Exploring Consumer Acceptance of AI-Generated Advertisements: From the Perspectives of Perceived Eeriness and Perceived Intelligence

Abstract

The rapid popularity of ChatGPT has brought generative AI into broad focus. The content generation model represented by AI-generated content (AIGC) has reshaped the advertising industry. This study explores the mechanisms by which the characteristics of AI-generated advertisements affect consumers’ willingness to accept these advertisements from the perspectives of perceived eeriness and perceived intelligence. It found that the verisimilitude and imagination of AI-generated advertisements negatively affect the degree of perceived eeriness by consumers, while synthesis positively affects it. Conversely, verisimilitude, vitality, and imagination positively affect the perceived intelligence, while synthesis negatively affects it. Meanwhile, consumers’ perceived eeriness negatively affects their acceptance of AI-generated advertisements, while perceived intelligence positively affects their willingness to accept AI-generated advertisements. This study helps explain consumers’ attitudes toward AI-generated advertisements and offers strategies for brands and advertisers for how to use AI technology more scientifically to optimize advertisements. Advertisers should cautiously assess the possible impact of AI-generated advertisements according to their characteristics, allowing generative AI to play a more valuable role in advertising.

1. Introduction

With the rapid and continued development of AI, AI-generated content (AIGC)—which refers to the use of AI to generate content efficiently and automatically [1,2]—has emerged as one of the most forward-looking technologies. The high expected growth rate of AIGC has triggered a new round of technological revolutions in various industries. iiMedia Research [3] forecasts that the core market size of AIGC in China is expected to reach RMB 7.93 billion in 2023 and RMB 276.74 billion in 2028. In the advertising industry, various enterprises have flocked to the market. For example, in 2023, Meta has combined its new large-scale language model LLaMA 2 with advertising product Advantage+ to bring automation to advertisers. Despite currently being in their infancy, AI-generated advertisements are highly effective at assisting in the production of content compared with advertisements created entirely by humans. However, do AI-generated advertisements really capture consumers’ hearts and minds? What kind of AI-generated advertisements are more acceptable to consumers and what are the underlying mechanisms? This paper aims to address these questions in depth.

The advertising process can be broadly summarized into four key aspects: market positioning, creation, evaluation, and feedback [4]. Advanced AI technology has the potential to enhance each of these stages. It is worth mentioning that the creative output is the primary aspect that directly engages with consumers, while the other stages are typically conducted internally within a company, often remaining less transparent and less visible to consumers [5,6]. Therefore, this study specifically emphasizes the visual characteristics of AI-generated advertisements, with a particular focus on the creative output.

Although contemporary AI generation techniques have reached a certain level in closely mimicking human creativity, the study of AIGC visual arts has, to date, been neglected [5]. It has been suggested that the extent to which AIGC images match reality influences users’ attitudes towards them [7]. If the generated content deviates excessively from reality, the sense of alienation it creates makes it difficult for consumers to find it realistic and identify with it; in contrast, the more realistic the AI advertisement is, the more it can immerse consumers and generate a sense of conviction [8]. This inspires us to examine the characteristic of verisimilitude of AI-generated advertisements and its impact on consumers’ willingness to accept such ads.

The presentation of advertisements in life can be seen as an artwork, with its vitality deriving from the dynamic reflection of reality and continuous innovation. Separation from the real world might lead to ossification [9], impacting the vitality of AI-generated advertisements. A lack of vitality in advertising could be an indication of rigidity and stereotyping, potentially making it challenging to elicit empathy from consumers. The abundant material resources and extensive database content underlying AI have the potential to enhance advertisement vitality and establish an emotional connection with consumers, but this outcome remains to be seen. Therefore, it is necessary to study the effect of the vitality of AI-generated advertisements.

The extension of AIGC’s imagination may fill in the gaps of human thinking, potentially assisting advertisers in broadening the boundaries of their creations and presenting consumers with the intriguing possibilities of advertising imagination [10]. While human imagination is rich, it is inherently limited by the framework of personal cognition. The powerful integration ability of AI might break through this limitation, possibly enabling the creation of advertisements that transcend people’s existing cognition. Overall, the imagination of AI-generated advertisements could open up new avenues for advertising creativity, but the full extent of this potential remains to be explored.

With its robust integration capabilities, AI mainly derives its creations from memories acquired through the machine’s storage of vast data training sets. AIGC models, serving as a technological tool for batch content generation, undergo iterative training on an extensive collection of existing data assets, which are subsequently synthesized to create new content [10]. However, since AI’s ability to fully comprehend the human mind is still a subject of exploration, its creations may convey a perception of being synthetic [8,11]. Although the overall output appears satisfactory, there may still be imperfections in the accuracy of the details. The synthesis of AI-generated advertisements may become a problem.

In summary, this study aims to examine the characteristics of verisimilitude, vitality, imagination, and synthesis identified through existing research and cases of AI-generated advertisements, and to explore how these characteristics affect consumer acceptance of AI-generated advertisements.

Consumers are the ultimate judges of advertising effectiveness [5]; accordingly, consumers’ perceptions of AI-generated advertisements are covered in our conceptual model. On the one hand, the logic of AIGC creation is based on the synthesis and processing of the training set [12], and the generated content differs from manually produced content in terms of how it is perceived. Previous research found that people perceive a sense of eeriness when confronted with images that are not lifelike, especially when they see human-like characters [13]. This sense of eeriness is a subjective negative attitude and is often accompanied by fear [14]. Perceived eeriness negatively affects consumers’ appreciation [5] and trust [13] of AI advertisements, which, in turn, might affect their acceptance. On the other hand, AI, through advanced technologies such as deep learning and computer vision, generates content with a high degree of automation, capable of addressing a variety of advertising needs and demonstrating a high level of intelligence [15]. Intelligence is a functional dimension. AI-generated advertisements rely on AIGC software to realize fully automated generation from creativity to content, and the degree of human involvement is much lower than that of traditional ad production. With advanced AIGC tools, not only has the efficiency of advertisement production been significantly improved, but it has also given advertisements an unprecedented level of intelligence. Malhotra and Ramalingam [16] noted that consumers’ perception of subjective intelligence is closely related to the capabilities of AI, therefore, this study also considers consumers’ perceived intelligence of AI-generated advertisements.

