Abstract
The use of AI-driven recommendations is becoming more prevalent in data-driven digital markets, particularly in areas with personal data protection laws governing the gathering, use, and sharing of consumer data. However, it is not well understood when such recommendations encourage consumer value co-creation and when they do so in a manner that feels intrusive, resisted or withdrawn. This study investigates the combined influence of AI recommendations, perceived intrusiveness, information disclosure, and consumer engagement on consumer value co-creation in the context of Personal Data Protection Law (PDPL) in Saudi Arabia’s e-commerce industry. Based on the Technology Acceptance Model (TAM), Uses and Gratifications Theory (UGT), the Privacy Paradox, and recent studies of human-AI interaction, a moderated-mediation model is proposed and tested empirically via a cross-sectional online survey. The researchers employed a non-probability stratified quota sampling technique with 486 Saudi online consumers, who were recruited using various social media platforms, to gather responses. The model was estimated using PLS-SEM with bootstrapping for the direct, indirect and moderation effects, using SmartPLS 4. Results show that AI-powered recommendations are a positive predictor of consumer value co-creation (β = .483, p < .001), and disclosure is a strong predictor of reduced perceived intrusiveness (β = .684, p < .001). Intrusiveness is found to be a factor that reduces the relationships of value co-created between the AI and the consumer as predicted by the model. The model has an excellent predictive ability for all the outcomes: R2CVC = 0.574; R2IN = 0.683; and R2EN = 0.211. The study explicitly connects TAM and UGT to the Privacy Paradox, as the perceived intrusiveness is not a sufficient reason to prevent consumers from co-creating value when the perceived relevance, gratifications and benefits for control outweigh the perceived discomfort with privacy. In practice, these results can be used to inform the design of platform managers and policymakers on the design of transparent and privacy-aware recommendation strategies that can continue to support consumer choice while promoting sustainable value co-creation.
Keywords
Introduction
Yes, the Artificial Intelligence has revolutionized the search, evaluation, and selection process of products and services in a digital environment. AI recommendation systems can provide consumers with personalized content, products, and services that can significantly improve their experience and provide organizations with better revenue and loyalty results. For new digital economies such as Saudi Arabia, those systems are a key component in the spectacular growth of new ecommerce sites and super-apps (Alahmari & Awad 2025).
There are also several concerns and challenges with personalization in AI-powered recommendation engines, such as privacy risks and concerns about manipulation by the organizations. As consumers increasingly feel the pressure of organizations leveraging data and AI for personalization and recommendation systems, they increasingly feel AI-based recommendation systems to be intrusive and over-intrusive for consumers. In this context, this study explores a unique boundary condition of AI-based recommendation systems by studying a comprehensive regulatory landscape of data protection and personalization as reflected in Saudi Arabia’s Saudi Personal Data Protection Law (PDPL). Moreover, this study looks at a specific regulatory environment for AI-based recommendation systems whose explicit consent and disclosure requirements and its direct marketing provisions for personalization boundaries are unique.
While the current literature has provided compelling insight into personalization, acceptance of technology and customer experience in the context of AI, there are some gaps in knowledge. First, while a lot of literature has been dedicated to the adoption outcomes (perceived usefulness, satisfaction, or intention to use), consumer value co-creation represents a higher order of consumer participation in the creation of value, a participation that involves consumers in sharing their knowledge, providing feedback, and investing effort. Second, the role of intrusiveness and disclosure have been studied separately in previous privacy and persuasion studies but has not yet been fully specified in the context of AI-powered recommendations, especially in legal settings. Third, the Privacy Paradox has not been properly incorporated in TAM/UGT models of AI recommendation. While consumers can feel uneasy about privacy, they can also feel it is intrusive, and feel they can’t avoid giving up information, engaging, or co-creating value when they feel it will be a benefit for relevance, convenience, enjoyment, and control is seen as greater than any perceived privacy risks.
Third, most of the current studies have focused on the context of a more established Western economies, while emerging economies have received little attention, and this is not just due to the difference in the regulatory context, but there is a big difference in the cultural context as well. We therefore have very little knowledge of the impacts of recommendations based on AI on value co-creation rather than de-creation, as the consumer is aware of their data rights, and an organization has strict data protection regulations that it must follow.
To address these gaps, this research explores the role of AI-based recommendations in consumer value co-creation in a PDPL-regulated market, as well as the mechanisms and conditions. Specifically, we answer question(s) that address the following research questions:
The goal of the study is to develop and empirically validate a model that integrates AI-powered recommendations, perceived intrusiveness, information disclosure, and consumer engagement to explain consumer value co-creation under PDPL.
This research has a number of contributions. First, it promotes the Service-Dominant Logic (SDL), in which consumers are not passive recipients of algorithmic recommendations, but active partners in value creation (Vargo & Lusch, 2004). Second, it merges the TAM and UGT into a single process framework along with the Privacy Paradox. The functional acceptance of useful and relevant recommendations is explained by TAM, the motivational gratifications that drive consumers to participate is explained by UGT, and the presence of co-creation despite a sense of intrusiveness is explained by the Privacy Paradox, which states that transparency and expected benefits are important enough to drive participation even in the face of intrusiveness. Third, the model introduces engagement as a mediator between disclosure and intrusiveness as boundary conditions affecting whether AI recommendations result in value co-creation. The study applies empirical direction to study personalization and co-creation, in the context of a regulated Saudi market of PDPL.
