Overconfident youth,underconfident elders: Age differences in algorithmic knowledge confidence and information perceptions

Abstract

Age-related differences in algorithmic knowledge are typically interpreted as competence gaps that render younger “digital natives” better equipped and older users more vulnerable in algorithmically curated environments. This study challenges these assumptions by examining (a) algorithmic knowledge confidence—whether users’ subjective algorithmic knowledge exceeds or falls short of their algorithmic awareness and knowledge—and (b) what consequences such misalignment has for information environment perceptions. Using US survey data (N = 1205), this study found systematic misalignment between subjective and awareness-based algorithmic knowledge: younger users overestimate their understanding, whereas older users underestimate theirs. Social media use partially mediated the age–knowledge confidence relationship. Knowledge confidence, in turn, predicted perceived information reliability and diversity. These findings reveal that critical assessment of information environment is determined not by knowledge level alone but by alignment between subjective and awareness-based knowledge, highlighting the need for age-tailored interventions addressing knowledge overconfidence and underconfidence.

Keywords

algorithmic knowledge knowledge confidence digital natives older adults social media information perceptions

Introduction

Algorithms have become central gatekeepers of information in digital societies (Wallace, 2018). From news feeds to content recommendations, algorithmic systems determine what information billions of users encounter daily on social media. This has sparked growing concern about whether users possess adequate knowledge to navigate these systems critically and recognize potential biases or manipulations.

Research on algorithmic knowledge has increased in response, examining how users understand algorithmic processes and how such knowledge shapes information behaviors. While studies have documented important relationships between algorithmic knowledge and critical information awareness or behavior (e.g. Chung and Wihbey, 2024; Gran et al., 2021; Oeldorf-Hirsch and Neubaum, 2023), it has predominantly focused on knowledge levels alone (i.e. how much users know or believe they know). Yet a fundamental question remains underexplored: do users accurately assess what they know? Across domains from consumer behavior to scientific debates to political knowledge, substantial evidence shows that perceived and actual knowledge frequently diverge (Alba and Hutchinson, 2000; Fischer et al., 2019; Lee et al., 2022). Such misalignment—the gap between what people think they know and what they actually know—has meaningful consequences for decision-making and information processing (Fischer et al., 2023; Fischer and Fleming, 2024). Whether similar misalignment exists for algorithmic knowledge, and what consequences it carries, remains largely unknown.

This gap is particularly notable given persistent age-related patterns in reported algorithmic knowledge. Younger users consistently report higher algorithmic knowledge than older users (Chung and Wihbey, 2024; Gran et al., 2021), a pattern often interpreted as evidence of superior “digital native” competence. However, growing evidence challenges this interpretation: younger users’ higher knowledge scores do not necessarily translate into more critical information behaviors (Chung, 2025; Wineburg and McGrew, 2019), while older users are more vigilant about the credibility of online information than many assume (Baines et al., 2025; Seo et al., 2021). These puzzles suggest that reported age differences in algorithmic knowledge may reflect how users assess their knowledge rather than what they demonstrably know. Moreover, little research has examined how users’ knowledge assessment shapes their perceptions of information environments. If users systematically overestimate or underestimate their algorithmic understanding, this may bias their evaluation of information quality, diversity, and risks—yet these relationships remain unexplored.

The present study addresses these gaps by examining algorithmic knowledge confidence—the degree to which users’ subjective algorithmic knowledge exceeds or falls short of their algorithmic awareness and knowledge—across age groups and its consequences for information environment perceptions. Specifically, we investigate whether age-related differences in reported algorithmic knowledge reflect systematic patterns of overconfidence and underconfidence, what mechanisms drive these patterns, and how algorithmic knowledge confidence shapes perceptions of information reliability and diversity on social media. By distinguishing what users believe they know from what they demonstrably know, this study provides a more precise account of age-related differences in algorithmic knowledge and reveals how metacognitive calibration shapes users’ trust in and assessments of algorithmically curated information environments.

Algorithmic knowledge: Subjective versus objective

Prior research on algorithmic knowledge has approached the concept in diverse ways. Early work conceptualized algorithmic knowledge primarily as awareness; recognizing that algorithms curate content (e.g. Eslami et al., 2015). More recent studies have expanded this to include understanding of how algorithms function and the ability to critically evaluate algorithmic outputs (Chung, 2025; Gran et al., 2021; Zarouali et al., 2021). Despite this diversity in conceptualization, most scholarly work has operationalized algorithmic knowledge as a cognitive ability—what users know or believe they know about algorithmic systems.

However, a critical distinction that has received limited attention in algorithmic knowledge research is the separation between subjective and objective knowledge.

This distinction, well established in other domains such as consumer behavior and political communication (Carlson et al., 2009; Lee et al., 2022), differentiates between what individuals think they know and what they actually know. Subjective knowledge refers to individuals’ self-assessed beliefs about how well they understand a particular subject matter (Carlson et al., 2009), while objective knowledge reflects accurate information stored in memory, typically measured through the ability to correctly answer knowledge questions (Brucks, 1985). Drawing on this broader literature, we define subjective algorithmic knowledge as users’ self-assessed understanding of how algorithms select, rank, and personalize content, and objective algorithmic knowledge as users’ ability to accurately evaluate statements about algorithmic mechanisms and their implications.

Although these two forms of knowledge are often correlated, extensive research demonstrates that they can diverge substantially at the individual level, particularly when systems are complex, opaque, and difficult to verify (Alba and Hutchinson, 2000). For example, research on political knowledge reveals that social media news use increases users’ subjective knowledge without improving their actual political knowledge (Lee et al., 2022). Similarly, studies of privacy literacy among college students show no significant correlation between self-assessed and objective privacy literacy, with systematic overconfidence observed (Li et al., 2023). These patterns suggest that in algorithmically mediated environments characterized by opacity and limited feedback, subjective and objective algorithmic knowledge may diverge as well.

Algorithmic knowledge gaps by age

Existing research consistently documents age-related patterns, with younger users reporting higher algorithmic knowledge than older users (Chung and Wihbey, 2024; Cotter and Reisdorf, 2020; Gran et al., 2021). However, these studies have typically relied on single measurement approaches—either subjective self-assessments or objective evaluations—without systematically examining both dimensions simultaneously or investigating their calibration. As a result, whether reported age differences reflect actual competence gaps, self-assessment biases, or both remains unclear.

