Indirect Media Effects on the Adoption of Artificial Intelligence: The Roles of Perceived and Actual Knowledge in the Influence of Presumed Media Influence Model

ABSTRACT

To examine the indirect media effects on the adoption of artificial intelligence (AI), we employed the influence of presumed media influence (IPMI) model and its extensions. The results of an online survey of adults (N = 1,360) in Hong Kong revealed that both their media attention to AI-related information and interpersonal communication about AI influenced presumed others’ media attention. In turn, that dynamic shaped perceived descriptive norms and injunctive norms of using AI, which influenced individuals’ intention to use AI for work and/or learning. Furthermore, although persumed media influence was stronger among individuals with high perceived knowlegdge than ones with low perceived knowledge about AI, no difference emerged between individuals with high versus low actual knowlege about AI.

Artificial intelligence (AI) technology encompasses a wide range of computer technology capable of performing tasks or making decisions that require human intelligence (Zhang & Dafoe, 2019). Although the emergence and prevalence of AI technology have brought substantive concerns, such as data security and copyright infringement, the technology has also become imperative in the development of society. For example, AI technology has created more jobs than anticipated across various industries such as technology, financial services, and health care (Loukides, 2022). Beyond that, AI is expected to significantly impact economic growth. For instance, AI is expected to double annual global economic growth rates by 2035 (Szczepański, 2019). Thus, the adoption of AI technology, especially for learning and working purposes, is crucial to society’s economic growth.

To facilitate the adoption of new technology, effective communication and the diffusion of information are important (Rogers, 1995). Diffusion refers to the process by which a novel technology spreads through different communication channels, including mass communication, interpersonal communication, and network communication within a social system (Rice, 2009). This diffusion process is influenced by several underlying factors, such as the extent to which actions, perceptions, communication processes, social norms, and social structures jointly to alleviate adopters’ uncertainty regarding the new technology (Rice, 2009; Venkatesh & Davis, 2000). In addition, at an early stage of diffusion, individuals may rely on multiple communication channels to gather information and make informed decisions regarding the adoption of such a technology. As of writing this paper, OpenAI had launched ChatGPT 3.5 and subsequently ChatGPT 4.0, which made AI technology unprecedently available and accessible among laypeople. Amid the increasing popularity of ChatGPT, AI technology has become a trending topic. This timing presents an ideal opportunity to study how communication influences individuals’ technology adoption as well as its underlying mechanisms.

In this study, to understand how communication affects individuals’ intention to adopt AI for learning and/or work, we have consulted the influence of the presumed media influence (IPMI) model (Gunther & Storey, 2003) and its extensions (Gunther et al., 2006; Shi et al., 2022). Those frameworks aided us in particularly investigating how mass communication and interpersonal communication shape individuals’ perceived norms via presumed media attention from others, which subsequently influences their intentions. In detail, mass communication, which refers to the dissemination of information from one source to a large audience (Chaffee & Metzger, 2001; Napoli, 2010), takes place through various channels. These channels encompass traditional media such as television and newspapers, as well as new media channels like social media platforms and news applications (Lang & Lang, 2009; Napoli, 2010), which collectively represent the diverse avenues available for mass communication in today’s media environment. In addition to mass communication, interpersonal communication provides another channel to diffuse information, both of which collaboratively shape presumed media influence (Shi et al., 2022).

Furthermore, given that media effects are contingent (Valkenburg et al., 2016), we explored the roles of two types of knowledge – actual knowledge and perceived knowledge – in the IPMI model as a means to examine how such individual differences shape the IPMI. Our findings are expected both to advance the IPMI model by identifying its boundary conditions and to inform communication strategies for promoting the adoption of AI technology.

The IPMI Model and Its Extensions

The IPMI model captures the multistep process illustrating the indirect media effects (Sun, 2013). In that process, individuals who pay attention to media content form basic initial impressions of a specific topic, and based on their own level of attention, they estimate others’ level of attention to the content as well. Subsequently, they assess how others are influenced by such presumed attention, which further impacts their own cognitive, attitudinal, or behavioral responses to the media content (Gunther & Storey, 2003).