The existing literature on AIGC comprises three main areas of research. The first is the detection of deepfake technology, methods of generation, and the impact of AIGC in different industries [8,17]. The second is the copyright or ethical issues raised by AIGC [18]. The third relates to the design and algorithms of generative frameworks for AIGC-specific scenarios [19]. In recent years, the power of AIGC in the advertising industry has gained a lot of attention from academics [8,20,21]. Researchers have mainly focused on understanding the differences between AI-generated and human-generated advertisements, especially from the viewpoint of consumer perception. This includes studying whether consumers prefer AI-generated or human-generated ads [22] and whether advertising appeals affect these preferences [20,21]. Additionally, some researchers have approached this topic from an emotional angle, exploring how consumers’ perceptions of empathy, guilt, and similar emotions influence their responses to AI-generated content [11]. However, there remains a need to enhance understanding of the intricate mechanisms underlying the characteristics of AI-generated advertisements and their impact on perception [5]. Therefore, we focus on AI-generated advertisements, summarizing the characteristics of AI-generated advertisements and exploring the influencing mechanisms from the perspectives of perceived eeriness and perceived intelligence.

This paper distinguishes the multifaceted effects of different characteristics of AI-generated advertisements on consumer perceptions. The findings offer a new perspective on the application of AIGC in advertising and highlight the complexity of ad design features as well as their significance in shaping consumer perceptions. By systematically analyzing how specific attributes of AI-generated advertisements impact consumer perceptions, including intelligence and eeriness, and further influence their acceptance willingness, this study establishes a theoretical framework for assessing the effectiveness of AIGC applications in advertising. This framework not only enriches existing research on advertising effectiveness but also provides a theoretical basis and direction for future empirical studies and practical applications. In addition, this research reveals that consumers’ perception of intelligence in AI-generated advertisements positively affects their willingness to accept these ads, whereas perceived eeriness has a negative effect on consumer acceptance. This insight provides valuable strategic guidance for advertisers, emphasizing the need to enhance ad creativity and appeal through AIGC technology while avoiding the elicitation of discomfort or eerie sensations among consumers to foster acceptance and recognition.

2. Literature Review and Hypotheses

2.1. SOR Theory

The stimulus-organism-response (SOR) model is an important theoretical model in the study of individual behaviors [23,24,25]. The theoretical framework consists of three components: stage S represents the input of information to stimulate the environment, stage O is the mechanism for psychological transformation after being influenced by the external stimulus, and stage R is the corresponding response made by the organism [23,24]. The SOR theory proposes that external stimuli can influence people’s cognitive and emotional states, and thus their behavioral responses [25,26]. Accordingly, utilizing the SOR theory can facilitate a more thorough analysis of people’s perceptions, behaviors, and reactions to advertisements, ultimately enabling the design of more effective marketing programs.

The SOR theory is frequently employed to examine consumer behavior in fields such as marketing and e-commerce, as it offers a comprehensive and effective framework for understanding consumer responses [26,27,28]. Nagano et al. [28] used the SOR theory to explore the impact of advertising speech on consumers’ perceived emotional state, and thus their behavioral intention. Based on this theory, Li et al. discussed strategies to improve advertising effectiveness and enhance user engagement [29]. There have also been studies using the SOR theory to examine the relationship between the characteristics of new technology and people’s perceptions and acceptance of it, and they have found that the characteristics of new technology can influence people’s perceptions, which subsequently affect their attitudes towards acceptance [30,31]. It is evident that the SOR theory is widely used to study the relationships between external stimuli, consumers’ psychological state, and consumers’ responses. Therefore, this study is based on the SOR theory, using it to explore the mechanism of the influence of characteristics of AI-generated advertisements on consumers’ willingness to accept these advertisements.

2.2. AIGC and AI-Generated Advertisements

AIGC, a branch of AI that originates from machine learning, is an AI technology that is trained with a large amount of data and can automatically create new content based on what has been learned [32,33]. Midjourney, Stable Diffusion, Make-A-Video, and other AIGC tools have drawn the attention of scholars worldwide to AI’s powerful ability to generate text, speech, images, and video, spurring related research [34].

Recently, scholars are increasingly focusing on AIGC research in business domains, such as advertising content design [5,35,36]; product innovation [37]; virtual influencers [38,39]; and brand voice [40], among others.

Consumer attitudes toward AIGC are also a concern in the commercial context. Previous research has found that people can harbor stereotypes about machines [5,41,42], which affects their reactions to AIGC [5]. Common stereotypes of machines have both negative and positive aspects [5]. Negative bias towards AI [43] and algorithmic aversion [44] are factors that negatively influence users’ acceptance of AIGC. Negative emotions such as perceptions of eeriness towards AIGC and general uneasiness about machines [5] generate negative evaluations. In addition, deepfakes—although time- and cost-efficient [45,46]—creates concerns among consumers owing to the inherent deception involved [11,47]. However, there is also a positive side to consumers’ reactions to AIGC: this is mainly reflected in the processing of information, where it is often assumed that machines will be more objective and less biased than humans [5,41,48]. This positive stereotype of machines is known as the “machine heuristic” and refers to mental shortcuts attributed to machine features or machine-like operations [42,49]. Positive evaluations of AI are often associated with machine heuristics [50]. Sivathanu et al. found that deepfake has made advertisements so realistic that they were believed to provide accurate information about the product—increasing consumer trust—which positively influenced the willingness to shop online [17].

In recent years, the application of AIGC in the field of advertising has attracted the attention of academia. Vakratsas and Wang argued that AI is flexible enough to support the generation of creative advertising ideas in an automated manner or with the help of human input [51]. Campbell et al. constructed a framework for understanding consumer responses to AI-manipulated advertisements, highlighting the impact on perceptions of verisimilitude and creativity [8]. Besides, Chaisatitkul et al. found that both users and industry professionals have a generally positive attitude towards AI-generated advertisements, seeing AI as enhancing efficiency and acceptance by the audience [52]. Some scholars are concerned about the differences between AI- and human-generated content. Ananthakrishnan and Arunachalam demonstrated that AI-generated content is preferred by a significant percentage of respondents and is comparable to human-created content in terms of creativity [22]. Bakpayev et al. revealed that consumers have favorable attitudes towards AI-created cognitive-oriented advertising but form lower evaluations of AI-created emotion-oriented content [21]. Song et al. found that advertisements with rational appeals improved visit intention for AI-generated advertisements more effectively, while emotional appeals were more attractive when the declared creator was human [20]. Arango et al. showed that, when consumers recognize AI-generated images in charitable donation advertisements, their empathic response may be reduced, leading to decreased willingness to donate [11].