This ambiguity is consequential. The consistent pattern of higher reported algorithmic knowledge among younger users has been widely interpreted to support two complementary assumptions: first, that younger “digital natives” are better equipped to navigate algorithmic systems and therefore less vulnerable to algorithmic manipulation or misinformation; second, that older users’ lower reported knowledge signals deficient understanding, rendering them more vulnerable in algorithmically curated environments.

However, both assumptions confront empirical puzzles that suggest the relationship between age and algorithmic knowledge may be more complex than it appears. Studies revealed that younger users demonstrate persistent difficulties in evaluating online information critically; two thirds of US high school students cannot distinguish news articles from sponsored content (Breakstone et al., 2021). In the algorithmic context specifically, higher algorithmic knowledge scores among young adults do not consistently translate into more critical information behaviors (Chung, 2025). If younger users genuinely possess superior algorithmic understanding, why does this advantage not manifest in more discerning engagement?

Similarly, the assumption of older adults’ deficient understanding warrants scrutiny. Research on aging and metacognition indicates that metacognitive monitoring accuracy is generally preserved or even enhanced in older adulthood across many cognitive domains (Hertzog and Dunlosky, 2011). Studies also revealed that older users demonstrate great vigilance regarding the credibility of online information (Baines et al., 2025; Seo et al., 2021). These findings suggest that age-related patterns in reported algorithmic knowledge may not straightforwardly reflect actual understanding differences. Instead, they point to the possibility that users across the age spectrum may assess their competence differently—systematic differences in knowledge confidence—independent of their actual knowledge levels.

Exposure, processing fluency, and confidence

We propose that age-related puzzles can be explained by differential exposure to algorithmically curated content and its metacognitive consequences. The mechanism operates through processing fluency—the subjective ease with which information is processed (Alter and Oppenheimer, 2009). Younger users engage with algorithm-intensive platforms at substantially higher rates: 95% of 18- to 29-year-olds use YouTube and approximately 63% use TikTok daily, compared to 64% and 12% among users 65 and older (Pew Research Center, 2025). This intensive and repeated exposure may create heightened processing fluency for algorithmic content.

Research demonstrates that such fluency is often misattributed to genuine understanding even when actual knowledge remains limited (Koriat and Bjork, 2005). Across multiple domains, fluent processing inflates perceived knowledge: television news viewers overestimate comprehension when production features enhance fluency (Schneider and Schwarz, 2017), and students overestimate learning from fluent video lectures despite poor transfer performance (Carpenter and Geller, 2020). In algorithmic environments, frequent interaction with algorithmically generated content may similarly foster an illusion of knowing (Glenberg et al., 1982), a sense of understanding that exceeds factual grasp of algorithmic mechanisms.

Conversely, older users’ substantially lower engagement with algorithmically curated platforms provides fewer opportunities to develop fluency-based cues of understanding. As a result, they may be less likely to experience the subjective ease that often fosters confidence in one’s knowledge. Without such fluency cues, older users may rely on the absence of familiarity as a signal of limited understanding, leading them to judge their algorithmic knowledge more conservatively. In this sense, just as younger users may mistake familiarity for understanding, older users may mistake unfamiliarity for incompetence.

Building on this logic, this study conceptualizes these patterns in terms of algorithmic knowledge confidence—the degree to which subjective algorithmic knowledge exceeds or falls short of objective algorithmic knowledge. When individuals’ self-assessed understanding exceeds their actual understanding, they exhibit algorithmic overconfidence; when self-assessed understanding falls short, they exhibit algorithmic underconfidence. Importantly, knowledge confidence should not be interpreted as a stable personality trait but as a context-dependent metacognitive phenomenon shaped by exposure, experience, and available cues for self-assessment. Accordingly, we propose the first hypothesis:

H1: Age will be negatively associated with algorithmic knowledge confidence, such that younger users are more likely to exhibit overconfidence and older users are more likely to exhibit underconfidence.

While H1 proposes a relationship between age and algorithmic knowledge confidence, the proximate mechanism driving this relationship could be social media use. As the primary context in which users encounter algorithmically curated content, social media platforms serve as the pathway through which age translates into differential algorithmic exposure. Given the strong age gradient in platform engagement documented above, we predict:

H2: Social media use will mediate the relationship between age and algorithmic knowledge confidence.

Consequences of knowledge confidence

If younger users exhibit algorithmic knowledge overconfidence while older users demonstrate algorithmic knowledge underconfidence (H1), and if these miscalibrations are driven by differential social media use (H2), what are the downstream consequences for how users perceive the social media information environment? We propose that algorithmic knowledge confidence systematically shapes two key dimensions of information environment perceptions: assessments of information reliability and diversity on social media.

These two dimensions represent complementary aspects of how users evaluate algorithmically curated information environments. Reliability perceptions reflect users’ trust in the quality and accuracy of information encountered on social media, while diversity perceptions reflect users’ beliefs about the breadth and balance of perspectives to which algorithms expose them. Together, these perceptions constitute users’ overall assessment of the information environment’s fitness for informed decision-making.

Theoretically, we conceptualize algorithmic knowledge confidence as an antecedent metacognitive condition shaping how users evaluate their social media information environment. First, we expect that algorithmic overconfidence may increase perceived reliability: when users overestimate their understanding of how algorithms curate content, they may perceive algorithmic outputs as more predictable and manageable, reducing epistemic vigilance. Research shows that perceived control reduces systematic processing and increases reliance on heuristic cues (Tiedens and Linton, 2001), and algorithmic explanations can enhance users’ sense of understanding and control (Rader et al., 2018). Overconfident users who believe they possess adequate understanding may therefore evaluate social media information as reliable even without critical scrutiny.

Conversely, algorithmic underconfidence may decrease perceived information reliability. Users who underestimate their understanding may experience heightened uncertainty about their ability to navigate algorithmic systems effectively. This perceived lack of competence can amplify skepticism and reduce trust, as individuals question whether they can adequately assess information quality when they believe they lack sufficient algorithmic knowledge. Research on self-efficacy indicates that low confidence in one’s capabilities increases perceived vulnerability and risk aversion (Bandura, 1997). Applied to algorithmic contexts, underconfident users may approach social media information with greater caution and lower trust, thereby reducing perceived reliability.

