Algorithmic personalization: a study of knowledge gaps and digital media literacy

The study addresses the surge of algorithmically personalized online content and its implications for privacy, manipulation, and societal information security. Personalization leverages user data (search history, behavior, demographics) to tailor content, raising concerns about surveillance and misuse. The research aims to assess public knowledge about personalization principles, technologies, risks, benefits, and data management. The primary objective is to develop an information-analytical framework to evaluate knowledge levels across social classes in the digital media ecosystem. The core hypothesis posits that within a social class, high awareness of personalization processes, technical methods, and control over online content correlates with substantial knowledge about personalized content. The work seeks to inform targeted media literacy programs by exploring socio-demographic, socio-political, ideological, and psychological factors.

Prior research shows varied public understanding and acceptance of personalization techniques. Users often lack awareness of how personalization operates despite knowing about cookies (Ur et al., Segijn). Techniques such as psychographic segmentation, hashtag tracking, geofencing are perceived as intrusive (Segijn & Van Ooijen, 2020), with younger individuals generally more tolerant than older adults. Benefits of personalization include reduced information overload and enhanced relevance, improving satisfaction and decision efficiency (Liang et al.; Chandra et al.). The privacy paradox and personalization trade-offs highlight tensions between perceived surveillance and benefits like discounts or relevance (Awad & Krishnan; Karwatzki et al.; Siraj). Online behavioral advertising (OBA) is effective yet raises ethical concerns due to covert tracking (Boerman; Chen & Stallaert; Stanton). Socio-demographic factors (age, gender) influence attitudes, with mixed findings on gender differences (Smit et al.; Milne et al.; Boerman). Motivations, transparency, and trust shape acceptance (Aguirre et al.). Algorithmic literacy varies by demographics, with misconceptions prevalent among elderly and less educated groups (Zarouali et al.). Personalization in political contexts is less accepted than commercial/entertainment recommendations (Sehl & Eder). Personalization can exacerbate polarization and disinformation (Perra & Rocha; Brkan; Buiten).

Design: A three-stage information-analytical system employing fuzzy logic and multidimensional membership functions was developed and verified using Czech population data. Sample and data collection: 1975 invited; 1346 completed an online CAWI survey (20–27 Feb 2023; 68.2% return). After quality control (16 excluded; 117 incomplete), final N=1213 respondents. Quota sampling (sex, age, education, region) from adMeter panel; age ≥15. Average completion time 17:51. Demographics: 604 men, 609 women; 81.8% working age; 65% with secondary/professional education; 25.7% with higher education. Instrument: 19 questions assessing: (a) knowledge/awareness of personalization, (b) preferences and trust, (c) perceived rights impacts, (d) desired control mechanisms, (e) views on monitoring of online activities. System roles: respondents; system analyst (configures); decision-maker (DM) (uses outputs). Stage 1 (Information model Mkpc): Three criterion groups.

• G1: Awareness of content personalization differences across content types, 12 statements (K11–K112), each rated 1–10 (1=strongly disagree; 10=strongly agree). Aggregate score λ1(ci)=Σp O1p(ci). Fuzzification μ1(ci) via harmonic Z-spline: μ1=1 for λ1<12; μ1=(1/2)+(1/2)cos((λ1−12)/108·π) for 12≤λ1<120; μ1=0 for λ1≥120.
• G2: Awareness of technical mechanisms (K21–K25): multiple-choice; μ2(ci) mapped by number of selected answers: 0.1 (none), 0.2 (1), 0.4 (2), 0.6 (3), 0.8 (4), 1 (5).
• G3: Perceived control over content (K3) with heuristic scale: 0, 10, 15, 20, 25, 30 points; fuzzification μ3(O31(ci)) via S-shaped function: μ3=0 if O31≤0; (O31)^2/450 for 0<O31≤15; 1−(30−O31)^2/150 for 15<O31<30; μ3=1 if O31≥30. Stage 2 (Fuzzy aggregation FMkpc):
• Step 1: DM sets weights α1,α2,α3∈[1,10]; normalize βh=αh/Σαh ∈[0,1]. Example: α1=9, α2=10, α3=8 → β1=0.33, β2=0.37, β3=0.30.
• Step 2: Aggregate knowledge per citizen using one of four convolutions: pessimistic YST1(ci)=1/Σh(1/μh(ci)^{βh}); careful YST2(ci)=∏h(μh(ci)^{βh}); average yST3(ci)=Σh βh·μh(ci); optimistic yST4(ci)=√(Σh βh·(μh(ci))^2). yST(ci)∈[0,1]. Stage 3 (Social class derivation SMkpc):
• Step 3: Demographics S: Gender σ1 (man, woman); Age σ2 (15–24; 25–34; 35–44; 45–54; 55–64; 64–74; ≥75); Education σ3 (unfinished basic; basic; secondary general (without diploma); full secondary (with diploma); professional; higher (bachelor/master)).
• Step 4: Aggregate by demographic characteristic j: ΨST(sj)=(1/rj)Σq yST(cq) for all citizens in sj.
• Step 5: Social classes ST formed by combinations (e.g., gender×age×education). Use multidimensional (3D) conical or pyramidal membership function centered at (1,1,1) scaled to (3,3,3). Conical function: CKPC=1−Δ if Δ<1 else 0, where Δ=(1/3)√[(yST(σ1*)−1)^2+(yST(σ2*)−1)^2+(yST(σ3*)−1)^2].
• Step 6: Linguistic mapping of CKPC to KPC term set: (0.89,1] high; (0.77,0.89] above average; (0.65,0.77] average; (0.54,0.65] low; [0,0.54] very low. Verification and example: System verified with full dataset of 1213 respondents; approbation shown for subset of 124 respondents from Jihomoravský region.