From a practical perspective, this research can contribute to marketers, platforms, and policymakers in terms of how to develop AI-powered recommendations that are privacy-friendly, increase transparency, and enable participation rather than withdrawal or passive consumption. In doing so, it is aligned with the legal and ethical considerations and challenges facing a country in terms of its digital transformation agenda in relation to personal data (Saudi Vision 2030, 2016).
Theoretical Background
The rapid proliferation of APR systems in e-commerce has created a significant increase in the understanding of psychological, behavioral, and relational implications of these systems. A dual-theory approach—TAM (Davis, 1989) and UGT (Ruggiero, 2000) is used as an interpretative lens to understanding the evaluation and response of customers to an AI-mediated interaction (Buhalis & Cheng, 2020). The openness to interaction, however, in the context of AI, relies on the usual TAM beliefs of perceived usefulness and perceived ease of use, but also on trust, control and ethical explainability. If the consumers are given suggestions about the action with AI algorithms, then they have to cross the threshold of trust, and as the intrusiveness increases the value of the system and satisfaction will diminish. On the other hand, when the AI is Responsive, Transparent, and Explainable, the consumers demonstrate greater trust, acceptance, and sustainable use of the platform, leading to opportunities for ongoing value co-creation (Choudhary et al., 2025).
UGT complements this view by providing an understanding of why customers would want to connect. AI recommendations can meet utilitarian goals like efficiency, convenience, and decision making and hedonic and social goals, such as entertainment, relevance and social proof. According to Lou et al. (2022), Rehman (2023), and Rehman et al. (2024), in these more interactive and personalized contexts, value co-creation is realized when these motivations are fulfilled and when consumers respond in turn with feedback, co-design input, ratings, and community participation. The feeling of intrusion is related to loss of autonomy and comfort and is generally linked to psychological reactance (Mo et al., 2024). Conversely, when individuals can access and understand how data is used and how it is personalized—and that information is presented in a clear and contextually meaningful way—they are likely to feel more in control, feel fair, and can participate and co-create (Ahmed & Aziz, 2025; A. Awad & Mahmoud, 2024; Chandra & Rahman, 2024; Ghonim, & Awad, 2024).
The Privacy Paradox fills in the missing link between these two theories. It refers to the phenomenon of privacy concerns correlating with online actions in a different way from what consumers say they want; actions such as disclosing personal information, using platforms, and engaging in data-heavy services (Barth & de Jong, 2017; Kokolakis, 2017; Norberg et al., 2007). The paradox is particularly pertinent in personalized digital services as users may simultaneously appreciate the relevance of the recommendations they receive, and feel uncomfortable about the data practices that enable it (Acquisti et al., 2015; N. F. Awad & Krishnan, 2006). Perceived usefulness, ease, and acceptance are the benefit-side factors that explain why consumers may choose to continue to use and co-create despite their discomfort with privacy concerns (Jin & Zhang, 2025; Ostrom et al., 2019), as is captured by TAM; convenience, entertainment, relevance, and social interaction are the gratification-side factors that explain why consumers may choose to continue to use and co-create despite their comfort with privacy concerns, as captured by UGT; and the Privacy Paradox captures the risk-benefit inconsistency, which explains why consumers may choose to continue to use and co-create even when they are uncomfortable with privacy concerns. Finally, in the current model, disclosure lessens the perceived risk side by reestablishing control and fairness, while engagement turns the perceived benefits and gratifications into co-creative behavior. The paradox, therefore, is not regarded as a contradiction, but rather as a condition: value co-creation can be continued under intrusiveness as long as values of transparency, perceived utility and gratifications are high enough to be able to outweigh the friction that comes with privacy (Rehman, 2025b).
With the use of TAM and UGT, AI platforms can be reconceptualized as socio-technical ecosystems where system qualities, user motives and privacy-risk evaluations all interact to influence behavior (Hermann et al., 2024). Recent work focusing on consumer-technology interaction and emerging-market branding highlights the importance of respecting users’ expectations, norms, and privacy boundaries in terms of technological affordances when creating value through interaction (Agarwal et al., 2020; Rehman, 2025a; Rehman et al., 2023; Rehman et al., 2024). For this ecosystem, co-creating behaviors relies on trust, psychological acceptance, motivational engagement, and privacy-risk appraisal as key mechanisms that either facilitates or hinders the co-creating behaviors.
In practice, AI recommendations systems, which leverage machine learning and predictive models, are gaining significance regarding the content that consumers consume, the manner in which they consume it, and the activities they engage in, such as content creation, rating, sharing, etc., and how they “give back.” While personalization is sometimes used to increase convenience and efficiency, sometimes it is seen as “too personal,” and can raise concerns around manipulation and privacy issues. Personalization, if not adequately disclosed or contained, can result in a sense of surveillance, loss of autonomy and lowered engagement, ultimately lowering the amount of co-production and co-creation. Providing disclosures about personalization (how it is being achieved, types of data being used, or control options) increases consumers’ sense of control, perception of fairness, or cognitive ease and can bolster their relational anchors, which make it easier to create value, eliminating personalization fatigue.