Although the IPMI model originates from the theory of third-person effect (Davison, 1983), it is considered to be a “more general” model with “broader application” (Gunther & Storey, 2003, p. 201) because it focuses on the presumed influence of others as the antecedent of the audience’s reaction. In contrast to the theory of third-person effect, which emphasizes self – other differences in media’s influence, the IPMI model’s core maintains that an audience’s reactions depend on media content’s presumed influence on others, not on media’s actual influence on self – other gaps in perception (Gunther & Storey, 2003; Sun, 2013). In addition, the IPMI model releases the constraint of “negative influence corollary” (p. 201), suggesting that indirect media effects can be generalized to media content with a broader range of presumed social influence, regardless of its valance (Gunther & Storey, 2003).

The first proposition of the IPMI model contends that a positive association exists between individuals’ personal media exposure and presumed others’ media exposure (Gunther & Storey, 2003; Gunther et al., 2006). From an audience-centered perspective, media exposure can be conscious or unconscious as well as concentrated or unconcentrated. Based on the audience’s attention and consciousness, the state of exposure can be categorized as passive exposure or active attention (Potter, 2009). Whereas media exposure refers to exposure to media content, media attention, meaning attention paid to specific media content, involves processing information with certain cognitive efforts (Potter, 2009). In other words, in a state of passive exposure, audiences can perceive the element of the message but process it unconsciously, whereas in active attention, audiences need to devote cognitive effort to process the information, and thus media attention has been regarded as a better proxy for cognitive effects than media exposure (Ho et al., 2020; Potter, 2009). Therefore, in the current study, we opted to examine media attention instead of media exposure. This allowed us to capture, to some extent, valid cognitive efforts of the audience members in forming the basic impressions of a given media content.

Based on their own media attention, individuals develop a subjective estimation of others’ attention to the same media content, known as the “presumed reach” (Gunther et al., 2006, p. 56). Then, individuals make persuasive inferences about the influence of such media content on others’ responses, assuming that it affects others’ attitudes and behaviors (Gunther, 1998). In other words, based on their own media attention, individuals infer the media attention of other audience members, as well as how others’ attitudes or behaviors are influenced by the same media content. In studies examining how media content shapes individuals’ adoption of new technology, the positive association between media attention and presumed others’ attention has been documented. For instance, individuals who frequently paid attention to messages about the benefits of plant-based meat or nano-enabled food presumed that their peers paid frequent attention to those messages as well (Ho et al., 2020, 2022). Based on the first proposition of the IPMI model and those empirical findings, we proposed Hypothesis 1 (H1):

H1:

Individuals’ media attention to information related to AI technology is positively associated with presumed others’ attention to similar information.

The diversity of communication channels and the frequency of attention matter in forming the impression of media content. Individuals are more likely to passively or actively access information delivered through various channels when an issue generates an information-rich environment (Avery, 2010). Whether intended or not, individuals can engage in repetitive communication about one specific topic through multiple communication channels to form a basic impression about the specific media content (Jang & Park, 2018). By contrast, if an issue emerges in an information-poor environment, then individuals are less likely to access that information from any channel, which leads to a weak impression of the media content.

To ensure that a message makes an impression, the literature addressing IPMI model often presents messages as emerging in information-rich environments, including political campaigns (e.g., Cohen & Tsfati, 2009) and public health issues (e.g., Hong, 2021). Because AI technology is currently a trending topic, both mass communication and interpersonal communication matter for the diffusion of AI technology and its acceptance. In particular, as information is diffused via mass communication channels, interpersonal communication provides a supplementary channel to diffuse information from mass media to those who pay little attention to the information on such media (Katz & Lazarsfeld, 1955). In a recent study, interpersonal communication was introduced and found to serve as the antecedent of presumed others’ media attention within the context of COVID-19 pandemic (Shi et al., 2022). That finding indicates that interpersonal communication matters for the diffusion of information in information-rich environments and can consequently shape individuals’ perceptions of others’ media attention and subsequently their own behavioral intentions. Therefore, we proposed Hypothesis 2 (H2):

H2:

Individuals’ frequency of engaging in interpersonal communication about AI technology is positively associated with presumed others’ media attention to information related to AI technology.