In summary, existing research points to substantial inconsistency in consumer responses to AIGC and AI-generated advertisements. The two types of evaluations, positive and negative, also illustrate the complexity of the mechanisms by which AIGC affects consumers’ perceptions and the importance of further research.

2.3. Characteristics of AI-Generated Advertisements

As AI technology continues to evolve, content generated using AI algorithms is likely to play a dominant role in the future of advertising and marketing [8,11,53]. When presented with AI-generated advertisements, consumers’ most direct response should be to process the advertisement in terms of its effect and the message conveyed—non-specialized consumers are seldom interested in what the invisible programming and instruction inputs in the AI-generated advertisements process are like [6]. Therefore, the artistic characteristics of AI-generated advertisements are indispensable research topics.

2.3.1. Verisimilitude

Verisimilitude in advertising refers to the degree of authenticity that consumers can perceive [8]. Although AI-generated advertisements have prompts from designers, they—unlike advertisements produced entirely by humans—lack the blend of human rational thought and artistic sensibility that manually generated advertisements possess.

Although machine-generated works may be highly accurate, there is still room for improvement in the handling of details. In particular, to make products more relevant to real life and for consumers to feel more immersed, most advertisers choose real people to appear in their advertisements [6]. The degree of verisimilitude of the characters can be reflected by low-tolerance features such as hair, skin texture, facial expressions, and hand movements. Inconsistencies in these aspects are often the root cause of the questionable verisimilitude of AI-generated advertisements. Li points out that the verisimilitude of AI works affects the user’s feelings and experience [54]. Research by Song and Shin suggested that characterization stimulates people’s inner sense of eeriness when it is not close to reality [13]. At the same time, the greater the verisimilitude, the less likely consumers are to perceive the advertisement as false [8]. Based on the above analysis, we propose the following hypothesis:

H1a: 

Verisimilitude has a negative influence on the perceived eeriness of AI-generated advertisements.

H1b: 

Verisimilitude has a positive influence on the perceived intelligence of AI-generated advertisements.

2.3.2. Vitality

Vitality is defined as the vital force, power, or principle possessed or manifested by creatures [55]. In artwork, vital force is interpreted as a vibrant, uplifting, and vivid presentation of the work [56]. Currently, AI-generated advertisements are capable of generating vivid images [11]. However, compared with human-produced advertisements, many AI-generated advertisements are repurposed image resources within a database [8]. It is difficult to reflect human thinking and emotions in an automated manner. Therefore, the vitality conveyed by AI-generated advertisements will be different from that of manually produced advertisements.

Vitality is often reflected in works of art. In order to create great works of art, artists must infuse vitality into their works to reflect the level and quality of their work [57]. When advertisements are given new connotations and vitality, they appear lively and vivid, and consumers receive the messages conveyed by the advertisements in a good mood [58]. Therefore, the more vividly AI-generated advertisements present the product or convey the message, the more consumers’ perception of the quality of the advertisements is improved, which in turn enhances the perception of the intelligence of AI-generated advertisements’ capabilities. At the same time, consumers would not perceive the advertisements as stiff or rigid [59]. However, if the AI-generated artwork lacks vitality, then people will perceive the colors and lines of this artwork as dull, stereotypical [60], and detached from real life, which would create a sense of eeriness [5]. Based on the above analysis, we propose the following hypothesis:

H2a: 

Vitality has a negative influence on the perceived eeriness of AI-generated advertisements.

H2b: 

Vitality has a positive influence on the perceived intelligence of AI-generated advertisements.

2.3.3. Imagination

Imagination is not the ability to produce or copy existing works, but the ability to create novel works [61]. Imagination enables one to create new things, such as novel works of art and scientific achievements [62]. AI shows strong potential for artistic creation [63], generating works that can produce unexpected effects after deep learning from a large amount of data within the framework of algorithms [6,8]. With the assistance of AI, it is possible to expand the limited creative ideas of humans to tens of thousands [64], or even more, instantaneously, reflecting the extensive imagination of AI advertising.

Imagination is one of the most important factors that makes an artistic painting unique and impressive [63]. Research shows that creativity has a positive impact on various consumer perceptions such as the assessment of advertising quality [65]. Evaluations of advertising creativity are sometimes combined with one or more dimensions related to the quality or artistic value of the produced advertisement [66,67]: when advertisements are more creative, viewers tend to perceive high creativeness of the AIGC technology involved, which increases their perception of intelligence. Therefore, when AI-generated advertisements are highly creative compared with other advertisements, viewers will favor the highly creative advertisements significantly more, and their perceived intelligence may appear higher [68]. In contrast, when AI-generated advertisements are low in creativity or output pieces are appropriate, they will be perceived as weird or eerie [69]. The lack of imagination in AI-generated advertisements may lead consumers to perceive the content as less thoughtful on the part of the brand or marketer. Because AI algorithms rely heavily on existing databases and works, the resulting output may become increasingly homogenized over time [22], reducing the perceived quality and intelligence of the ads. Based on the above analysis, we propose the following hypothesis:

H3a: 

Imagination has a negative influence on the perceived eeriness of AI-generated advertisements.

H3b: 

Imagination has a positive influence on the perceived intelligence of AI-generated advertisements.

2.3.4. Synthesis

The synthesis is the bricolage characteristic experienced in AI-generated advertisements. It arises because, in the process of generating AI-generated advertisements, already existing resources in the database and newly added resources can both be used [8]. AI-generated works are often described via text or diffused on the basis of existing images, which can lead to the overall picture being well-grasped; however, the details are complex and variable owing to the lack of detailed description. As a result, individual elements of the picture may appear to be incongruous and distorted.