From this perspective, consumers are more likely to turn into brand advocates and share resources to enable co-creation via experiences, product ideas, and involvement in a brand community (Lee et al., 2022). Engagement is instead presented as a crucial process rather than a mere response—a complex phenomenon that is a combination of cognitive attention, affective, and investment aspects (Maga & Bodlaj, 2025). When consumers see the use of AI as useful, respecting their privacy, and ethically sound, this interaction is magnified to drive value creation at a higher level, including enhanced user-generated content, feedback loops, and loyalty. The feeling of being “seen and respected” by algorithmic systems contributes to dialogical interaction and to social and reputational capital.
However, when behavior goes against the norm or is unclear, there is a risk of affective disengagement and cognitive resistance which can hinder co-creation opportunities (Bălan, 2023). Adding perceived fairness, algorithmic justice and emotional comfort to the goals oriented gratifications of UGT can also account for consumer acceptance and repeated behavior, as it is also a part of the perceived disclosure and intrusiveness. By ensuring that the recommendations made by AI are aligned with the gratifications that consumers are seeking, in an ethical and transparent way, consumers are more likely to move toward co-creation behavior.
These mechanisms take on a specific salience given the Saudi context of a digital transformation-oriented agenda for Vision 2030 and a Personal Data Protection Law. Beyond the legal obligation, disclosure is also an indicator of procedural justice and cultural sensitivity which makes it easier for people to engage, adds value to the personal, and minimizes intrusiveness. In this respect, the salience of disclosure is further strengthened by the consent requirements of PDPL (Art. 11) and by the direct marketing restrictions of PDPL (Art. 29 as implemented) which highlight the intrusiveness of direct marketing.
AI-Driven Recommendations and Consumer Value Co-Creation
AI-driven recommendation systems have become indispensable tools in the e-commerce landscape, as they create personalized portfolios, streamline search functionality, and help consumers navigate through choice overload. APR can act as a catalyst for engagement and co-production activities like reviews, Q&A, social sharing, and preference feedback (Abbad et al., 2021), all of which serve as forms of value co-creation in the context of Service-Dominant Logic (Solakis et al., 2024). From a privacy-paradox perspective, the consumer might be aware of the cost of privacy and choose to give it even if the functional value, convenience, enjoyment level, or social utility of the recommendation is high enough.
But if customers feel that there is an invasion of privacy or that the recommendations are “creepy,” they may choose to opt-out of the platform and disengage from co-creation activities altogether (Belanche et al., 2024; Wen et al., 2022). According to Hardcastle et al. (2025); Wang et al. (2025) and recent studies in the field of human–AI interaction (Rehman, 2023: Rehman et al., 2024), the overall idea is that an appropriate design of an APR should lead to a net positive effect on the value co-creation of consumers in a PDPL regulated market.
The intrusiveness of a boundary condition.
APR → CVC (AI-driven recommendations → consumer value co-creation) relationship depends on intrusiveness (IN) perception. Intrusiveness is defined as the level of personalization is not suitable, personalization is too persistent, or personalization breaches privacy boundaries (Nguyen et al., 2025; Omar Zaki et al., 2024; Qadeer et al., 2025; Wahid & Awad, 2025). Even when AI’s recommendations are relevant, high levels of IN can hinder interaction with the system and thus hinder value co-creation (Sun et al., 2025). For instance, users who are annoyed due to the lack of full disclosure or consent over the use of much of the data tend to reduce their interaction with the platform and its co-creative capacity (Bălan, 2023; Wen et al., 2022).
Low interaction reveals the affective and cognitive blockages to personalization-to-co-creation flow in a high-IN condition (Zhu et al., 2022). But, as the Privacy Paradox points out, intrusiveness does not necessarily kill co-creation, it just dampens the personalization reward. When the consumer feels the recommendation serves a purpose, or the platform affords meaningful disclosure and control, or participation offers immediate gratifications, the consumer can continue to provide reviews, ratings, and feedback. In contrast, interaction is enhanced when interfaces and messages honor autonomy, do not make inferences which violate norms, and provide some control. To summarize: IN does reduce the marginal returns of personalization, as the user will not combine their resources with the platform, particularly in the context of sensitive data, like markets regulated by PDPL.
Disclosure as a Transparency Lever (Moderator)
According to Sajan and Giri (2025), DS in the APR environment is defined as an understanding/explanation of a platform in terms of methods of data collection or information/data, personalization policies, and types of user information used in recommendation systems. Low perceived intrusiveness and good relationship management have been identified as key attributes of DS. To minimize perceived intrusiveness of the information frontier and to promote positive interaction of the data subject, it is necessary to present an accurate and clear disclosure, such as “why this recommendation?,” or “types of data used” or a clear option for consent (Ahmed & Aziz, 2025; Bălan, 2023). When such disclosures or understanding of such information are not provided, there may be increased occurrences of perceived privacy concerns or violations (Al-Ramahi et al., 2024; Peltier et al., 2024).
Being transparent in enhancing the perceived legitimacy and usefulness of personalization can increase users’ appreciation and encourage them to engage in behaviors that facilitate the co-creation of personalization, such as providing more feedback and joining platform communities (Lou et al., 2022; Qadeer & Awad, 2025; Sun et al., 2025; Wang et al., 2025). The difference between secrecy and transparency is that transparency builds trust and the feeling of autonomy, perceived fairness, and cognitive dissonance is created with secrecy (Al-Ramahi et al., 2024; Alghizzawi et al., 2025; Ylilehto et al., 2021). In the Saudi PDPL context, DS also signifies regulatory compliance and institutional norms that should enhance the APR-engagement pathway and thus contribute to the resolution of the Privacy Paradox by giving individuals more control over the use of their personal data for personalization.