Normative Mechanisms in the IPMI Model

The presumed media influence on others refers to the subjective estimation of how a given media content influences others (Gunther & Storey, 2003). This presumed influence operates through two underlying mechanisms: ecological influence and normative influence. In the mechanism of ecological influence, individuals’ responses are driven by the consideration of the potential influence of media content on others’ responses. It involves estimating the magnitude of media influence on others or subjectively assessing how others may behave in a certain way due to media effects (Sun, 2013). On the other hand, normative influence assumes media to be a source of normative perceptions and extends the concept of presumed influence to encompass normative perceptions, including perceived descriptive norms and injunctive norms, which further shape individuals’ responses (Gunther et al., 2006; Sun, 2013).

In detail, perceived descriptive norms refer to the estimated prevalence of a particular behavior, whereas perceived injunctive norms refer to the estimated social approval of performing that behavior (Rimal & Lapinski, 2015). Particularly, in an early stage of new technology diffusion, individuals may lack direct experience with or be uncertain about the technology. People may rely on the behaviors or opinions of others in their social environments to make decisions regarding technology adoption (Rice, 2009; Venkatesh & Davis, 2000). We thus focused on examining the normative mechanism of presumed media influence in the current study. The positive association between presumed others’ media attention and normative perceptions is thoroughly documented in the literature addressing the IPMI across media content. For instance, in the context of HIV prevention, individuals’ perceptions of the prevalence of the use of pre-exposure prophylaxis prompted their intention to seek information about using pre-exposure prophylaxis (Hong, 2021). Moreover, presumed attention to pro-smoking and pro-drinking media content is positively associated with perceived norms of the risky behavior, which in turn prompts behavioral intentions (Gunther et al., 2006; Ho et al., 2014). Thus, we proposed Hypotheses 3 and 4 (H3 and H4) as follows:

H3:

Presumed others’ media attention to AI technology is positively associated with perceived descriptive norms of using AI technology for work and/or learning.

H4:

Presumed others’ attention paid to media content related to AI technology is positively associated with perceived injunctive norms for work and/or learning.

Moreover, normative perceptions have been theorized as important antecedents of behavioral intention in the IPMI (Gunther et al., 2006; Sun, 2013). Additionally, the theory of technology acceptance holds that individuals tend to rely on their normative perceptions in deciding whether to use new technology at the early stage (Venkatesh & Davis, 2000). Therefore, we also hypothesized:

H5:

Perceived descriptive norms of using AI technology for work and/or learning is positively associated with the intention to use AI technology for work and/or learning.

H6:

Perceived injunctive norms of using AI technology for work and/or learning is positively associated with the intention to use AI technology for work and/or learning.

Knowledge As a Boundary Condition in the IPMI Model

To investigate the boundary conditions for the indirect media effects illustrated by the IPMI model, we followed the recommendation in Valkenburg et al. (2016) review that such indirect media effects can be enhanced or reduced by individual differences. In particular, individuals differ in their ability to process relevant media content (Krcmar, 2009), which consequently prompts different patterns of media effects. In fact, the ability to process media content can manifest as knowledge about the focal topic (Asghar et al., 2022; Driscoll & Salwen, 1997).

Within the context of IPMI, when media attention has occurred, individuals engage in the assessment and interpretation of media content (Krcmar, 2009). This interpretation and assessment process of media information is influenced by both subjective knowledge and actual knowledge. Indeed, knowledge contributes to the variations in understanding and construal of the media content under consideration (An, 2007; Krcmar, 2009). Based on the metacognitive framework (Nelson & Narens, 1990), knowledge can be understood at two underlying levels: the object level, which refers to actual knowledge, and the meta level, which pertains to perceived knowledge. On the one hand, actual knowledge is specific to a particular topic and serves as an objective measure of competence regarding the topic (Park et al., 1988). By contrast, perceived knowledge refers to individuals’ self-assessment of their knowledge (Radecki & Jaccard, 1995), which reflects the combination of actual knowledge and one’s self confidence (Raju et al., 1995). In other words, individuals may exhibit discrepancies between their actual and perceived knowledge.