Although AI technology has facilitated the reorganization of digital images [8,11], it has also provided the conditions for false pastiche. We are concerned about the negative impact of synthesis, given it is often a source of false consciousness for consumers [11]. Advertisements are an effective way to create a brand image for consumers, and smoothly presented advertisements increase consumer goodwill toward the brand. Suppose the advertisement presents nothing more than a combined output of input footage with algorithmic augmentation, distortions will give consumers a negative view of the visual quality [70], which would make them skeptical about the intelligence. Material reorganization can also blur the line between real and fake, triggering a sense of eeriness in consumers [5]. Fluent and coordinated advertisements should take into account the overall layout and detailed design of the content to reduce the sense of being synthetic and enhance the aesthetic quality of the advertisement [71]. Based on the above analysis, we propose the following hypothesis:

H4a: 

Synthesis has a positive influence on the perceived eeriness of AI-generated advertisements.

H4b: 

Synthesis has a negative influence on the perceived intelligence of AI-generated advertisements.

2.4. Consumer Perceptions

Advertising is a consumer-oriented activity aimed at information dissemination, and consumers are the judges and receivers of the visual effects of advertisements [5]. Therefore, it is crucial to study consumers’ perceptions. AIGC is perceived differently from manually generated content in that people have stereotypes about machines [5], which affect their responses to AIGC [41,45]. Common stereotypes of machines may be both negative and positive [5]. Therefore, the characteristics of AI advertisements may affect consumers’ acceptance of them by influencing the perceived eeriness and perceived intelligence both negatively and positively.

Perceived eeriness refers to the sense of horror that people may experience when viewing AIGC [5]. This feeling of eeriness is regarded as a negative emotional state [13]. Research has found that the feeling of eeriness significantly affects people’s trust, and when the level of eeriness is higher, their trust is lower [13]. When consumers perceive AI advertisements as scary and eerie, they feel distrustful of the AI advertisements. The level of trust affects consumers’ acceptance and willingness [72], and if consumers develop negative emotions of distrust, they will have a defensive reaction toward these messages [73], which reduces their acceptance and willingness towards AI advertisements.

H5a: 

Perceived eeriness has a negative influence on the consumers’ willingness to accept AI-generated advertisements.

Perceived intelligence refers to the consumer’s overall evaluation of the ability of AI automatically generated advertisements, which acts as a trigger or driver of consumer behavior and attitude [74,75]. The greater the perceived intelligence of AI-generated advertisements, the higher the perceived quality of the generated advertisements, and the more the consumer perceives the advertisements to be trustworthy; consequently, the consumer has a higher willingness to accept the advertisements [76,77]. Gao et al. indicated that a high quality of advertisements can directly increase consumers’ willingness to accept them [78]. Therefore, consumers will have a higher willingness to accept AI-generated advertisements when they perceive these advertisements to be more intelligent.

H5b: 

Perceived intelligence has a positive influence on the willingness to accept AI-generated advertisements.

2.5. Mediation of Perceived Eeriness and Perceived Intelligence

The SOR theory states that when an individual is exposed to an external stimulus, it causes a cognition in the organism, which in turn leads to a response. In other words, behavior is not directly elicited by a stimulus, but undergoes a sequential process from the stimulus to the organism’s cognition [79].

In this study, we have presented the characteristics that may influence consumers’ willingness to accept AI-generated advertisements. These characteristics—verisimilitude, vitality, imagination, and synthesis—are derived from a combination of realistic advertisement content and a review of the existing literature. The verisimilitude of an AI work refers to how closely it resembles reality [8]. However, due to current technological limitations, AI creations are not yet perfect in every detail. When these imperfections deviate from existing conventions, such as having extra fingers or a missing arm, they can evoke a sense of eeriness [13]. Similarly, advertisements lacking vitality can lead to psychological rejection, making consumers feel uncomfortable and experience eeriness. When AI-generated advertisements perform poorly in terms of imagery, it may indicate deficiencies in their technical aspects, leading to a perception of eeriness [69]. Given that AIGC is based on the reorganization of database material [8], elemental incongruity occurs frequently. This abrupt synthesis of elements is also a source of eeriness. Eeriness is a negative feeling for consumers which affects their appreciation of AI works [5] and subsequently influences their willingness to accept AI-generated advertisements. In summary, we propose the following hypothesis:

H6: 

Perceived eeriness mediates the path from verisimilitude (H6a), vitality (H6b), imagination (H6c), and synthesis (H6d) to willingness to accept the AI-generated advertisements.

Consumers’ perception of AI is not solely negative; it also has positive aspects [5]. Intelligence is a significant functional dimension that distinguishes AI positively from traditional artificial methods. In the AI-related literature, intelligence is defined as the ability to achieve goals automatically [15]. If AI-generated advertisements are life-like, consumers will marvel at the powerful capabilities of AI technology and it will promote the generation of an intelligent sense [80]. AI can expand imagination and generate novel advertising content [64], often breaking through the framework of traditional advertising. In addition, AI advertisements with high vitality can deeply understand human emotions and thus show more humanized elements in advertisements. These allow consumers to experience the powerful intelligence of AI. But synthesis in AI-generated advertisements has a negative influence on perceived intelligence. Wang and Li indicated that perceived intelligence influences consumer behavior [74]. The greater the perceived intelligence of AI-generated advertisements, the higher the willingness to accept them among consumers [76,77]. Considering the above factors, we propose the following hypothesis:

H7: 

Perceived intelligence mediates the path from verisimilitude (H7a), vitality (H7b), imagination (H7c), and synthesis (H7d) to willingness to accept the AI-generated advertisements.

The research model of this study is shown in Figure 1.

3. Methodology

3.1. Survey Procedure

To test the theoretical model proposed in this study, the authors adopted a questionnaire survey. To help subjects understand the theme of the survey, a brief introduction describing the AI-generated advertisement is presented at the beginning of the formal questionnaire. In the points for attention, we noted that only those who have viewed an AI-generated advertisement before can participate in the research. Subjects were required to indicate whether they had ever watched an AI-generated advertisement. The following set up was subsequently presented: please recall an AI-generated advertisement that impressed you deeply; recall as many details and feelings as possible. This imagined scenario could not only mobilize the characteristics of an AI-generated advertisement in the subjects’ memory, but also help the research team to screen subjects who answered the survey conscientiously.

Other key variables in the research model were measured after subjective questions; these included verisimilitude, vitality, imagination, synthesis, perceived eeriness, perceived intelligence, and willingness to accept. At the end of the questionnaire, demographic information of the subjects was collected. The key variables in the questionnaire were designed based on established scales used in previous studies (stated in Section 3.2). To enhance its relevance and accuracy, we incorporated specific details and adapted the questions to suit our research context of AI-generated advertisements while maintaining the integrity of the original scales.