H3a: Algorithmic overconfidence will be positively associated with perceived reliability of information on social media.

H3b: Algorithmic underconfidence will be negatively associated with perceived reliability of information on social media.

Algorithmic overconfidence may also inflate perceptions of information diversity. Users who overestimate their algorithmic understanding may mistakenly believe that algorithms are designed to expose them to diverse perspectives, failing to recognize how personalization and engagement optimization narrow their feeds. Research suggests that users may hold implicit or idealized beliefs about the rationality and objectivity of algorithms, expecting them to act fairly and without bias despite limited transparency (Beer, 2017; Bucher, 2017). Such beliefs may lead overconfident users to infer that algorithmic curation promotes exposure to diverse viewpoints, attributing perceived diversity to algorithmic design rather than recognizing structural constraints.

In contrast, algorithmic underconfidence may deflate perceptions of information diversity through its effects on engagement with the information environment. Users who doubt their algorithmic understanding may experience lower self-efficacy, which research has shown to reduce motivation, persistence, and active engagement in learning contexts (Zimmerman, 2000). In algorithmically curated environments, such reduced engagement may translate into less exploratory interaction with content, resulting in narrower exposure that users may perceive as limited diversity. Alternatively, uncertainty about algorithmic functioning may heighten sensitivity to structural constraints, leading underconfident users to perceive information diversity as low even when their exposure is comparable to that of more confident users.

H4a: Algorithmic overconfidence will be positively associated with perceived informational diversity on social media.

H4b: Algorithmic underconfidence will be negatively associated with perceived informational diversity on social media.

Together, these hypotheses suggest that algorithmic knowledge confidence shapes how users perceive the information environment across two dimensions. Given that age predicts algorithmic knowledge confidence (H1) through social media use (H2), and such knowledge confidence predicts information environment perceptions (H3, H4), it follows that social media use and algorithmic knowledge confidence should serially mediate the relationship between age and these perceptions:

H5: The relationship between age and (a) perceived information reliability and (b) perceived information diversity will be serially mediated through social media use and algorithmic knowledge confidence.

Methods

Sample

A national survey of US adults (N = 1205, Female 51.5%, M age = 46.16, SD = 15.73) was conducted via an online survey platform Prolific in May 2025. Prolific was selected as the recruitment platform due to its demonstrated data quality advantages over other online panels. Comparative studies have found that Prolific consistently delivers higher-quality, more attentive responses than other widely used platforms, including Amazon Mechanical Turk and Qualtrics, across key metrics such as attention check compliance, response validity, and participant engagement (Esch et al., 2025; Peer et al., 2022). A non-probability quota sample was used; participants were recruited based on stratified quotas for three demographic characteristics (sex, age, and ethnicity) to approximate the US population distribution as reported by the 2021 American Community Survey (ACS). No regional quotas were applied. The sample for this study had a median age of 46 (39.2 in the ACS), with 51.5% being female (50.5% in the ACS), and a median income range of $75,000 to $99,999 ($77,719 in the ACS). Our data underrepresented Hispanics (8.0% vs 18.8% in the ACS) and non-Hispanic Whites (66.4% vs 74.8% in the ACS), but proportionately represented Black (13.4% vs 13.7%) and Asian (6.5% vs 6.7%) populations. The remaining 5.7% identified as other racial/ethnic groups.

Measures

Subjective algorithmic knowledge was measured using a 5-point Likert-type scale, in which participants indicated the extent to which they believed they understood social media algorithms. The items were designed to correspond to the five dimensions of algorithmic awareness and knowledge (Chung, 2025; Zarouali et al., 2021). Example items included: “I am able to recognize the presence of algorithms in social media feeds,” “I understand the types of data that social media algorithms use to recommend content to me,” and “I understand how algorithmic decisions are made on social media” (1 = Strongly disagree, 5 = Strongly agree). Responses were averaged to create a composite index (M = 3.81, SD = 0.76, α = .85).

Algorithmic awareness and knowledge was measured with Zarouali et al.’s (2021) Algorithmic Media Content Awareness (AMCA) scale, a validated instrument that allows researchers to assess users’ understanding of algorithms in online platforms in a standardized way (Chung, 2025; Oeldorf-Hirsch and Neubaum, 2023; Voorveld et al., 2024). The AMCA scale was used as an awareness-based proxy for objective algorithmic knowledge.¹ This scale measured understanding of social media algorithms across four dimensions, such as content filtering, automated decisions, human-algorithm interplay, and ethical considerations. We also added a fifth dimension to capture motivations behind algorithms, as called for in Zarouali et al. (2021). The overall scale is reliable (M = 3.95, SD = 0.67, α = .83), as are sub-dimensions (α = .77–.83). The 16 measurement items are presented in Table 1, organized by dimension.

Table 1.

Questions to measure algorithmic awareness and knowledge.

Dimension	Items
	Please indicate to what extent you are aware of the following statements about algorithms on social media.
Content filtering(1 = Not aware at all,5 = Completely aware)	1. Algorithms are used to recommend content to me on social media.2. Algorithms are used to prioritize certain content over others.3. Algorithms are used to tailor certain content to me on social media.4. Algorithms are used to show someone else different content than I get to see on social media.
Automated decision	1. Algorithms are used to show me content on social media based on automated decisions.2. Algorithms do not require human judgments in deciding which content to show me on social media.3. Algorithms make automated decisions on what content I get to see on social media.
Human-algorithm interplay	1. The content algorithms recommend to me on social media depend on my online behavior on that platform.2. The content algorithms recommend to me on social media depend on my online behavioral data.
	3. The content algorithms recommend to me on social media depend on the data that I make available online.
Ethical considerations	1. It is not always transparent why algorithms decide to show me certain content on social media.2. The content algorithms recommend to me on social media can be subjected to human biases such as prejudices and stereotypes.3. Algorithms use my personal data to recommend certain content on social media, and this has consequences for my online privacy.
Motivations behindalgorithms	1. Social media platforms use algorithms to induce clicking behavior and user engagement.2. Algorithms are designed to promote content that serves the social media platform’s business interests.3. The content I see on social media is prioritized based on what benefits the platform financially or strategically.