In the literature on the media effects, perceived knowledge is often regarded to be undesirable because it leads to the illusion of knowledge and could amplify the presumed negative effects of media content on others (e.g., Schäfer, 2020; Yang & Tian, 2021) or enlarge the self – other difference in perception (e.g., Chen & Atkin, 2020; Driscoll & Salwen, 1997). However, in McLeod et al. (1997) study on rap music, perceived knowledge about rap music was not found to be associated with the self – other difference in perception. The findings regarding actual knowledge have also been mixed. For instance, Driscoll and Salwen (1997) found that actual knowledge was not correlated to the third-person effect, whereas Huh and Langteau (2007) discovered that individuals with a high level of actual knowledge tended to perceive the influence of direct-to-consumer prescription drug advertising on others to be greater than individuals with a low level of actual knowledge.

Although little research to date has investigated the role of perceived and/or actual knowledge in the IPMI model, the mentioned finding on the third-person effect indicates that the two types of knowledge could play a role in shaping the perceived magnitude of media influence on others, which might further shape the outcomes of the IPMI. In this study, drawing upon the IPMI model and taking into account previous, mixed empirical findings, we asked the following research questions (RQs):

RQ1:

Does actual knowledge (high vs. low) about AI moderate the pattern of the IPMI? If so, then how does it moderate that pattern?

RQ2:

Does perceived knowledge (high vs. low) about AI moderate the pattern of the IPMI? If so, then how does it moderate that pattern?

Method

Participants and Procedures

Using quota sampling, we administrated an online survey via Qualtrics with a total of 1,360 citizens (677 women, 49.8%) in Hong Kong at least 18 years old (See Table B1 in the Supplementary Materials). The survey took place from May 9 to May 19, 2023 (i.e., approximately 5 months after the launch of ChatGPT 3.5 and 5 days prior to the launch of ChatGPT 4), and the survey’s quota were set with reference to the census data of adults in Hong Kong regarding age and gender. After participants provided their consent, they reported their actual and perceived knowledge about AI. Next, to ensure that participants had a consistent definition of AI, we provided them with a definition of AI adapted from Zhang and Dafoe (2019) with several examples of AI applications to facilitate their understanding. Thereafter, participants answered the questions addressing the IPMI model’s variables in the context of AI in a randomized order as a means to eliminate the potential order effects. The institutional research ethics committee approved the questionnaire and procedure.

The participants had a mean age of 49 years old (SD = 15.66, range = 19–83) and a median of monthly household income between HKD 50,001 to 60,000 (approx. USD 6,400 to 7,600). By level of education, 46% of participants (n = 750) had at least a bachelor’s degree. Regarding their computer programming proficiency, 33.1% of participants (n = 450) reported having no basic knowledge of or experience in programming.

Measures

Actual Knowledge About AI

Participants answered 10 true-or-false statements (1 = true, 2 = false,3 = unknown) concerning their actual knowledge about AI (Roland Berger, 2020), such as “The key technology of AI is based on machine learning.” For each statement, the correct answer was assigned a score of 1, whereas wrong or “unknown” answers were assigned a score of 0 (see Table A1 in the Supplementary Materials). We totaled the scores for those 10 statements to create the scale for actual knowledge about AI (M = 5.72, SD = 2.13, range = 0–10).

Perceived Knowledge About AI

We assessed participants’ perceived knowledge about AI using a single item (Akin et al., 2020). Participants indicated the extent to which they believed that they understood AI on a scale ranging from 0 (a complete lack of understanding) to 100 (complete understanding everything related to AI). The mean score was 58.78 (SD = 21.93, range = 0–100).

Media Attention to AI-Related Information

Integrating the media typology proposed by Gil de Zúñiga et al. (2012) and practical media usage among citizens in Hong Kong (Kemp, 2023), we assessed media attention to AI-related information using five items. Those items covered the following media channels: (a) traditional media (e.g., television, newspapers, and magazines), (b) social media for video sharing (e.g., YouTube), (c) photo-sharing social media (e.g., Instagram and Xiaohongshu), (d) social networking sites/applications (e.g., Facebook and WeChat), and (e) news applications (e.g., HK01, Channel C HK, and TVB NEWS). Participants reported the frequency with which they had paid attention to AI-related information via those five media channels in the past month on a 7-point scale ranging from 1 (never) to 7 (always). We calculated the average score of the five items to create the scale for media attention (M = 4.46, SD = 1.29, α = .89).