3.2. Measurement Development

Seven key variables were examined in this study, namely, verisimilitude (VE), vitality (VI), imagination (IM), synthesis (SY), perceived eeriness (PE), perceived intelligence (PI), and willingness to accept (WA). All variables were measured with the Likert seven-point scale. Verisimilitude was measured using three items adopted from Campbell et al. [8]. For the measurement items on the vitality of AI-generated advertisement displays, we refer to the studies by Yan [9] and Zheng et al. [81]. For the assessment of imagination, we refer to the studies by Zhou and George [82], Scott and Bruce [83], and Tierney, Farmer, and Graen [84]. For the assessment of synthesis, we refer to Arango, Singaraju, and Niininen [11] and Whittaker et al. [85]. Perceived eeriness was measured using four items adopted from Wu and Wen [5]. Perceived intelligence was measured using three items adopted from Parayitam, Kakumani, and Muddangala [86], and Moussawi and Koufaris [87]. Willingness to accept was measured using three items adopted from Vijayasarathy and Jones [88], Pavlou [89], and Chen and Tan [90].

Prior to the formal survey, we conducted a pre-test to ensure that the respondents could understand our questions correctly and to guarantee the reliability and validity of the scale items. During this phase, we conducted the questionnaire with a sample of 110 participants who were similar to our target population. Based on their feedback, we revised the questionnaire, making necessary changes to improve clarity and understandability. Some items that did not meet the desired standards of reliability or validity were modified or removed. The final content of the items can be found in Appendix A.

3.3. Sampling and Data Collection

This study utilizes a popular and professional questionnaire distribution platform in China, Credamo, “https://www.credamo.com/home.html#/ (accessed on 6 June 2024)”. The platform covers a variety of sections such as questionnaire design, matching survey, eye movement experiment, and data collection. Credamo is currently widely used in market research by both universities and enterprises. In the design of questionnaires on the Credamo platform, the quality of data can be improved by using screening questions. The survey data were collected in China, and this study started collecting experimental data in September 2023 and ran until December 2023 when a total of 1154 responses were collected. To ensure the quality of the questionnaire, the items adapted from English were translated into Chinese by researchers. After confirming that the items were correct, we distributed the questionnaire. The subjects who denied having watched any AI-generated advertisements were screened out. Accordingly, we retained 1147 valid responses, and the effective sample rate was 99.4%. The response time of these subjects exceeded 300 s, and the sources covered most provinces in China.

4. Data Analysis and Results

Structural equation modeling (SEM) is a multivariate statistical analysis method whose core lies in the integration of factor analysis and path analysis. Structural equation modeling can be primarily categorized into covariance-based structural equation modeling (CB-SEM) and partial least squares-based structural equation modeling (PLS-SEM). PLS-SEM was used in this study to validate the theoretical model. The reasons for this is that the PLS-SEM is suitable for this study. First, the PLS-SEM method is applicable to data presenting a non-normal distribution [91], relaxing the data requirements of CB-SEM [92]. The skewness statistics of the variables range from −2.745 to 2.038 and the kurtosis statistics range from 1.131 to 9.199, both of which do not qualify for normality. Second, PLS-SEM has a good effect on the verification and evaluation of theoretical models, especially for complex models [93]. We conducted PLS-SEM analysis using SmartPLS 4.0.9.2. Hair et al. [94] proposed that two model tests should be followed when using PLS-SEM: (i) the measurement model and the (ii) structural model.

4.1. Subject Demographic Information

The demographic information of the valid subjects is listed in

Table 1. Demographic Information.

Demographic	Frequency	%
Gender
Female	689	60.1
Male	458	39.9
Age (years)
25 and under	331	28.9
26–35	557	48.6
36–45	148	12.9
46–55	60	5.2
56–65	41	3.6
66 and above	10	0.9
Education level
Senior high school, secondary vocational school and below	35	3.1
Undergraduate, junior college	913	79.6
Master	153	13.3
PhD or above	46	4.0
Career
Student	202	17.6
Office worker	765	66.7
Civil servant	61	5.3
Staff of public institutions	82	7.1
Freelance	18	1.6
Other	19	1.7
Income
RMB 2000 or less	126	11.0
RMB 2001–RMB 5000	195	17.0
RMB 5001–RMB 8000	274	23.9
RMB 8001 or more	552	48.1
Total	1147	100.0

. There were 689 female subjects (60.1%) and 458 male subjects (39.9%). The majority of the subjects were in the 26–35 years age group (48.6%).

4.2. Common Method Variance and Multicollinearity

Common method variance (CMV) may occur when research adopts same-source and self-reported data [95]. Liang et al. [96] proposed a method based on the PLS model to test the CMV on MIS Quarterly, adding a common method factor that contains all items of the constructs to the PLS model. For this research, the average value of the square of substantive factor loading (0.744) is significantly greater than the average value of the square of method factor loading (0.017). The results show that CMV is unlikely to be an issue for this research.

Multicollinearity occurs when two or more predictor variables have a highly linear relationship. The variance inflation factor (VIF) is an index used to measure whether multicollinearity exists. The VIF values of all items are in the range of 1.722 to 3.062, which are less than the currently recognized threshold of 10 in the literature [97]. We contend that multicollinearity is not an issue in the data.

4.3. Measurement Model

4.3.1. Reliability and Validity

Four indicators were selected to evaluate the reliability and validity, namely, the standardized loading, composite reliability (CR), Cronbach’s alpha, and average variance extracted (AVE), respectively. Detailed information on these indicators is presented in

Table 2. Construct validity, factor loading, CR, and AVE.

