users in four grades in two rural schools in China, which may result in a relatively concentrated and unrepresentative sample for this study that does not cover other types of adolescent groups. Future research can be conducted by adopting an online survey to make the participants more diverse and to expand the sample size at the same time. Second, although the influence of psychological factors, personal ability factors, and interpersonal factors on adolescents’ algorithmic resistance in rural areas has been investigated, little is known about whether differences in personal digital capital would have an impact on their resistance practices (Ren et al., 2022). As the accumulation of digital competencies (information, communication, safety, content creation and problem solving), and digital technology (Ragnedda et al., 2020), the lack of examination of digital capital may make the model of this study incomplete, as well as neglecting the impact of digital capital on algorithmic literacy between urban and rural adolescents. So future studies should take this variable into account and explore this issue through a comparative study of adolescents’ algorithmic resistance in urban and rural areas. Third, the dependent variable of the study explored resistance behavior as a holistic object, focusing on the differences in the degree of algorithmic resistance, but lacking the distinction between different types of resistance practices, which may result in the differences between them not being distinguished effectively, however, the influences that induce diverse resistance behaviors may be quite distinct. Future research could explore the role of resistance intention and willingness for different types of resistance behaviors based on a breakdown of the types of algorithmic resistance behaviors. Finally, the study is a cross-sectional one, lacking longitudinal comparisons of resistance intentions and behaviors of the same group over time. This may make it difficult for this study to longitudinally explore the stability of algorithmic resistance psychology and behavior in the same adolescent group across time. Future research can focus on the changes in algorithmic resistance through multiple data collection and analyses.

Conclusion

This study is significant for the expansion of the existing literature in at least three aspects.

First, the study fills the gap in previous research on algorithm-human interactions that has been mostly qualitative by quantifying the concept of algorithmic resistance empirically. Concretely, the study further explores the embodiment of an emerging contemporary representation of users’ digital practices in adolescent group. The overall establishment of the model implying that algorithmic resistance, an interventional digital practice that few studies have previously addressed, is widespread among adolescent short video app users of different ages and genders in rural China.

Second, switching from the “algorithm-centric” perspective to an “algorithm-user interaction” perspective, a comprehensive model integrating psychological states, personal capabilities, and interpersonal interactions is developed to explore the different impact of various aspects on algorithmic resistance by changing the research object from the content to the interaction between the user and the recommendation algorithm. By further validation of the model, the results find that adolescents’ resistance to recommendation algorithms is simultaneously affected by both positive and negative factors. In more detail, algorithmic resistance is positively influenced by the combination of algorithmic literacy, peer influence, and perceived threat to freedom, but is weakened by their dependent psychology in the process of using recommendation algorithms. The examination of the three dimensions, covering both positive and negative factors, makes the model construction more comprehensive and systematic, compensating for the shortcomings of previous studies that focused too much on the performance of resistance behaviors instead of the influencing factors, and also implying that adolescents’ algorithmic resistance is an extremely complex process of practice based on agency.

Finally, this study develops a dual process linking the psychology of algorithmic resistance and the behavior of algorithmic resistance, based on the integration of the prototypical willingness model(PWM) and the theoretical model of planned behavior (TPB), enabling a detailed examination of the psychology of adolescents’ resistance in two different states of rationality. At the same time, the interaction between the two paths was verified. The results show that adolescent users’ algorithmic resistance practices follow both rational and irrational paths. The rational path is manifested by adolescents’ behavioral intentions based on the comprehensive evaluation of their algorithm literacy, psychological condition (containing both positive and negative ones), which in turn influences their resistance practices toward algorithms. The irrational path, on the other hand, is more significantly affected by adolescents’ peers and also affected by algorithm literacy and psychological perception to generate resistance intention, which consequently shapes their resistance practice. In addition, the resistance willingness may positively influence the resistance intention, further facilitating the implementation of adolescents’ algorithmic resistance practice behaviors. The distinction between willingness to resist and intention to resist in two different rational states of resistance not only breaks away from the single psychological variable design in previous studies, but also fits better with the mindset of adolescents, which has significant meaning for reminding later researchers to pay more attention to the double-side of adolescents’ psychological and behavioral models in the process of media interaction.

Data Availability Statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author/s.

Ethics Statement

The studies involving human participants were reviewed and approved by Renmin University of China. Written informed consent to participate in this study was provided by the participants’ legal guardian/next of kin. Written informed consent was obtained from the individual(s), and minor(s)’ legal guardian/next of kin, for the publication of any potentially identifiable images or data included in this article.

Author Contributions

XL wrote the manuscript. YC and WG collected the data and analyze it and revised the manuscript. All authors contributed to the article and approved the submitted version.

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s Note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.