Interpersonal Communication About AI

We assessed participants’ interpersonal communication about AI with three items (Shi et al., 2022). Participants reported the frequency with which they had discussed AI with (a) family members, (b) classmates and/or colleagues, and (c) friends in the past month on a 7-point scale ranging from 1 (never) to 7 (always). We averaged the responses to those three items to create a scale for interpersonal communication (M = 4.22, SD = 1.48, α = .89).

Presumed Others’ Media Attention to AI Related Information

We assessed participants’ presumption of others’ attention to media reporting AI-related information with three items (Ho et al., 2020). Participants reported the estimated frequency with which (a) family members, (b) friends, and (c) the Hong Kong public paid attention to AI-related information in the past month on a 7-point scale ranging from 1 (never) to 7 (always). We averaged the responses for those three referent groups to create a scale for others’ presumed media attention (M = 4.43, SD = 1.23, α = .81).

Perceived Descriptive Norms

Participants rated their level of agreement with three statements on a 7-point scale ranging from 1 (strongly disagree) to 7 (strongly agree): (a) most people close to me, (b) most people I know, and (c) most people important to me use AI technology for work and/or learning (Fishbein & Ajzen, 2010). We averaged the responses for the three items to create the scale for perceived descriptive norms (M = 4.62, SD = 1.27, α = .89).

Perceived Injunctive Norms

Participants rated the extent to which they thought the referent groups would approve of their use of AI technology for work and/or learning on three items using a 7-point scale ranging from 1 (strongly disapprove) to 7 (strongly approve) adapted from Fishbein and Ajzen (2010). The referent groups included (a) people close to me, (b) people I know, and (c) people important to me. We averaged the responses for the three items to create the scale for perceived injunctive norms (M = 4.89, SD = 1.09, α = .87).

Intention to Use AI Technology for Work And/Or Learning

We assessed participants’ likelihood of using AI technology for work and/or learning in the following month using a 7-point Likert scale ranging from 1 (very unlikely) to 7 (very likely), M = 4.94, SD = 1.40.

Table A2 in Supplementary Materials reports the correlations between study variables.

Analytical Procedure

Structural Equation Modeling for IPMI Model

We used the lavaan package (version 0.6–15) in R studio (version 4.2.2) to perform structural equation modeling (SEM) with the maximum likelihood method (MLM) estimator, which is robust with non-normal data. To test H1–H6, we examined the relationships among the IPMI model’s variables using SEM (see Figure 1). Age was introduced as a control variable because the use of AI and perceptions of AI vary across age groups (Chien et al., 2019). The overall model remains the same after introducing more demographic variables (See Figure D1 in Supplementary Materials). In addition, all reverse models did not fit the data well (See Table C1–C3 in Supplementary Materials).

Figure 1. The overall influence of the presumed media influence model.

N = 1,360. Age was introduced to the model as control variable.

***p < .001.

Multigroup SEM Between Individuals with Different Levels of Actual/Perceived Knowledge

To answer RQ1 and RQ2, we first employed multigroup CFA models for the latent variables to examine the measurement invariance between individuals with different levels of actual/perceived knowledge. In the multigroup CFA models, configural invariance, metric invariance, and scalar invariance were tested in three hierarchical steps.

After establishing the measurement invariance, we conducted multigroup comparisons between people with low and high levels of actual/perceived knowledge in SEM. The baseline model was established by restricting item loadings, intercepts, and all of the path coefficients as being equal across the low and high actual/perceived knowledge groups. Next, a series of multigroup comparisons were performed in which we released the constraint of each path coefficient one at a time to compare the difference in chi-squared values with the baseline model and to examine whether the specific path significantly differed between the low and high actual/perceived knowledge groups.

Results

The Overall IPMI Model

The SEM model for the overall IPMI model presented in Figure 1 revealed an acceptable model fit: χ²/df = 6.52, CFI = .944, TLI = 0.933, RMSEA = .064, 90% CI of RMSEA = [.060, .067], SRMR = .053.