Engagement as the Mediating Mechanism
It can be defined here as being the level of mental, emotional and behavioral engagement with the platform or the brand, manifested through things such as opinion-giving, sharing experiences, and reacting to tailored recommendations and communities. EN, in turn, serves as a moderator of the processing and interpretation of the stimulation for co-creation. Hence, the higher the degree of presence of EN, the greater its potential to shape the reduction of perceived moderation of privacy concerns and the resulting positive behavior. For consumers with low EN, the platform’s personalization can be misinterpreted, causing them to feel uncomfortable, and ultimately closing the platform down.
However, users of the AI service are also value multipliers. Personalization can enhance the EN of consumers by delivering more personalized and meaningful experiences at the right time (Payne et al., 2021; Shams Eldin et al., 2025; Vo et al., 2024; Yin et al., 2025). The arguments above are rooted in the framework and underpinned by the TAM and UGT theories, and latest research on human-AI interaction and co-production in digital services (Rehman & Zeb, 2023; Rehman et al., 2024).
The conceptual model of these relationships is presented in Figure 1.
Theoretical framework.
Methodology
Sampling and Data Collection
The research targets Saudi e-commerce consumers who are often exposed to AI-based recommendation systems on top e-commerce websites in Saudi Arabia, including Noon, Amazon.sa, Jarir, Nahdi, Extra, as well as fintech platforms like stc pay, mada-linked bank apps, etc. A quantitative cross-sectional research design has been adopted for the study based on the research objectives and hypothesized relationships between APR, perceived intrusiveness (IN), disclosure (DS), engagement (EN), and consumer value co-creation (CVC). The research design has been found appropriate for testing directional relationships between research constructs (APR → EN → CVC, where IN and DS are moderators), which is usually adopted for measuring effect sizes through PLS-SEM.
Sampling and recruitment: As social media recruitment does not offer a visible population sampling frame, the study design cannot be considered a stratified random sampling design. Rather, we employed a non-probability stratified quota sampling design to represent the key strata of APR exposure (e.g., frequent vs. occasional exposure) while recruiting participants via multiple social media platforms. To address potential selection bias, we employed eligibility screening, attention/consistency checks, and recruited participants via multiple social media platforms. These approaches help to reduce, but not eliminate, the risk of self-selection bias, which is noted as a limitation.
Sample size and statistical power: We did not employ the “10-times rule,” which is outdated. Rather, we performed an a priori power analysis (using the G*Power 3.1 approach) for multiple regression analysis with up to 8 predictors, assuming a medium effect size (f2 = 0.15), α = .05, and power = .95 (Faul et al., 2009). The power analysis suggests that a minimum sample size of N = 160 (power ≈ 0.951) is required; hence, our sample size (n = 486) is substantially beyond the minimum required. As a secondary PLS-SEM-specific check, we also point to the inverse square root rule for minimum sample size requirements in PLS-SEM studies (Kock & Hadaya, 2018).
Non-response bias was checked by comparing early and late respondents on the main study constructs by using independent-samples t-tests at α = .05. No significant differences were found across the key variables, indicating that the threat of non-response bias to the validity of the findings is not of a serious nature. Table 1 reflects the demographic profile of the respondents.
Demographic Profile (n = 486).
This study used adult human subjects by conducting an anonymous online survey, and was conducted as minimal-risk research. Potential harm was minimized because adults were screened for eligibility, consented to participate voluntarily, completed the questionnaire anonymously, no directly identifying information was collected, the questionnaire was reported collectively and a discontinue item was provided to prevent the completion of the questionnaire for adults who did not wish to participate. The questionnaire did not ask for any sensitive personal information beyond the demographic categories required for the purpose of the research and the responses were analyzed only in an aggregated form. The potential benefits of the study to the society and the participants (enhanced personalization of the AI, greater consumer control, and enhanced design of privacy-conscious platforms) outweighed the small risk of participating in the survey. The informed consent was obtained electronically prior to the completion of the questionnaire: participants read an information sheet that explained the purpose of the study, the estimated time, voluntary participation, confidentiality, the handling of the data and the right to withdraw and only those who clicked the button that displayed the consent confirmation answered the questionnaire. Data handling was done in line with PDPL principles of transparency, consent, purpose limitation and data minimization. The protocol was discussed and approved by the authors’ institution’s ethics committee prior to data collection.
Evaluation of Common Method Bias (CMB)
Since all the data was gathered by only one questionnaire and self-reporting scales were used, the presence of CMB was dealt with by both procedural and statistical remedies. Ex ante, we preserved the anonymity of the respondents, introduced various anchors for the scales as needed on the constructs and mixed conceptually distinct sets of items to reduce evaluation apprehensions and consistency motifs.
We conducted two procedures to test for CMB. To this end, Harman’s single factor test was employed with unrotated principal component analysis, and five factors were extracted which accounted for 63.9% of the total variance. The first factor accounted for 28.7% of the variance and this value is less than 50%. This does not imply the existence of a general factor accounting for most of the covariances among the items. Second, we calculated the full collinearity variance inflation factor (VIF) for all the constructs (Kock, 2015; Kock & Lynn, 2012). The results showed that all the VIFs were lower than 3.3, which indicates that neither vertical nor lateral collinearity was a problem. On the whole the above indicates that the presence of CMB is not a variable affecting the relationship of the studied variables.