Construct	Items	Std. Loading	$\mathbf{Cronbach}\mathsf{\alpha }$	CR	AVE
Verisimilitude	VE1	0.881	0.838	0.841	0.755
	VE2	0.858
	VE3	0.868
Vitality	VI1	0.882	0.844	0.846	0.763
	VI2	0.871
	VI3	0.867
Imagination	IM1	0.815	0.849	0.853	0.688
	IM2	0.816
	IM3	0.860
	IM4	0.827
Synthesis	SY1	0.882	0.912	0.916	0.792
	SY2	0.875
	SY3	0.896
	SY4	0.905
Perceived eeriness	PE1	0.843	0.892	0.896	0.755
	PE2	0.883
	PE3	0.854
	PE4	0.893
Perceived intelligence	PI1	0.865	0.808	0.810	0.723
	PI2	0.849
	PI3	0.836
Willingness to accept	WA1	0.835	0.814	0.816	0.729
	WA2	0.840
	WA3	0.885

. The standardized loadings are above 0.700 and in an acceptable range. Cronbach’s alpha ranged from 0.808 to 0.912, with the values also exceeding the threshold of 0.700 [98]. The CR ranged from 0.810 to 0.916, exceeding the threshold of 0.700. These results indicate that the measurements are reliable. In addition, each AVE is above 0.500, showing good convergence validity.

4.3.2. Discriminant Validity

To test the discriminant validity between constructs, we used the Fornell–Larcker standard and cross-loading method.

Table 3. Discriminant validity and Fornell–Larcker standard.

Construct	IM	PE	PI	SY	VI	WA	VE
IM	0.830
PE	−0.629	0.869
PI	0.727	−0.695	0.850
SY	−0.574	0.753	−0.684	0.890
VI	0.754	−0.631	0.729	−0.628	0.873
WA	0.707	−0.693	0.806	−0.618	0.716	0.854
VE	0.622	−0.665	0.731	−0.664	0.680	0.699	0.869

Note(s): Diagonal elements are the square root of AVE for each construct. Imagination (IM), perceived eeriness (PE), perceived intelligence (PI), synthesis (SY), vitality (VI), willingness to accept (WA), verisimilitude (VE).

displays the values of the Fornell–Larcker standard. Diagonal elements are the square root of the AVE for each construct. The number below the diagonal represents the correlation coefficients. As shown in the table, the number on the diagonal of each column is the largest, indicating good discriminant validity.

Table 4. Discriminant validity and cross-loading.

Construct	IM	PE	PI	SY	VI	WA	VE
IM1	0.815	−0.502	0.574	−0.424	0.590	0.562	0.464
IM2	0.816	−0.481	0.588	−0.461	0.611	0.562	0.483
IM3	0.860	−0.584	0.663	−0.532	0.678	0.633	0.582
IM4	0.827	−0.514	0.580	−0.483	0.618	0.583	0.528
PE1	−0.477	0.843	−0.507	0.584	−0.469	−0.532	−0.501
PE2	−0.547	0.883	−0.610	0.661	−0.544	−0.618	−0.576
PE3	−0.611	0.854	−0.678	0.734	−0.630	−0.645	−0.646
PE4	−0.538	0.893	−0.603	0.621	−0.532	−0.600	−0.573
PI1	0.608	−0.633	0.865	−0.619	0.639	0.702	0.668
PI2	0.621	−0.598	0.849	−0.598	0.647	0.708	0.617
PI3	0.626	−0.539	0.836	−0.523	0.570	0.643	0.576
SY1	−0.461	0.625	−0.559	0.882	−0.543	−0.507	−0.563
SY2	−0.490	0.627	−0.573	0.875	−0.527	−0.537	−0.562
SY3	−0.542	0.701	−0.646	0.896	−0.585	−0.570	−0.620
SY4	−0.544	0.719	−0.646	0.905	−0.577	−0.581	−0.614
VI1	0.616	−0.522	0.622	−0.557	0.882	0.602	0.594
VI2	0.634	−0.532	0.623	−0.557	0.871	0.592	0.572
VI3	0.717	−0.594	0.663	−0.533	0.867	0.676	0.613
WA1	0.587	−0.589	0.652	−0.500	0.578	0.835	0.573
WA2	0.608	−0.574	0.692	−0.507	0.597	0.840	0.583
WA3	0.615	−0.611	0.719	−0.573	0.655	0.885	0.632
VE1	0.503	−0.560	0.618	−0.573	0.589	0.601	0.881
VE2	0.604	−0.632	0.682	−0.626	0.630	0.641	0.858
VE3	0.505	−0.533	0.599	−0.524	0.546	0.574	0.868

Note(s): For loadings and cross-loadings, the bold values shaded in gray are the indicator loading values corresponding to the constructs and others are cross-loading values. All indicator loadings are significant (p < 0.001). Imagination (IM), perceived eeriness (PE), perceived intelligence (PI), synthesis (SY), vitality (VI), willingness to accept (WA), and verisimilitude (VE).

displays the values of cross-loading. Each variable’s loading on its own items is significantly higher than for its cross-loading with other items. This result also indicates good discriminant validity.

4.4. Structural Model

4.4.1. Path Coefficient

After the measurement model was evaluated, PLS-SEM was used to examine the structural model. The analysis results are shown in Figure 2.

Table 5. Results of the path analysis.

H	Path	$\mathsf{\beta }$	M	S.E.	T	p	f2	Remarks
H1a	VE → PE	−0.186	−0.185	0.043	4.371	0.000	0.042	Supported
H1b	VE → PI	0.285	0.285	0.037	7.780	0.000	0.116	Supported
H2a	VI → PE	−0.053	−0.053	0.039	1.344	0.179	0.003 NS	Unsupported
H2b	VI → PI	0.185	0.185	0.036	5.079	0.000	0.039	Supported
H3a	IM → PE	−0.196	−0.197	0.044	4.412	0.000	0.044	Supported
H3b	IM → PI	0.288	0.287	0.038	7.576	0.000	0.111	Supported
H4a	SY → PE	0.484	0.484	0.046	10.432	0.000	0.325	Supported
H4b	SY → PI	−0.213	−0.213	0.037	5.834	0.000	0.074	Supported
H5a	PE → WA	−0.256	−0.257	0.036	7.158	0.000	0.107	Supported
H5b	PI → WA	0.628	0.627	0.034	18.403	0.000	0.645	Supported

Note(s): ^NS: not significant. Verisimilitude (VE), vitality (VI), imagination (IM), synthesis (SY), perceived eeriness (PE), perceived intelligence (PI), willingness to accept (WA).