The results of SEM showed that media attention to AI-related information was positively associated with presumed others’ attention to AI-related information (β = .249, p < .001), as was interpersonal communication(β = .730, p < .001). Presumed others’ media attention to AI-related information was positively associated with perceived descriptive norms (β = .769,p < .001) and perceived injunctive norms (β = .764, p < .001). Meanwhile, perceived descriptive norms (β = .344, p < .001) and perceived injunctive norms (β = .412, p < .001) were positively associated with intention to use AI technology for work and/or learning. Therefore, the data were consistent with H1 to H6.

Multigroup SEM: High versus Low Actual Knowledge About AI

For actual knowledge, we used participants’ mean score for actual knowledge about AI (M = 5.72, SD = 2.14) to categorize participants as having low(M ≤ 5.72, n = 589, 43%) or high (M > 5.72, n = 771, 57%) actual knowledge about AI. The results of multigroup CFA showed that measurement invariance was established across the low and high actual knowledge groups(see Table A3 in Supplementary Materials).

However, a series of chi-squared tests for the multigroup comparisons revealed that the differences in the path coefficients of the IPMI model’s variables across the low and high actual knowledge groups were non-significant (see Table A4 in Supplementary Materials and Figure 2)

Figure 2. Multigroup analyses for low versus high actual knowledge groups (on the left) and low versus high perceived knowledge groups (on the right).

The paired path coefficients in bold represent significantly different coefficients between two groups. Age was introduced to the model as a control variable.

***p < .001

Multigroup SEM: High versus Low Perceived Knowledge About AI

For perceived knowledge, we used participants’ mean score for perceived knowledge about AI (M = 58.78, SD = 21.93) to categorize participants as having low (M ≤ 58.78, n = 579, 43%) or high (M > 58.78, n = 781, 57%) perceived knowledge about AI. The results of multigroup CFA showed that measurement invariance was established across the low and high perceived knowledge groups (see Table A3 in Supplementary Materials).

A series of chi-squared tests for the multigroup comparisons revealed that the difference in the association between media attention and presumed others’ media attention to AI-related information was significant, χ²(1) = 24.171, p < .001. In particular, the positive influence of media attention on presumed others’ media attention was stronger in the high perceived knowledge group (β = .331, p < .001) than in the low perceived knowledge group (β = .181, p < .001). The results also showed a significant difference in the association between interpersonal communication and presumed others’ media attention, χ²(1) = 23.513, p < .001. The positive influence of interpersonal communication on presumed others’ media attention was stronger in the high perceived knowledge group (β = .773, p < .001) than in the low perceived knowledge group (β = .669, p < .001). In addition, the differences in the association between presumed others’ media attention with the perceived norms were significant, χ²(1) = 6.558, p = .010 for perceived descriptive norms, and χ²(1) = 6.471, p = .011 for perceived injunctive norms. In detail, the positive influence of presumed others’ media attention on perceived descriptive norms was stronger in the high perceived knowledge group (β = .794, p < .001) than in the low perceived knowledge group (β = .680, p < .001). The positive influence of presumed others’ media attention on perceived injunctive norms was also stronger in the high perceived knowledge group (β = .775, p < .001) than in the low perceived knowledge group (β = .677, p < .001). The difference in the remaining paths between the IPMI model’s variables across the low and high perceived knowledge groups was non-significant (see Table A5 in Supplementary Materials and Figure 2).

Discussion

By consulting the IPMI model and its extensions, our study revealed the indirect media effects on individuals’ intention to use AI technology for work and/or learning. Moreover, we advanced the IPMI model by distinguishing actual knowledge from perceived knowledge and documented the latter as a boundary condition for the IPMI model through a series of multigroup SEM analyses. In detail, found that the model’s pattern did not manifest differently for individuals with low versus high levels of actual knowledge about AI. However, the pattern did differ between individuals with low and high levels of perceived knowledge about AI. Those findings theoretically advance the IPMI model and contribute to the literature about indirect and conditional media effects on the adoption of AI technology.