Measurement Scales
All constructs were measured with multi-item scales with high level of validation and on a seven-point Likert scale (1 = strongly disagree; 7 = strongly agree). A translation-equivalence procedure was used: the items were translated into Arabic first, and then back translated into English. All items were reviewed for cultural appropriateness and PDPL relevance by a bilingual expert panel. A pre-test was carried out with the Saudi respondents to check the reliability and clarity, prior to the data collection. QuillBot was used solely as a language tool that assisted in the polishing of texts to be translated and the clarity in English; no AI tool was used to create research content, data analysis, interpretation, or conclusions, which were all reviewed, edited, and approved by the authors.
- APR: Artificial Intelligence-Powered Recommendations. Perceived personalization, relevance, and facilitation of discovery were captured by five items, for example, “The system offers personalized recommendations based on my past experiences”; “It helps me discover products I would not have otherwise known.”
- Co-creation of Consumer Value. CVC was measured using six items adapted from Wen et al., 2022 which reflected co-production and value-in-use such as: “The company values my input in the creation of offerings.”
- Perceived Intrusiveness: IN. IN was measured using four items drawn from Sun et al. (2025), which capture discomfort with overly personal or invasive recommendations (e.g., “Recommendations feel overly personal, which makes me uncomfortable”).
- DS: DS was assessed with four items adapted from Xia et al. (2024). The items are related to transparency about the collection and usage of personal data for personalization, such as “The company clearly explains how my data are used to generate recommendations.”*-Engagement (EN).
The EN scale had four items which were adapted from Ylilehto et al. (2021) and measured cognitive, affective, and behavioral commitment to the APR system (e.g., “I like using the APR system when I shop online”). Pilot and main samples of the instrument had good reliability and validity. All constructs demonstrated internal consistency with Cronbach’s alphas ranging above acceptable levels and composite reliabilities for all constructs being above the acceptable level. Convergent validity was demonstrated with the average variance extracted higher than 0.50 for each construct. The discriminant validity was evaluated by the heterotrait-monotrait ratio and all HTMT values were less than 0.85. The results showed that the measurement model is valid and reliable in relation to its constructs in the context of the APR in Saudi Arabia.
Results
These hypothesized relationships were tested by using partial least squares structural equation modeling (PLS-SEM) in SmartPLS 4. The PLS-SEM is suitable in the present study as it can handle complex models with multiple latent constructs, indirect and interaction effects, and does not require multivariate normality, which are common in AI personalization and digital commerce studies.
Reliability & Validity
The internal consistency reliability, convergent validity, and discriminant validity of the reflective measurement model were first investigated. Internal consistency reliability was determined with Cronbach’s alpha (α) and composite reliability (CR) with the criterion of 0.70. Convergent validity was tested using the average variance extracted criterion (AVE ≥ 0.50). The Fornell-Larcker criterion (where the square root of AVE should be greater than the inter-construct correlations) and HTMT values < 0.85 were used to check for discriminant validity (Fornell &Larcker 1981).
Table 2 gives an overview of the measurement diagnostics. As observed, the loadings of the indicators are all high and unambiguous, most of them ≥ 0.85, the Cronbach’s alpha and CR values are all ≥ 0.70 for all the constructs which means convergent validity has been satisfied with AVE range from 0.604 to 0.854. The square root of AVE of all constructs is higher than the interconstruct correlation and the HTMT values which are not exceeding 0.845, are within the limit of 0.85. Although conceptually related, overall this is a strong support as the measurement indicators of APR, DS, IN, EN, and CVC measure the constructs with precision and consistency. On a general note, this measurement model is a good foundation for subsequent analyze.
Measurement Model Diagnostics and Structural Validity Indicators.
Note. AVE > correlations = Fornell–Larcker criterion for discriminant validity. HTMT < 0.85 = acceptable discriminant validity. f2 (Cohen, 1988): 0.02 = small; 0.15 = medium; 0.35 = large effect.
The close proximity of the loading intervals (i.e., primarily ≥ 0.85) together with the large α/CR values and AVE values between 0.604 and 0.854, all confirm that the indicators are measuring their respective targeted latent constructs with accuracy and reliability. Achievement of both Fornell-Larcker and HTMT conditions further confirms the validity of evidence of the discriminating nature of APR, DS, IN and EN as statistical constructs, which is a prerequisite for making valid inferences regarding the complex and indirect/conditional relationships between APR, and CVC, in an AI-mediated Saudi e-commerce context.
Model Characteristics
The R2 (coefficient of determination), effect size (f2), Stone-Geisser’s predictive relevance (Q2) using blindfolding method and the standardized root mean square residual (SRMR) as global model fit were computed (Hair et al., 2019; Hu & Bentler, 1999; Stone, 1974).
Variance explained (R2). From a behavioral research perspective, standard cut-off for R2 values are .75, .50 and .25 which represent strong, medium and weak models, respectively, and this model includes a significant amount of variance in all key endogenous variables. Consumer value co-creation (CVC) has an R2 of .574 showing a medium-strong relationship between APR and DS and CVC, whereas perceived intrusiveness (IN) has an R2 of .683 which means that the combined effect of APR and DS on perceived intrusiveness (IN) is significant. Furthermore, the proximal process variable of engagement (EN) yielded a very modest R2 of .211, which is certainly not trivial given an exploratory mediation-moderation design, indicating that at least some part of this variable is affected by the structural antecedents as detailed in this model.