Algorithmic knowledge confidence was operationalized with a residual-based approach. Specifically, subjective algorithmic knowledge scores were regressed on algorithmic awareness and knowledge scores, and the unstandardized residuals were saved as the confidence index (M = 0.00, SD = 0.73). Positive residuals indicate that a respondent rated their subjective knowledge higher than would be expected given their algorithmic awareness and knowledge level (overconfidence), while negative residuals indicate the reverse (underconfidence). This approach was adopted in preference to a simple difference score, as it accounts for the empirical relationship between the two components and avoids the psychometric limitations associated with standardized subtraction-based indices.

Perceived reliability of social media content was measured by asking respondents how confident they are that the social media content shows (a) factual, true, or accurate information and (b) balanced and objective (modified from Reisdorf and Blank, 2021;1 = Very doubtful, 5 = Very confident, M = 2.86, SD = 1.09, r = .81, p < .001).

Perceived information diversity was measured by asking respondents the extent to which they believe the content recommendation system on social media provides them with (a) news or information that helps them discover new perspectives they wouldn’t have found elsewhere and (b) news or information that presents opinions and worldviews that are different from mine (modified from Matt et al., 2014; 1 = None at all, 5 = A great deal, M = 3.06, SD = 1.01, r = .70, p < .001).

Social media use was measured by asking respondents how often they use the following social media platforms in a typical day: Facebook, Instagram, LinkedIn, Pinterest, Reddit, Signal, Snapchat, TikTok, Twitter, WhatsApp, and YouTube. Scores for each platform were averaged to create a composite social media use score (Oeldorf-Hirsch and Neubaum, 2023; 1 = Never, 5 = Several times a day, M = 2.89, SD = 0.84, α = .83).

Results

Analytical approach

All hypotheses were tested using SPSS 29.0. Simple and multiple linear regression analyses were used to examine direct relationships, and one-way analysis of variance (ANOVA) with Tukey Honestly Significant Difference (HSD) post hoc comparisons was conducted to assess age group differences. Mediation and serial mediation hypotheses were tested using PROCESS macro (Hayes, 2022), with 5000 bootstrap samples and 95% bias-corrected confidence intervals. Although the study was not preregistered, the data, survey materials, and analysis code are publicly available on the Open Science Framework (OSF; https://doi.org/10.17605/OSF.IO/K2F68).

Hypotheses test

H1 predicted a negative association between age and algorithmic knowledge confidence. Simple linear regression confirmed this hypothesis (β = -.220, p < .001, R² = .048, F(1, 1201) = 61.154, p < .001). A polynomial regression including a quadratic term did not improve model fit (ΔR² = .000, F change = .501, p = .479), confirming linear specification. One-way ANOVA further indicated significant differences across age groups, F(5, 1197) = 12.846, p < .001, η² = .051. Table 2 presents descriptive statistics by age group, revealing systematic variation from overconfidence among younger users to underconfidence among older users, and Figure 1 presents systematic variation from overconfidence among younger users to underconfidence among older users. Post hoc comparisons (Tukey HSD) revealed that younger groups (ages 18–44) exhibited significantly greater overconfidence than older groups (ages 45+), with the largest difference between the 25–34 and 65+ groups (M_diff = .45, p < .001). Notably, differences within the younger cluster (Age groups 1–3) and within the older cluster (Age groups 4–6) were non-significant (ps > .45), indicating homogeneous overconfidence and underconfidence regions with a gradual transition between them (see Table 2 for full pairwise comparisons). These convergent findings across analytical approaches support H1.

Table 2.

Algorithmic knowledge by age group: Descriptive statistics and pairwise comparisons.

Age group	Subjective knowledge	Algorithmic awareness/knowledge	Knowledge confidence	Pairwise comparisons (confidence mean difference)
Age group	M (SD)	M (SD)	M (SD)	1	2	3	4	5
1. 18–24	3.99 (0.64)	3.88 (0.70)	0.19 (0.61)
2. 25–34	4.04 (0.65)	3.97 (0.61)	0.22 (0.65)	-.02
3. 35–44	3.91 (0.71)	3.95 (0.63)	0.09 (0.69)	.10	.13
4. 45–54	3.76 (0.75)	3.98 (0.64)	-0.07 (0.73)	.26*	.28**	.16
5. 54–65	3.68 (0.81)	3.97 (0.71)	-0.14 (0.77)	.34***	.36***	.23**	.08
6. 65+	3.56 (0.83)	3.89 (0.73)	-0.24 (0.78)	.43***	.45***	.33***	.17	.10

Note. Subjective knowledge and algorithmic awareness and knowledge measured on 1–5 scale. Knowledge confidence = residual from regressing subjective algorithmic knowledge on algorithmic awareness and knowledge. Positive values indicate overconfidence; negative values indicate underconfidence. Lower triangle shows mean differences (Row-Column) from Tukey HSD post hoc tests.

p < .05, **p < .01, ***p < .001.

H2 predicted that social media use would mediate the relationship between age and algorithmic knowledge confidence. We tested this hypothesis using PROCESS Model 4 (Hayes, 2022) with 5000 bootstrap samples. Results revealed significant effects for both component paths. Age negatively predicted social media use (β = -.345, SE = 0.014, t = -12.73, p < .001, 95% CI [-0.21, -0.15]), and social media use positively predicted algorithmic knowledge confidence after controlling for age (β = .347, SE = 0.025, t = 12.28, p < .001, 95% CI [0.25, 0.35]). The total effect of age on algorithmic knowledge confidence was significant (β = -.220, p < .001), and the direct effect remained significant when controlling for social media use (β = -.100, SE = 0.013, p < .001, 95% CI [-0.07, -0.02]). The standardized indirect effect through social media use was significant (β = -.120, Bootstrap 95% CI [-0.15, -0.10]). These results indicate partial mediation, demonstrating that age differences in algorithmic knowledge confidence are partly driven by differential patterns of social media use. Hence, H2 is supported.