shows the detailed data for the path coefficients. At the 0.05 level of significance, consumers’ perceived eeriness of AI-generated advertisements was significantly negatively affected by verisimilitude ($\beta$ = −0.186, p < 0.001) and imagination ($\beta$ = −0.196, p < 0.001), and significantly positively affected by synthesis ($\beta$ = 0.484, p < 0.001). Vitality ($\beta$ = −0.053, p = 0.179) has no significant negative effect on perceived eeriness. Consumers’ perceived intelligence was significantly positively affected by verisimilitude ($\beta$ = 0.285, p < 0.001), vitality ($\beta$ = 0.185, p < 0.001), and imagination ($\beta$ = 0.288, p < 0.001), and significantly negatively affected by synthesis ($\beta$ = −0.213, p < 0.001). H1a,b, H2b, H3a,b, and H4a,b were supported. H2a is not supported. Perceived eeriness ($\beta$ = −0.256, p < 0.001) significantly negatively affects consumers’ willingness to accept AI-generated advertisements. Perceived intelligence ($\beta$ = 0.628, p < 0.001) significantly positively affects consumers’ willingness to accept AI-generated advertisements. H5 (a, b) is supported. Furthermore, it is widely recognized in the literature that R² values of 0.500 and 0.750 represent moderate and strong explanatory power, respectively [99]. Based on the values of R², vitality, imagination, synthesis, and verisimilitude jointly explained 64.5% of the variance in perceived eeriness and 69.7% of the variance in perceived intelligence. Perceived eeriness and perceived intelligence jointly explained 68.3% of the variance in willingness to accept.

This study also evaluated the cross-validated redundancy (Q²) and effect size (f²) of the structural model. The Q² values of perceived eeriness (Q² = 0.476), perceived intelligence (Q² = 0.498), and willingness to accept (Q² = 0.493) all exceeded 0, showing good predictive correlation of the prediction model [100]. Regarding effect size (f²), synthesis (f² = 0.325) was the most important predictor of perceived eeriness, and verisimilitude (f² = 0.116) was the most important predictor of perceived intelligence. The effect of perceived intelligence (f² = 0.645) on consumers’ willingness to accept exceeded 0.35 [94], indicating a large effect size.

4.4.2. Mediating Effect

We tested for indirect effects with 5000 samples using the Bootstrapping program.

Table 6. Indirect and mediating effects.

H	Path	$\mathsf{\beta }$	M	S.E.	T	p	95% CI		Remarks
							Lower	Upper
H6a	VE → PE → WA	0.048	0.048	0.014	3.486	0.000	0.023	0.076	Supported
H7a	VE → PI → WA	0.179	0.179	0.025	7.058	0.000	0.129	0.227	Supported
H6b	VI → PE → WA	0.014	0.014	0.011	1.258	0.208	−0.006	0.037	Unsupported
H7b	VI → PI → WA	0.116	0.116	0.024	4.826	0.000	0.070	0.166	Supported
H6c	IM → PE → WA	0.050	0.051	0.014	3.668	0.000	0.026	0.080	Supported
H7c	IM → PI → WA	0.181	0.180	0.027	6.794	0.000	0.129	0.234	Supported
H6d	SY → PE → WA	−0.124	−0.124	0.018	6.726	0.000	−0.162	−0.089	Supported
H7d	SY → PI → WA	−0.134	−0.133	0.023	5.900	0.000	−0.179	−0.090	Supported

Note(s): Verisimilitude (VE), vitality (VI), imagination (IM), synthesis (SY), perceived eeriness (PE), perceived intelligence (PI), willingness to accept (WA).

displays the results of the mediation effect test. In regard to the relationship between the verisimilitude and willingness to accept, there was a significant mediating effect of both perceived eeriness (β = 0.048, 95% CI = [0.023, 0.076]) and perceived intelligence (β = 0.179, 95% CI = [0.129, 0.227]). For the relationship between the vitality of the AI-generated advertisements and consumers’ willingness to accept, there was a significant mediating effect of perceived intelligence (β = 0.116, 95% CI = [0.070, 0.166]). The mediating effect of perceived eeriness (β = 0.014, 95% CI = [−0.006, 0.037]) was not significant. Concerning the relationship between the imagination and willingness to accept, there was a significant mediating effect of both perceived eeriness (β = 0.050, 95% CI = [0.026,0.080]) and perceived intelligence (β = 0.181, 95% CI = [0.129,0.234]). In regard to the relationship between synthesis sensibility and willingness to accept, there was a significant mediating effect of both perceived eeriness (β = −0.124, 95% CI = [−0.162, −0.089]) and perceived intelligence (β = −0.134, 95% CI = [−0.179, −0.090]). H6a,c,d and H7a,b,c,d were supported. H6b is not supported.

5. Conclusions and Discussions

5.1. Conclusions

Recent times have seen rapid advancements in the development of AIGC. The advertising industry, as a business that relies heavily on content output, has taken advantage of this opportunity to accelerate the efficiency of content production [5,8]. This article uses the SOR theory to explore the impact of the characteristics of AI-generated advertisements on consumers’ willingness to accept these advertisements through the perceived eeriness and perceived intelligence of the advertisements. The specific conclusions of this research are as follows.

First, the verisimilitude and imagination of AI-generated advertisements negatively affect consumers’ perceived eeriness, and synthesis has a positive effect on perceived eeriness. Meanwhile, verisimilitude, vitality, and imagination positively affect consumers’ perceived intelligence, and synthesis has a negative effect on perceived intelligence. These conclusions further confirm the studies of Campbell et al. [8], Wu and Wen [5], and Arango et al. [11], who believe that verisimilitude created by AI will weaken consumers’ awareness of falsity, and creation will negatively affect consumers perception of eeriness. The difference is that previous research has devoted less attention to the effect of these characteristics on perceived intelligence [5]. The synthesis of AI-generated advertisements is also noteworthy, and has yet to be fully explored by scholars. While academic attention has been directed towards synthetic media [85] and synthetic content [11], the accuracy of the majority of image generators is not yet satisfactory. This paper further investigates the characteristics of synthesis in the domain of AIGC. Therefore, consideration should be given to enhancing the verisimilitude, vitality, and imagination of AI-generated advertisements, and diminishing the synthesis of the picture presentation. Advertisers should be discerning about AIGC content, and strive to find high-quality data assets that fit the brand’s tone as well as performing high-quality model training. They should seek to make AI-generated advertisements more vivid and reduce dullness, bringing consumers closer to the advertising content. After receiving the AIGC advertisement produced by the advertisers, the brand can make the necessary post-production adjustments to improve the overall quality of the advertisement and diminish the sense of integration in the picture.