Overall, population knowledge of personalized content in the Czech sample is at an average to above-average level across key dimensions (process awareness, technical understanding, and perceived control). Sample statistics: N=1213 (604 men; 609 women); 81.8% working age; 25.7% higher education. Weights example: α1=9, α2=10, α3=8 → β1=0.33, β2=0.37, β3=0.30. Illustrative citizen-level aggregates (average convolution): yST3(c2)=0.723; yST3(c3)=0.51; yST3(c11)=0.858; yST3(c1209)=0.654. Demographic aggregates (Jihomoravský region example): gender: ΨST(s1=men)=0.649; ΨST(s2=women)=0.646. Age: ΨST(15–24)=0.735; 25–34=0.639; 35–44=0.629; 45–54=0.634; 55–64=0.681; 64–74=0.638; ≥75=0.68. Education: ΨST(basic)=0.674; secondary general=0.62; full secondary=0.62; professional=0.121; higher=0.697. The professional education group shows notably lower aggregated scores in this fragment. Social-class example (men, 35–44, higher education): CKPC=0.812 → linguistic level: above average. Cross-group insight: Young individuals (15–24) with higher education exhibit the highest knowledge levels; men and women aged 35–44 with professional education tend to score lowest. Findings support targeted educational programs focusing on vulnerable groups.

Findings substantiate the hypothesis that greater awareness of personalization processes, technical mechanisms, and content control correlates with higher knowledge levels. Demographic disparities—particularly lower scores among mid-age groups with professional education and higher scores among younger, higher-educated individuals—signal uneven algorithmic literacy. This aligns with prior literature noting age and education-related misconceptions and differing acceptance of personalization. The information-analytical system provides actionable granularity for decision-makers to design targeted media literacy interventions, address the privacy–personalization paradox, and mitigate risks of manipulation, polarization, and disinformation. By integrating fuzzy aggregation and multidimensional membership functions, the system translates individual assessments into interpretable linguistic levels for social classes, aiding NGOs, public agencies, and regulators in prioritizing resources and tailoring curricula to bolster resilience and informational self-determination.

The study introduces a comprehensive information-analytical system to assess public knowledge of personalized content, comprising: (1) an information model with criteria on process awareness, technical understanding, and perceived control; (2) a fuzzy aggregation method for collective knowledge assessment; and (3) a fuzzy derivation method for social-class-specific knowledge levels using multidimensional membership functions. Verified on data from 1213 Czech respondents, with an approbation in the Jihomoravský region, the system evidences average to above-average knowledge overall and highlights vulnerable groups. Future work will examine determinants of acceptance of personalized services, ethical dimensions of data collection, and disinformation risk perceptions across demographics and social classes. The team plans to implement the system’s algorithms into software to enable broader practical use and facilitate comparative assessments of digital media literacy across regions.

Verification relies on a specific respondent sample and a system analyst’s interpretation of real data to delineate initial knowledge levels. Results may exhibit slight ambiguities due to the choice of multidimensional membership function types and characteristic functions. Despite these constraints, the methodology is repeatable, mathematically grounded, and designed to be independent of the number of criteria and demographic characteristics, supporting adaptability across contexts.