The IPMI Model and Its Underlying Mechanisms

Our findings, in replicating the application of the IPMI model and its extensions, demonstrate that both media attention and interpersonal communication are key antecedents of presuming others’ attention to focal media content. Consistent with the IPMI model’s propositions, individuals’ media attention was positively associated with presumed others’ media attention to the same content across various media outlets (Gunther & Storey, 2003). Moreover, interpersonal communication was positively associated with presumed others’ media attention, which corroborates previous research involving the IPMI model (Shi et al., 2022). However, contrary to Shi et al. (2022) findings, the effect size of interpersonal communication in our study was greater than personal media attention’s, for both factors served as antecedents of others’ presumed media attention. Those findings suggest that in the early diffusion of a trending topic, whose information richness is less than that of mainstream topics (e.g., COVID-19), individuals might rely more on interpersonal communication to obtain information and consequently estimate others’ media attention.

In terms of the normative mechanisms of the IPMI model, we found that presumed others’ media attention was positively associated with perceived descriptive and injunctive norms, which, in turn, positively influenced individuals’ intention to use AI technology for work and/or learning. Those findings indicate that in the early diffusion of technology adoption in a society, individuals estimate others’ acceptance of the behavior under the influence of media content and subsequently make decisions to engage in such behavior. Indeed, the literature on social normative influences has long considered norms as communication phenomena and that individuals can internalize normative influence through various channels of communication (Rimal & Lapinski, 2015). In other words, theories on social normative influences tend to emphasize the direct influence of communication on individuals’ perceived norms about a certain behavior. On the contrary, the IPMI model and its empirical evidence, including our findings, indicate the indirect effects of communication on individuals’ perceived norms, mediated through one’s estimation of others’ exposure to the information regarding the focal behavior. Thus, individuals’ normative perceptions seem to stem from two sources: direct communication and their own estimation of others’ exposure to the communication. Although testing that assumption was beyond the scope of our research, we believe that it is valuable to combine the IPMI model and the theory on social normative influences in investigating how individuals construct and adjust their normative perceptions via communication.

Perceived Knowledge As a Boundary Condition

Our analytical technique (i.e., multigroup SEM technique) allowed comparing the full IPMI model between individuals with high vs. low actual and perceived knowledge, which enabled us to investigate the boundary condition for the pattern of, instead of a single path within, the IPMI model. The findings showed that perceived knowledge, not actual knowledge, served as the boundary condition for the IPMI model. This suggests that the indirect media effects posited by IPMI model were stronger among individuals with high perceived knowledge than ones with low perceived knowledge.

Although actual knowledge and perceived knowledge are associated, they are conceptually distinct from each other (Driscoll & Salwen, 1997; Radecki & Jaccard, 1995). For one, perceived knowledge but not actual knowledge has been regarded as a factor of metacognition, meaning one’s awareness of their knowledge in a specific domain (Koriat & Levy-Sadot, 1999). Furthermore, our findings suggest that perceived knowledge enhanced the associations between media attention, presumed others’ media, and normative perceptions. However, the relationship between normative perceptions and the intention to adopt AI technology was not significantly different between the low vs. high perceived knowledge groups. Such a consistent relationship between perceived norms and AI use intention across groups echoed with the theoretical framework as well as empirical studies about technology adoption (Venkatesh & Davis, 2000). In addition, several meta-analyses have revealed that normative perceptions have consistently shown moderate effect sizes on behavioral intentions to accept technology (Chong et al., 2022; Schepers & Wetzels, 2007). These findings align with the pattern observed in our study. It suggests that the influence of normative perceptions on the intention to adopt technology may be stable across groups with different levels of knowledge.

On the other hand, the pattern of the IPMI model was similar between individuals with low and high levels of actual knowledge. This suggests that media effects are contingent on subjective perceptions, such as perceived knowledge, rather than solely relying on objective measures of competence. In other words, the process of media effects on behavioral outcomes can be seen as a cognitive procedure in which individuals are motivated to observe and analyze how reality operates based on their subjective estimations, rather than hinging on objective measures, such as actual knowledge.