H3a and H3b predicted directional relationships between algorithmic knowledge confidence and perceived information reliability on social media. Simple linear regression supported these hypotheses: algorithmic knowledge confidence significantly predicted perceived information reliability (β = .309, SE = 0.041, t = 11.25, p < .001, R² = .095, F(1, 1201) = 126.53, p < .001). The positive coefficient indicates that overconfident users perceived social media information as more reliable, while underconfident users perceived it as less reliable. This relationship accounted for 9.5% of the variance in information reliability perceptions. To ensure this effect was not confounded by age or social media use, we conducted multiple regression including these variables as controls. Algorithmic knowledge confidence remained a significant predictor (β = .177, SE = 0.042, t = 6.29, p < .001), even after controlling for age (β = -.018, SE = 0.019, p = .519) and social media use (β = .334, SE = 0.038, p < .001). The full model explained 19.4% of variance in reliability perceptions (R² = .194, F(3, 1199) = 96.069, p < .001), representing a 10 percentage point increase over the simple model (ΔR² = .098, p < .001). These findings support both H3a and H3b, demonstrating that algorithmic knowledge confidence shapes information reliability perceptions independent of age and social media usage factors. The robust effect of algorithmic knowledge confidence (β = .177) even when controlling for social media use—which itself showed a strong effect (β = .334)—indicates that both metacognitive confidence and actual platform engagement independently contribute to users’ trust in social media information.

H4a and H4b predicted directional relationships between algorithmic knowledge confidence and perceived information diversity. A simple linear regression supported these hypotheses. Algorithmic knowledge confidence significantly predicted perceived informational diversity (β = .322, SE = 0.038, t = 11.81, p < .001, R² = .104, F(1, 1202) = 139.429, p < .001). The positive coefficient indicates that overconfident users perceived social media as exposing them to more diverse viewpoints and balanced perspectives, while underconfident users perceived less information diversity. This single relationship accounted for 10.4% of the variance in diversity perceptions, demonstrating that algorithmic knowledge confidence shapes users’ assessments of the breadth and balance of information they encounter on social media platforms.

Multiple regression including age and social media use as controls revealed a more complex pattern. Algorithmic knowledge confidence remained a significant positive predictor (β = .150, SE = 0.037, t = 5.58, p < .001), supporting H4a and H4b. However, social media use emerged as the strongest predictor of information diversity perceptions (β = .425, SE = 0.033, t = 15.25, p < .001), with age showing a non-significant direct effect (β = -.047, SE = 0.017, t = -1.78, p = .075). The full model explained 27.1% of variance in diversity perceptions (R² = .271, F(3, 1199) = 148.394, p < .001), representing a 16.7 percentage point increase over the simple model (ΔR² = .167, p < .001). These findings suggest that perceived information diversity is driven more strongly by behavioral patterns of platform engagement than by metacognitive assessments of algorithmic understanding.

H5a and H5b predicted that the relationship between age and information perceptions would be serially mediated through social media use and algorithmic knowledge confidence. We tested these hypotheses using PROCESS Model 6 (Hayes, 2022) with 5000 bootstrap samples. For perceived information reliability (H5a), age negatively predicted social media use (β = -.345, SE = 0.014, t = -12.73, p < .001, 95% CI [-0.21, -0.15]), and social media use positively predicted algorithmic knowledge confidence after controlling for age (β = .347, SE = 0.025, t = 12.28, p < .001, 95% CI [0.25, 0.35]). Algorithmic knowledge confidence positively predicted perceived information reliability after controlling for age and social media use (β = .177, SE = 0.042, t = 6.29, p < .001, 95% CI [0.18, 0.35]), and social media use also directly predicted perceived reliability (β = .334, SE = 0.038, t = 11.38, p < .001, 95% CI [0.36, 0.50]). The total effect of age on perceived information reliability was significant (β = -.172, SE = 0.019, t = -6.05, p < .001, 95% CI [-0.15, -0.08]), but the direct effect was reduced to non-significance (β = -.018, SE = 0.019, t = -0.646, p = .519, 95% CI [-0.05, 0.02]), indicating full mediation. The total indirect effect was significant (β = -.154, Bootstrap 95% CI [-0.18, −0.13]), with the serial indirect pathway (age → social media use → algorithmic knowledge confidence → reliability) also significant (β = -.021, Bootstrap 95% CI [-0.03, -0.01]). In addition, the indirect pathways through social media use alone (β = -.115, Bootstrap 95% CI [-0.14, -0.09]) and algorithmic knowledge confidence alone (β = -.018, Bootstrap 95% CI [-0.03, -0.01]) were significant, respectively. Figure 2 demonstrates the serial mediation paths. These findings support H5a and indicate that age effects on information reliability perceptions operate through multiple pathways. While the dominant pathway runs through social media use (β = -.115), knowledge confidence also independently mediated the relationship (β = -.018), and the serial pathway through both social media use and knowledge confidence was significant (β = -.021). These findings confirm that age differences in information reliability perceptions are largely explained by differential patterns of social media engagement and the knowledge confidence such engagement generates, with younger users’ greater platform use fostering overconfidence that inflates perceived reliability, and older users’ lower engagement contributing to underconfidence that reduces it.

For perceived informational diversity (H5b), the serial mediation pattern was similar. Age negatively predicted social media use (β = -.345, SE = 0.014, t = -12.73, p < .001), and social media use positively predicted algorithmic knowledge confidence (β = .347, SE = 0.025, t = 12.28, p < .001). Algorithmic knowledge confidence positively predicted perceived information diversity after controlling for age and social media use (β = .150, SE = 0.037, t = 5.58, p < .001, 95% CI [0.13, 0.28]), and social media use showed a strong direct effect on information diversity perceptions (β = .425, SE = 0.033, t = 15.25, p < .001, 95% CI [0.44, 0.57]). The total effect of age on perceived information diversity was significant (β = -.227, SE = 0.018, t = -8.06, p < .001, 95% CI [-0.18, -0.11]), with a non-significant direct effect (β = -.047, SE = 0.017, t = -1.78, p = .075, 95% CI [-0.06, 0.00]), indicating full mediation. The total indirect effect was significant (β = -.180, Bootstrap 95% CI [-0.21, -0.15]), with the serial indirect pathway (age → social media use → algorithmic knowledge confidence → diversity) also significant (β = -.018, Bootstrap 95% CI [-.03, -.01]). The indirect pathway through social media use alone was dominant (β = -.147, Bootstrap 95% CI [-0.18, -0.12]), while the indirect pathway through knowledge confidence alone was also significant (β = -.015, Bootstrap 95% CI [-0.03, -0.01]). Figure 3 demonstrates the serial mediation paths. These findings support H5b, demonstrating that age effects on information diversity perceptions also operate through multiple pathways. While the dominant pathway runs through social media use, algorithmic knowledge confidence also independently mediates the relationship, with social media use playing a somewhat stronger mediating role in shaping information diversity perceptions (β = -.147) than information reliability perceptions (β = -.115).