Predictive Relevance (Q2). Blindfolding was used to determine the Q2 values. The standard cut-offs are 0.35/0.15/0.02 which corresponds to large, medium and small predictive relevance. The models show predictive significance for CVC (Q2 = 0.366) and IN (Q2 = 0.397) as shown in Table 3. It has a small to medium predictive relevance for EN (Q2 = 0.127). This suggests not only that the model has a strong ability to explain the data but also that it has strong out-of-sample predictive ability for the key outcome (CVC) and the key boundary construct (IN).
Model Characteristics.
Note. SRMR: 0.060. EN* = consumer engagement/proximal process variable; the asterisk is explanatory only and does not denote p < .05.
Effect sizes (f2). We also explored the effect size to examine the incremental contribution of each exogenous construct to the endogenous target construct. According to the values shown in Table 2, all the variables APR, DS and IN have small to moderate impact on their respective dependent variables. This is theoretically consistent in that in part DS will reduce perceived intrusiveness and increase engagement whereas IN will act as a boundary condition that will moderate the APR–co-creation relationship without overpowering it.
A rough mark of model fit, called SRMR (standard root mean square residual). While PLS-SEM does not seek to assess the exact fit of the model, the SRMR index can be used as an approximation of the goodness of the approximation of the correlation matrix suggested by the model to the matrix of the actual correlations. Presented the SRMR value is 0.060, lower than the recommended value of 0.08 (Hu & Bentler, 1999), which helps to reduce possible concerns about significant misspecifications of the structural or measurement models.
All results, R2, Q2, f2, and SRMR, give evidence to the explanatory as well as predictive sufficiency of the proposed model of the relation between APR and CVC using EN, while IN and DS are important contextual factors in a PDPL-controlled Saudi e-commerce context.
In the estimation run outlined here, the proximal process variable explained at R2 = .211 is the engagement/communication channel construct in the process layer (listed as “DS/EN” in some previous intermediate drafts). Its variance explained and Q2 are interpreted as modest yet sizable in an exploratory mediation–moderation design.
Hypotheses Testing
The structural relation was further validated with the bootstrapping of the PLS-SEM analysis with the 5,000 resampling with SmartPLS 4.0, which even included a bias-corrected p-value and 95% CI for not only direct but also indirect (mediation) and interaction (moderation) effects (Hair et al., 2019). The main path coefficients, p values, and CI are presented in Table 4, the focal moderation relationship is presented graphically in Figure 2, and the complete structural model is presented in Figure 3.
Hypothesis Testing.
Moderating effect of perceived intrusiveness on the APR–CVC relationship.
Structural model showing direct, indirect, and conditional paths.
A direct influence of APR on CVC (H1). APR and CVC are positively and statistically significantly related (p < .001, β = .483), which is in line with H1. Consumers value co-creation more as the quality, relevance, and usefulness of AI-based recommendations increase, which leads to higher levels of consumer value co-creation. Consumers in Saudi Arabia are more likely to provide feedback, share experiences, and engage in activities on the platform that have value for both parties if the APR is timely, relevant, and suitable for the consumer.
Mediated process through engagement (EN; H4). To justify the proposed mechanism, APR is positively correlated with EN and EN is positively correlated with CVC. The indirect effect APR → EN → CVC is statistically significant, which means that part of the effect of APR on CVC is indirect via EN. The above evidence lends support to H4 and highlights the crucial importance of engagement as the psycho-behavioral bridge that turns the AI-based personalization into co-creation outcomes.
Perceived intrusiveness (IN; H2). The moderation analysis reveals that the positive relationship between APR and CVC is reduced by IN. The interaction term APR × IN is significant in the predicted direction and the simple-slope plot (Figure 2) indicates that the APR–CVC relationship is greater in the lower levels of IN and less in the higher levels of IN. This is consistent with the Privacy Paradox: though consumers may not stop co-creating when they feel intrusive, having a sense of privacy decreases the marginal value of personalization for them, provided there is a high sense of relevance, gratifications, and control. The simple-slope plot also indicates that there is a possible attenuation (not reversal) within its range, reflecting a managerial understanding of intrusiveness as a friction effect rather than a tipping point into negative value co-creation.
Additional Relevant Links (Direction and Consistency of Ancillary Paths)
The APR-CVC relation is not reversed, but is rather attenuated, given that PR is positive at all levels of IN. This supports H2. Disclosure (DS) as a tool to enhance transparency (H3). The study revealed a two-fold positive effect of DS in the Saudi PDPL context: DS was negatively correlated with IN, due to the transparency demonstrated by the disclosure of the DS’s use of the data and the provision of control mechanisms to help reduce perceptions of intrusiveness. The same factors that make consumers more willing to use the platform and be more active with the recommendations also increase their willingness to use DS, as is evident from the positive relation between the two. Moreover, exploratory research indicated that the more positive this relation is the higher the DS is. This is consistent with the notion that the more transparent the platform is, the more motivational impact the recommendations offered by the platform will have.
Other structural relations. The other paths described in Table 4 provide more information about the process. Perceived disclosure is positively correlated with APR (APR → DS)—that is, users assume transparency when they experience recommendations that are well-explained and controllable. The relationships between DS and IN are negative as expected, with higher DS pointing to lower IN (disclosure scale coded toward transparency/control), while the relationship between APR and IN is also negative but weaker, as the disclosure scale is coded toward personalization and increased personalization can be perceived as intrusive without sufficient disclosure. The positive direct association with CVC in the full model does not contradict the interaction effect that CVC strengthens APR–CVC slope, as both effects are present in the full model. From a Privacy Paradox perspective, when the potential benefits of intrusive personalization are great, some consumers act as co-creators, but the value of APR diminishes as intrusiveness increases.