Practical Implications

Our findings revealed that both media channels and interpersonal communication matter in promoting the adoption of AI technology for working and learning. Governments and other organizations could conduct large-scale media campaigns, which could educate the public to ethically and responsibly use AI technology. As media campaigns often induce interpersonal communication regarding the campaign topic (Southwell & Yzer, 2007), such media campaigns about AI would facilitate interpersonal discussion about AI. Both media campaigns and induced interpersonal communication are expected to facilitate the AI adoption. In addition, as the normative influence on AI adoption was found to be stable across different levels of actual/perceived knowledge groups, it indicates that social influence and environment are imperative in the adoption of AI for working and learning purposes. To promote AI use for such purposes, communication efforts could target at the organizational level, such as companies and universities, instead of the individual-level, to nurture supportive social norms to encourage AI use for personal development and employability.

Limitations and Future Research

First, the data we collected is cross-sectional in nature, which offers limited causal inferences. To address this limitation, we examined reverse models of our study model, and the results suggested the reverse models did not fit the data well. Nevertheless, we acknowledge that such statistical analyses still cannot determine causality among the IPMI variables we examined. In fact, Sun (2013) pinpointed that the current formulation of the IPMI model primarily serves as a depiction of how media effects occur, rather than providing a comprehensive explanation for why these effects occur and elucidating the underlying causality, despite containing causal propositions. To advance the theory building, future research on the IPMI needs to investigate the underlying causality between IPMI variables, such as through longitudinal surveys or experiments.

Second, due to the scarcity of literature on the role of actual knowledge and perceived knowledge in the IPMI framework, we investigated the role of the two types of knowledge in the framework separately to address the research gap. It also provides a parsimonious solution. However, these two types of knowledge are not orthogonal with each other. Future research efforts can segment individuals into four groups based on their levels (i.e., high and low) of these two types of knowledge and examine how the IPMI pattern differs across the four groups.

Third, two-thirds of our participants reported having at least some basic understanding of computer programming, which could be attributed to the well-developed curriculum on computer programming implemented in Hong Kong since 2013 (Cremer, 2023). Such a prevalence of digital and information literacy might be a practical reason why the normative mechanism of presumed media influence shapes individuals’ intention to adopt AI. However, the well-educated sample limits the generalizability of the current findings, making them less applicable to the countries or regions with limited computer education or underdeveloped information and communication technologies (ICTs). Future research could explore how indirect media effects shape AI adoption in those underdeveloped ICT areas.

Conclusion

In our study, we employed the IPMI model and its extension to examine the indirect effects of media attention and interpersonal communication on the adoption of AI technology through normative mechanisms. We discovered that both media attention and interpersonal communication acted as antecedents of the presumed media exposure of others, which influenced descriptive and injunctive norms and subsequently individuals’ intention to use AI for work and/or learning. Moreover, we identified that the indirect media effects posited in the IPMI model were stronger among individuals with high perceived knowledge than ones with low perceived knowledge. The findings theoretically advance the IPMI model and offer insights to promote the adoption of AI technology.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Data Availability Statement

The dataset, code for data analyses, and supplementary materials supporting the findings of this study are available on the OSF webpage (https://osf.io/h43cn/?view_only=aa677ec3f949445fa67ba57ff9dd4a43).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/08838151.2024.2377244

Additional information

Notes on contributors

Zixi Li

Zixi Li is a doctoral student in the School of Communication, Hong Kong Baptist University. Her research interests include persuasion, health and environmental communication, and meta-analysis.

Jingyuan Shi

Jingyuan Shi is an Associate Professor at the Department of Interactive Media, Hong Kong Baptist University. Her research interests converge at the intersection of persuasion, health communication, and new communication technology.

Yinqiao Zhao

Yinqiao Zhao is a doctoral student in the School of Communication, Hong Kong Baptist University. His research interests include journalism studies and media effects.

Bohan Zhang

Bohan Zhang is a doctoral student in the School of Communication, Hong Kong Baptist University. His research interests include media effects and mobile communication.

Bu Zhong

Bu Zhong is a Professor at the Department of Interactive Media and the Dean of the School of Communication, Hong Kong Baptist University. His research centers on the convergence of communication, technology, and human behavior.