Discussion

This study examined age-related differences in algorithmic knowledge through the lens of metacognition and its consequences. Results revealed systematic patterns of knowledge overconfidence and underconfidence: younger users overestimated their understanding while older users underestimated theirs, with social media use partially mediating this relationship. Algorithmic knowledge confidence, in turn, shaped perceptions of information reliability and diversity on social media. These findings demonstrate that previously reported age-related differences in algorithmic knowledge reflect differences in knowledge confidence driven in part by differential exposure to social media rather than actual competence gaps.

Theoretical and practical implications

The first key contribution of this study is a reconceptualization of age-related differences in algorithmic knowledge. Prior research has consistently documented that younger users report higher algorithmic knowledge than older users (e.g. Chung and Wihbey, 2024; Gran et al., 2021), a pattern often interpreted as evidence that “digital natives” possess superior understanding of algorithmic systems. Our results challenge this interpretation. By distinguishing subjective algorithmic knowledge from algorithmic awareness and knowledge and examining the discrepancy between them, this study demonstrates that reported age differences primarily reflect differences in knowledge confidence—specifically, overconfidence among younger users and underconfidence among older users—rather than actual competence disparities. This approach shifts attention from knowledge as a static resource to knowledge as a metacognitive judgment, emphasizing the alignment—or misalignment—between how much individuals believe they know and how much they demonstrate knowing through awareness-based measures.

Importantly, descriptive statistics in Table 2 reveal that this pattern of knowledge overconfidence and underconfidence emerges from age-related differences in subjective knowledge rather than awareness-based knowledge. Algorithmic awareness and knowledge remained stable across all age groups (p = .641), indicating no actual competence gap. In contrast, subjective knowledge declined systematically with age, F(5, 1197) = 11.995, p < .001, ranging from M = 3.99–4.04 among younger users (18–34 years) to M = 3.56–3.68 among older users (55+ years). This pattern demonstrates that younger users’ apparent knowledge advantage reflects inflated self-assessment rather than superior understanding, while older users’ lower reported knowledge reflects underestimation rather than deficient competence.

Specifically, younger users’ overconfidence appears to stem from processing fluency generated by intensive exposure to algorithmic content on social media. The finding that social media use partially mediates the age–knowledge confidence relationship corroborates this exposure–fluency–confidence pathway. Frequent interaction with algorithmically curated feeds, recommendations, and personalized content on social media creates a subjective sense of familiarity that individuals may misattribute to genuine understanding—a pattern consistent with research on the illusion of knowing (Glenberg et al., 1982; Koriat and Bjork, 2005; Schneider and Schwarz, 2017). Conversely, older users’ underconfidence reflects not a lack of metacognitive ability, but the absence of such fluency-based confidence inflation.

This finding fundamentally challenges the digital native narrative, which conflates confidence with competence (Kirschner and De Bruyckere, 2017). Rather than viewing algorithmic knowledge as a skill that younger generations naturally possess and older generations lack, our results suggest that frequent technology use generates fluency-based confidence without corresponding comprehension. Familiarity with algorithmic outputs does not equate to understanding of algorithmic mechanisms, and mistaking exposure for expertise may lead to misguided assessments of users’ capabilities. Our confidence framework makes this distinction explicit, offering a more nuanced understanding of what it means to be algorithmically literate.

Notably, the 25–34 age group exhibited the greatest overconfidence relative to their awareness-based knowledge level, surpassing even the youngest group (18–24). This pattern may reflect a combination of sustained exposure to social media environments and accumulated real-world experience. Adults in their mid-twenties to early thirties constitute a core user group of major social media platforms (Pew Research Center, 2025), while also having entered professional and social contexts that may reinforce a broader sense of digital competence. In line with this, research on digital competence suggests that confidence in navigating digital environments is shaped not only by age but also by broader educational and experiential factors (Ulfert-Blank and Schmidt, 2022; van Laar et al., 2017). Taken together, sustained exposure to algorithmically curated content and experiential confidence may contribute to fluency-based overconfidence beyond what is observed among the youngest users. Although speculative, this pattern is broadly consistent with accounts of overconfidence peaking at intermediate levels of experience rather than at the earliest stages of domain engagement (Kruger and Dunning, 1999).

Another noteworthy finding concerns the differential role of algorithmic knowledge confidence in shaping distinct dimensions of information environment perceptions. While algorithmic knowledge confidence predicted both perceived information reliability and perceived information diversity in the hypothesized directions, its relative importance varied substantially. For information reliability perceptions, algorithmic knowledge confidence emerged as a robust predictor (β = .177) even after controlling for age and social media use, suggesting that trust in social media information is closely tied to metacognitive assessments of one’s algorithmic understanding. Users who overestimate their understanding of how algorithms work appear to extend this confidence into trust in algorithmic outputs. This finding aligns with broader evidence that how accurately people assess their own knowledge has real implications for how they engage with information and make decisions in complex information environments (Fischer et al., 2019, 2023; Fischer and Fleming, 2024).

In contrast, for information diversity perceptions, social media use was the dominant predictor (β = .425), with algorithmic knowledge confidence playing a weaker role (β = .150). This pattern suggests that perceived information diversity may be more strongly driven by social media exposure volume than by metacognitive assessments. Users may infer information diversity from the sheer quantity and variety of content they encounter on social media, relatively independent of their assessment of algorithmic understanding. This distinction indicates that not all information environment perceptions are equally shaped by metacognitive processes—some may be more directly tied to behavioral patterns of media consumption.