Second, consumers’ perceived eeriness has a negative impact on their willingness to accept AI-generated advertisements, whereas their perceived intelligence positively influences their willingness to accept such ads. This study finds that perceived eeriness, often accompanied by feelings of strangeness and unease [14], represents a negative emotional perception, which is consistent with previous studies [5]. In terms of perceived intelligence, some prior studies have focused on the objective evaluation of ability [15]. However, this paper posits that perceived intelligence plays a mediating role in the acceptance of AI-generated advertisements, aligning with the positive impact on consumer attitudes observed in previous studies to some extent [74,75]. Advertisers should carefully check the content of AI-generated advertisements, focusing on both the overall message and finer details. Avoiding making consumers perceive them as ‘weird’ is one important way to enhance their acceptance of AI-generated advertisements. Perceived intelligence can serve as a criterion for measuring the quality of advertisements. High perceived intelligence enhances consumers’ favorable impression of AI-generated advertisement content, making them more willing to accept these ads.

5.2. Theoretical Implications

First, this study examines the characteristics of AI-generated advertisements, enriching existing theoretical research on AI-generated content (AIGC). As a cutting-edge technology currently attracting significant attention, AIGC has gained the interest of numerous scholars in academia [8,35]. Advertisements confer an image of the promoted product, making the content of the advertisement a crucial aspect of its generation [101]. Previous studies have primarily focused on the algorithmic aspects of AIGC [19], which is undoubtedly an important research direction. However, the technological processes involved in content generation are invisible and difficult for consumers to understand [6]. Therefore, this study shifts the focus to the visualization of AI-generated advertisements. As Ford [35] suggested, future research on AI advertisements should concentrate on enhancing consumers’ appreciation of such content. Unlike previous scholars who have focused on the differences in advertising attitudes between AI-generated content and human-generated content [21,22], this research enriches understanding of the effect of the visual characteristics of AI-generated advertisements on consumers’ willingness to accept these ads.

Second, this paper focuses on consumers’ perceptions of AI-generated advertisements, examining them from the dual perspectives of perceived eeriness and perceived intelligence. Given the widespread adoption of AI technology, it is necessary to understand consumers’ attitudes towards the advertisements generated by AI. Current research on consumer perspectives on AI-generated advertisements is still in its early stage [5]. While some previous studies have indicated that many consumers prefer AI-generated advertisements to human-generated ones [22], other studies have found human-generated advertisements to be more appealing in many instances [11,20]. This inconsistency in research findings highlights the necessity of this study. To address this gap, this study empirically demonstrates the mediating effect of perceived eeriness and perceived intelligence on consumers’ acceptance of AI-generated advertisements based on the SOR theory. Prior research has shown that perceived eeriness, the feeling of unease or strangeness consumers may experience when encountering AI-generated content, can negatively impact their appreciation of and trust in AI advertisements [5,13], potentially hindering acceptance. In addition, previous research has also suggested that perceived intelligence has the potential to positively influence consumer attitudes towards advertisements [75,76]. However, the existing literature has not simultaneously considered the dual mediating effect of perceived eeriness and perceived intelligence on AI-generated advertisements. By exploring these dual perspectives, this study provides a comprehensive understanding of consumers’ attitudes towards AI-generated advertisements and further enriches the research on consumer perspectives in the emerging field of AI-generated advertisements.

5.3. Managerial Implications

In addition to the theoretical implications, this study provides useful insights to support advertisers’ use of AIGC tools to produce advertisements.

First, introduce elements of real-life scenarios and add artificial optimization when producing AI-generated advertisements to increase vitality. Introducing elements of real-life scenarios in advertisements can make consumers perceive AI advertisements as vivid and life-like. However, although AI can automatically generate complete advertisements, AI-generated advertisements have a different vitality to that embodied in human-generated content, so they should be handed over to a human for review and adjustment after AI has completed the initial generation of advertisements. If the human reviewer identifies raw, mechanical details to be improved, these can be fine-tuned, modified, and optimized to ensure that the advertisements are vivid and imaginative to help to eliminate the sense of eeriness that consumers may perceive in regard to AI advertisements.

Second, diverse data and cross-domain learning should be provided to improve the imagination of AI-generated advertisements. AI-generated advertisements are based on learning from rich and diverse data. The input of advertising materials of different styles, themes, and emotions by advertisers can help AI broaden its scope of imagination; moreover, by introducing various types of images, text, audio, and other data, AI can learn and create from multiple dimensions.

Third, advertisers should seek to improve algorithms as well as continuously iterate and optimize to reduce the synthetic feel of AI advertisements. Advertisers can explore new ways of producing AI-generated advertisements by experimenting with different parameter settings, algorithm configurations, or generation techniques. For the initial results of AI-generated advertisements, iterations should continue. More advanced technologies can be introduced to automatically fix the processing of parts with a sense of segmentation as well as to choose the overall appropriate color scheme, font style, and layout for advertisements. This would not only make the advertisements more harmonious and unified as a whole, but also improve their quality and reduce the sense of eeriness caused by algorithmic synthesis.

Finally, advertisers should enhance model training in order to improve the verisimilitude of AI-generated advertisements. The quality of AI-generated advertisements is limited by the training data used to generate them. Providing high-quality and natural advertisement material as training data can help AI in better understanding and learning the principles and characteristics of advertisement design, and ensure that the generated content is more realistic and reliable. Furthermore, advertisers should seek to strengthen model training and increase the number of iterations of model training to allow AI to learn the details of advertising processing; strengthen AI’s realistic model; and improve the realism of AI-generated advertisements.

5.4. Research Limitations and Future Research

Although this study makes meaningful contributions to both theory and practice, the following areas for improvement remain. First, although this study focuses on consumers’ overall willingness to accept AI advertisements, it does not examine consumers’ product purchase willingness and brand attitudes. These topics could be explored in future studies. Second, different cultures may have different attitudes toward AI advertisements; in this study, the respondents were from an Eastern culture, and the results may not be generalizable to other populations. To obtain more generalized results, more consumers from different cultures should be included in future studies. Thirdly, only the intrinsic characteristics of AI advertisements were included as independent variables in the model of this study, while external variables that may affect consumers’ willingness to accept were not taken into account. Such relevant factors not included in the current research model can be explored in future research.