Figure 2 shows how perceived intrusiveness moderates the APR–CVC relationship, while Figure 3 depicts the complete structural model with standardized path coefficients and conditional paths. Together, the figures show that subtle and controllable personalization—high DS and managed IN—facilitates engagement and enables value co-creation in Saudi e-commerce platforms.
Discussion
We suggest that required disclosures under the PDPL serve as a “trust anchor,” with transparency about what happens to data signaling procedural fairness, legal compliance, and control, thus promoting engagement instead of defensive withdrawal. This is also where the Privacy Paradox materializes in this study: consumers might feel a sense of privacy risk or intrusiveness, but still behave and co-create, if disclosure generates the feeling of an understanding and manageable privacy threat and if the functional and motivational benefits of APR are strong enough.
This study aimed at exploring the role of APR on the promotion of CVC in the Saudi e-commerce context and the circumstances that can facilitate or undermine the benefits of APR. The APR affects CVC directly and indirectly through EN, and perceived IN and disclosure (DS) are critically important contextual variables in a PDPL-regulated environment, inspired by TAM, UGT, the Privacy Paradox and new insights into HAI. The results provide a number of insights. First, the positive and significant effects of APR on CVC indicate that Saudi customers are more likely to offer effort and knowledge (reviews, feedback, co-design input, or community participation) if they find the recommendations made by AI to be useful, relevant, and appropriate. Second, the psycho-behavioral mechanism of engagement is the one that creates the link between personalization and co-creation. Third, perceived intrusiveness is not just a constraint; it reduces, but does not eliminate, the value of personalization. This is aligned with the principles of the Privacy Paradox and the privacy calculus that consumers might still continue to engage if they believe that the perceived relevance, convenience, enjoyment, and control outweigh the discomfort they feel regarding privacy. Fourth, disclosure also lowers the risk side of the equation and enhances engagement by explaining why recommendations are provided, how data is used, and the options available to the user to modify or even remove themselves from the process (Ding et al., 2025; Lopes et al., 2025; Yang & Wang, 2025). This finding thus reinforces the idea that transparency and explainability are both fundamental and essential aspects of the design of an AI service, not just tasks to perform for compliance.
Practical Implications
Create designs for ‘explainable engagement’. Disclosure should not be a back-stage legal footnote, but a front-stage design element. One-tap controls (e.g., muting topics, adjusting sensitivity, or opting out by category), in-context data-use notices that are easy to notice, and “why this recommendation?” affordances can reduce perceived IN (IN) and increase engagement (EN), thereby improving the APR → CVC subpathway described here.
Implement the principles of the Personal Data Protection Law in the interactions, such as purpose limitation, data minimization and fine-grained consent, as a part of the product strategy. The topic-based consent-for example, electronics, groceries, travel-and context-based consent-for instance, Ramadan campaigns versus regular browsing-foster the protection of cultural and religious sensitivities and reduce IN in very sensitive periods. This is not a compliance requirement, but a basis for trusted product design, PDPL.
Personalization should not be done for the sake of personalization. Complaint trends, “not relevant” clicks, signals for privacy concern, and drops in dwell time after exposure to recommendations or ads make up “intrusiveness telemetry” that should be tracked by organizations. If the IN indicators fall outside of the threshold values, automatic rate limiting and category level dampening should be activated. The goal is to maintain the value of the AI-powered recommendations without going past the “creepy line” where personalization threatens to undermine EN and CVC.
Create an engagement rather than just clicks. Ranking logic should not be based exclusively on CTR (click through rate) but should include in EN-related proxies, that better capture the ideas of co-creation, such as: helpfulness votes on reviews, activity within forums/communities, idea submissions. Polls that enable co-creating the content, guided review prompts, community Q&A modules, and other features help transform the sense of relevance into continuous contribution and value co-creation.
Localization playbook for KSA.
Creative sensitivity: Ensuring imagery and copy is modest and sensitive to the religious calendar and local cultures.
Family-oriented bundles: Positioned on household utility, social cohesion, and community value to accord with family-oriented cultural values.
Arabic interface nuance: Avoid technical jargon and avoid implying judgment; use clear, judgment-free Arabic wording for disclosure, permissions, and controls.
Safe defaults: Take cautious, privacy-preserving default profiles and only incrementally reveal more advanced personalization options as trust and EN develop.
Governance and Oversight: Firms should establish a Personalization Review Board or similar governance structure composed of product, legal/PDPL, and ethics representatives. This body would review new personalization features for their potential impact on IN, oversee A/B testing of DS formats and consent flows, and maintain a “model card” style log documenting data sources, treatment of sensitive attributes, and user-control affordances. Such governance helps ensure that APR remains aligned with both PDPL and local ethical norms while supporting sustainable co-creation.
Limitations and Future Research
Design and generalizability. Tables 2 to 4 show that the Saudi sample is large and the diagnostics of the model are satisfactory. The research design, however, is cross-sectional in nature and, thus, causal inferences are theory-provided and not experimentally tested. Longitudinal or field-experimental research in the future might follow the development of EN and IN under different disclosure policies, consent mechanisms, and recommendation logics, such as before/after disclosure policy changes, or before/after the addition of privacy prompts.