The findings carry important implications for algorithmic literacy interventions. Current approaches to algorithmic literacy often focus on increasing users’ factual knowledge about how algorithms function. While such knowledge is valuable, our results suggest that improving calibration (i.e. helping users accurately assess what they do and do not understand) may be equally or more important. In this light, for younger users exhibiting overconfidence, de-biasing interventions that provide concrete feedback about the limits of their understanding would be helpful. For example, interactive tutorials that reveal common misconceptions about algorithmic processes or self-assessment quizzes that highlight gaps between perceived and actual knowledge could help recalibrate inflated confidence. For older users exhibiting underconfidence, confidence-building interventions that acknowledge and validate their existing understanding while providing scaffolded learning opportunities would be beneficial. The goal would not be to inflate confidence artificially, but to help users recognize and trust their actual competencies.

Beyond individual-level interventions, our findings highlight the role of exposure patterns in shaping both knowledge confidence and information environment perceptions. This suggests that structural factors, such as the prevalence of algorithm-intensive platforms and patterns of age-segregated platform adoption, contribute to systematic knowledge confidence gaps across demographic groups. Addressing these patterns may require platform-level or societal-level interventions rather than solely individual education.

Limitations and future research

Several limitations warrant consideration. First, this study employed a cross-sectional design, precluding causal inferences about the directionality of relationships. While our theoretical framework posits that social media exposure leads to algorithmic knowledge overconfidence or underconfidence, which in turn shapes perceptions, the reverse or reciprocal relationships are also plausible. Longitudinal research tracking how users’ knowledge confidence and perceptions evolve over time, particularly as they increase or decrease social media use, would provide stronger evidence for causal pathways.

Second, the algorithmic awareness measure employed in this study is based on the AMCA scale, which assesses awareness of algorithmic mechanisms rather than performance-based knowledge. As such, the knowledge confidence index captures the discrepancy between subjective knowledge and awareness-based knowledge, rather than metacognitive miscalibration in the classic psychometric sense. This reflects an inherent methodological challenge in the field: the opaque and proprietary nature of social media algorithms makes it difficult to establish a definitive ground truth against which users’ knowledge can be assessed (Gagrčin et al., 2026; Oeldorf-Hirsch and Neubaum, 2023). Furthermore, as an awareness-based measure, the AMCA may itself contain inaccuracies if respondents lack insight into what they know and do not know about algorithmic systems—a concern that is particularly salient given the opaque nature of social media algorithms and the possibility that respondents’ awareness judgments reflect misconceptions as much as genuine understanding. As the field continues to develop more sophisticated approaches to examining algorithmic literacy, future research may explore methods that move beyond self-report to capture users’ algorithmic knowledge more directly and accurately.

Third, while algorithmic knowledge confidence significantly predicted information environment perceptions, it accounted for modest proportions of variance. Our mediation analyses revealed that social media use, in particular, had strong direct effects on both information reliability and diversity perceptions, independent of algorithmic knowledge confidence. This indicates that other factors—including political identity, general media literacy, trust propensities, and platform-specific experiences—likely play some roles as well. Future research should examine how algorithmic knowledge confidence interacts with these other factors in shaping information assessments.

Finally, although Prolific has demonstrated strong data quality relative to other online panels (Esch et al., 2025; Peer et al., 2022), the use of a non-probability online sample limits the generalizability of the findings to the broader US population. Also, our sample was limited to US adults. Patterns of social media use, algorithmic exposure, and algorithmic knowledge confidence may vary across cultural contexts and platform ecosystems. Replication across diverse samples and contexts would strengthen confidence in the generalizability of these findings.

Conclusion

By revealing the role of algorithmic knowledge confidence in algorithmic literacy, this work highlights the need to move beyond questions of “how much do users know” toward questions of “how accurately do users assess what they know.” In algorithmically mediated information environments characterized by opacity and limited feedback, the accuracy of self-assessment may matter as much as, if not more, the level of knowledge itself. As algorithmic curation becomes increasingly central to information access, understanding and addressing algorithmic knowledge confidence will be essential for fostering informed, critical engagement with digital media.

Footnotes

ORCID iD

Myojung Chung

Funding

The author received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Notes

Author biography

Myojung Chung (Ph.D., Syracuse University) is an Associate Professor in Journalism and Media Innovation at Northeastern University. Her research explores how the digital media environment, including algorithms and automated systems, shapes the processing and spread of (mis)information.

References

Alba

Hutchinson

(2000) Knowledge calibration: What consumers know and what they think they know. Journal of Consumer Research 27(1): 123–156.

Alter

Oppenheimer

(2009) Uniting the tribes of fluency to form a metacognitive nation. Personality and Social Psychology Review 13(3): 219–235.

Baines

Hargittai

Palfrey

(2025) Older adults’ response strategies to misinformation on social media. Social Media + Society 11(4): 20563051251386360.

Bandura

(1997) Self-efficacy: The Exercise of Control. Freeman.

Beer

(2017) The social power of algorithms. Information, Communication & Society 20(1): 1–13.

Breakstone

Smith

Rapaport

, et al. (2021) Students’ civic online reasoning: A national portrait. Educational Researcher 50(8): 505–515.

Brucks

(1985) The effects of product class knowledge on information search behavior. Journal of Consumer Research 12(1): 1–16.

Bucher

(2017) The algorithmic imaginary: Exploring the ordinary affects of Facebook algorithms. Information, Communication & Society 20(1): 30–44.

Carlson

Vincent

Hardesty

, et al. (2009) Objective and subjective knowledge relationships: A quantitative analysis of consumer research findings. Journal of Consumer Research 35(5): 864–876.

10.

Carpenter

Geller

(2020) Is a picture really worth a thousand words? Evaluating contributions of fluency and analytic processing in metacognitive judgements for pictures in foreign language vocabulary learning. Quarterly Journal of Experimental Psychology 73(2): 211–224.

11.

Chung

(2025) When knowing more means doing less: Algorithmic knowledge and digital (dis)engagement among young adults. Harvard Kennedy School (HKS), Misinformation Review, 13 October. DOI: 10.37016/mr-2020-186

12.