User heterogeneity: We conducted this study without carrying out detailed multi-group analyses across the age, gender and experience with AI categories. Future work may examine if privacy concerned groups, younger users, and new-to-platform users have stronger DS → EN effects, or more pronounced attenuation from IN. The results of these would offer a more substrate basis for introducing more specific disclosure and personalization approaches to various user groups.
Algorithmic form. APR was not differentiated based on algorithm type, but rather was operationalized as a higher order construct. Future research could further explore the differences among collaborative filtering, content-based, and hybrid recommenders as well as generative versus heuristic explanation style. Thereby, non-uniform effects of EN and IN can arise, which can result in more complex recommendations for the design of recommender systems inregulated environments.
Perceived fairness and justice: Not separately modeled as a construct in the current study. The dimensions may help boost predictive validity for the other three dimensions (trust, EN, and ultimately, CVC) in future models and offer additional connections to the more contemporary human–computer interaction and human–AI communication frameworks of AI personalization.
Granularity of intrusiveness. In this treatment, IN has been dealt with as a single entity. Future research might be directed toward breaking down intrusiveness into data-scope, context intrusiveness, and frequency intrusiveness: What kinds of data are used, where and when personalization occurs, and how often? More granular operationalization will enhance the conceptualization of the “creepy line” and give clearer management interventions.
Behavioral telemetry and mixed methods. The data for this study is self-reported survey data. With explicit consent and anonymization, matching survey with behavioral telemetry could allow researchers to determine if increased or decreased self-reported EN correlates with increased or decreased observed co-creation practices including review quality, peer feedback, and idea submission, and whether changes in the use of DS practices result in measurable changes in behavior (Petrescu et al., 2024). Qualitative interviews or focus groups in mixed-method designs might be able to further elucidate consumer understanding of APR, IN, and DS in PDPL regulated markets.
Conclusion
The results of this study indicate that AI-driven recommendations in the Saudi E-commerce sector generates CVC either directly or via the activation of EN. The positive APR → CVC effect, however, does not vanish as IN increases, but rather becomes less positive; this pathway is perceived as intrusive. The study integrates TAM, UGT and Privacy Paradox to provide an explanation for why individuals might still create value through co-creation when the perceived benefits and gratifications associated with the intrusive personalization more than offset the discomfort of privacy loss, and disclosure-enabled control is experienced. The findings highlight disclosure as an important management tool. Disclosure and control processes that are transparent, culturally responsive and PDPL aligned can minimize IN, boost EN, and safeguard APR to CVC.
The Saudi context-R2, Q2, and SRMR, in turn, provide strong support for these inferences in terms of explanatory and predictive diagnostics of the model in the Saudi context. For practitioners, this means thinking about disclosure as an infrastructure of engagement, about personalization in a bounded way, and about optimizing for contribution and co-creation rather than clicks.
For researchers, this study aggregates the concepts of personalization, transparency and intrusiveness, in addition to the Privacy Paradox, under a single process model, the moderated–mediated one. This model can inform future longitudinal, multi-group and algorithm-specific studies of the joint value creation potential of AI recommendations in Saudi Arabia and other culturally diverse markets with regulation.
Footnotes
Acknowledgements
The authors are thankful to the Deanship of Graduate Studies and Scientific Research at University of Bisha for supporting this work through the Fast-Track Research Support Program.
Ethical Considerations
Ethical review and approval were obtained from the authors’ institutional ethics committee prior to data collection. The committee name, institution, and approval number are blinded in this peer-review version to preserve anonymity and will be provided in the non-anonymized title page/final manuscript in accordance with Sage Open requirements. The study involved an anonymous online survey of adult participants and was assessed as minimal risk. Risk of harm was limited by voluntary participation, anonymous response collection, non-collection of directly identifying information, the right to discontinue before submission, and aggregate reporting.
Consent to Participate
Electronic informed consent was obtained from all participants before the questionnaire began. Participants were informed about the study purpose, voluntary nature of participation, anonymity/confidentiality, minimal risk, expected benefits, PDPL-compliant data handling, and the option to withdraw before submission. Only participants who confirmed consent and eligibility as adults aged 18 years or above proceeded to the survey.
Consent for Publication
Not applicable; the manuscript reports only aggregate, non-identifiable survey results and includes no individual-level personal data, images, or identifiable quotations.
Author Contributions
Conceptualization: Abdelrehim Awad; Methodology: Bshair Alharthi; Formal analysis: Bshair Alharthi, Abdelrehim Awad; Investigation: Bshair Alharthi, Abdelrehim Awad; Data curation: Bshair Alharthi, Abdelrehim Awad; Writing—original draft: Bshair Alharthi, Abdelrehim Awad; Writing—review & editing: Bshair Alharthi, Abdelrehim Awad; Visualization: Bshair Alharthi, Abdelrehim Awad; Supervision: Bshair Alharthi, Abdelrehim Awad.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
The anonymized dataset supporting the findings of this study is available from the corresponding author* upon reasonable request, subject to institutional review requirements and PDPL-compliant data-sharing procedures. No directly identifying participant information is included in the dataset.
Declaration of AI-Assisted Language Editing
QuillBot was used only for translation polishing and English-language clarity. The authors reviewed and approved all wording and remain fully responsible for the scientific content, analyses, interpretations, and conclusions.