Chung

Wihbey

(2024) The algorithmic knowledge gap within and between countries: Implications for combatting misinformation. Harvard Kennedy School (HKS) Misinformation Review, 28 August. DOI: 10.37016/mr-2020-155

13.

Cotter

Reisdorf

(2020) Algorithmic knowledge gaps: A new dimension of (digital) inequality. International Journal of Communication 14: 745–765.

14.

Esch

Mylonopoulos

Theoharakis

(2025) Evaluating mobile-based data collection for crowdsourcing behavioral research. Behavior Research Methods 57(4): 106.

15.

Eslami

Rickman

Vaccaro

, et al. (2015) “I always assumed that I wasn’t really that close to [her]”: Reasoning about invisible algorithms in news feeds. In: Proceedings of the 33rd Annual ACM conference on human factors in computing systems, Seoul, South Korea, 18–23 April 2015, pp. 153–162. New York: ACM.

16.

Fischer

Amelung

Said

(2019) The accuracy of German citizens’ confidence in their climate change knowledge. Nature Climate Change 9(10): 776–780.

17.

Fischer

Fleming

(2024) Why metacognition matters in politically contested domains. Trends in Cognitive Sciences 28(9): 783–785.

18.

Fischer

Huff

Anders

, et al. (2023) Metacognition, public health compliance, and vaccination willingness. Proceedings of the National Academy of Sciences 120(43): e2105425120.

19.

Gagrčin

Naab

Grub

(2026) Algorithmic media use and algorithm literacy: An integrative literature review. new media & society 28: 423–447.

20.

Glenberg

Wilkinson

Epstein

(1982) The illusion of knowing: Failure in the self- assessment of comprehension. Memory & Cognition 10(6): 597–602.

21.

Gran

A-B

Booth

Bucher

(2021) To be or not to be algorithm aware: A question of a new digital divide? Information, Communication & Society 24(12): 1779–1796.

22.

Hayes

(2022) Introduction to Mediation, Moderation, and Conditional Process Analysis: A Regression-based Approach. 3rd ed. Guilford Press.

23.

Hertzog

Dunlosky

(2011) Metacognition in later adulthood: Spared monitoring can benefit older adults’ self-regulation. Current Directions in Psychological Science 20(3): 167–173.

24.

Kirschner

De Bruyckere

(2017) The myths of the digital native and the multitasker. Teaching and Teacher Education 67: 135–142.

25.

Koriat

Bjork

(2005) Illusions of competence in monitoring one’s knowledge during study. Journal of Experimental Psychology: Learning, Memory, and Cognition 31(2): 187–194.

26.

Kruger

Dunning

(1999) Unskilled and unaware of it: How difficulties in recognizing one’s own incompetence lead to inflated self-assessments. Journal of Personality and Social Psychology 77(6): 1121–1134.

27.

Lee

Diehl

Valenzuela

(2022) Rethinking the virtuous circle hypothesis on social media: Subjective versus objective knowledge and political participation. Human Communication Research 48(1): 57–87.

28.

Huang

Wang

, et al. (2023) Are digital natives overconfident in their privacy literacy? Discrepancy between self-assessed and actual privacy literacy, and their impacts on privacy protection behavior. Frontiers in Psychology 14: 1224168.

29.

Matt

Benlian

Hess

, et al. (2014) Escaping from the filter bubble? The effects of novelty and serendipity on users’ evaluations of online recommendations. In: Proceedings of the Thirty-fifth International Conference on Information Systems (ICIS), Auckland, New Zealand, 14–17 December, paper 67. Association for Information System.

30.

Oeldorf-Hirsch

Neubaum

(2023) Attitudinal and behavioral correlates of algorithmic awareness among German and US social media users. Journal of Computer-mediated Communication 28(5): zmad035.

31.

Peer

Rothschild

Gordon

, et al. (2022) Data quality of platforms and panels for online behavioral research. Behavior Research Methods 54(4): 1643–1662.

32.

Pew Research Center (2025) Americans’ social media use 2025. 20 November. Available at: https://www.pewresearch.org/internet/2025/11/20/americans-social-media-use-2025/

33.

Rader

Cotter

Cho

(2018) Explanations as mechanisms for supporting algorithmic transparency. In: Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, Montreal, QC, Canada, 21–26 April, paper 103. ACM.

34.

Reisdorf

Blank

(2021) Algorithmic literacy and platform trust. In: Hargittai

(ed.) Handbook of Digital Inequality. Edward Elgar Publishing, pp.341–357.

35.

Schneider

Schwarz

(2017) How perceived processing fluency influences the illusion of knowing in learning from TV reports. Journal of Media Psychology 29(4): 178–188.

36.

Seo

Blomberg

Altschwager

, et al. (2021) Vulnerable populations and misinformation: A mixed-methods approach to underserved older adults’ online information assessment. new media & society 23(7): 2012–2033.

37.

Tiedens

Linton

(2001) Judgment under emotional certainty and uncertainty: The effects of specific emotions on information processing. Journal of Personality and Social Psychology 81(6): 973–988.

38.

Ulfert-Blank

Schmidt

(2022) Assessing digital self-efficacy: Review and scale development. Computers & Education 191: 104626.

39.

Van Laar

Van Deursen

Van Dijk

, et al. (2017) The relation between 21st-century skills and digital skills: A systematic literature review. Computers in Human Behavior 72: 577–588.

40.

Voorveld

Meppelink

Boerman

(2024) Consumers’ persuasion knowledge of algorithms in social media advertising: Identifying consumer groups based on awareness, appropriateness, and coping ability. International Journal of Advertising 43(6): 960–986.

41.

Wallace

(2018) Modelling contemporary gatekeeping: The rise of individuals, algorithms and platforms in digital news dissemination. Digital Journalism 6(3): 274–293.

42.

Wineburg

McGrew

(2019) Lateral reading and the nature of expertise: Reading less and learning more when evaluating digital information. Teachers College Record 121(11): 1–40.

43.

Zarouali

Boerman

de Vreese

(2021) Is this recommended by an algorithm? The development and validation of the algorithmic media content awareness scale (AMCA-scale). Telematics and Informatics 62: 101607.

44.

Zimmerman

(2000) Self-efficacy: An essential motive to learn. Contemporary Educational Psychology 25(1): 82–91.