The Impact of User Experience on Continuous Usage Intention for AI-Generated Digital Music Platforms: Examining the Mediating Roles of Emotional Attachment and Technology Self-Efficacy

S-O-R Theory

One of the theoretical foundations of this study is the S-O-R (Stimulus-Organism-Response) model proposed by Mehrabian and Russell (1974). This model reveals the relationships between external stimuli (Stimulus), internal human states (Organism), and behavioral responses (Response), providing a framework for explaining how environmental factors influence human psychology and behavior. The logical rationale for selecting this theory lies in its unique advantages and high compatibility with the research context of this study. AI-generated digital music platforms are not merely tools but complex environments that provide rich sensory, social, and emotional experiences. The “Stimulus (S)” dimension of S-O-R theory can perfectly accommodate these multidimensional user experiences as inputs. More critically, its “Organism (O)” component provides us with a theoretical “black box,” allowing us to deeply explore how users’ internal psychological states—including emotions (such as emotional attachment) and cognition (such as technology self-efficacy)—are activated by these external experiences. Finally, this theory connects these internal psychological changes with external behavioral intentions (continuous usage intention) through the “Response (R)” component, forming a complete causal chain. Therefore, the process-oriented and integrative perspective of S-O-R theory, which can simultaneously accommodate experience, emotion, and cognition, makes it an ideal theoretical tool for interpreting user behavior on AI-generated digital music platforms.

In the context of this study, S-O-R theory provides a theoretical perspective for exploring user experience and continuous usage behavior. External stimuli (S) refer to the platform’s multidimensional user experiences, internal states (O) refer to users’ levels of emotional attachment and technology self-efficacy, and behavioral responses (R) refer to users’ continuous usage intention. Users perceive different levels of experiential stimuli in their interactions with the platform, triggering changes in emotional connection and self-efficacy, ultimately driving their behavioral decisions to continue using the platform. Numerous empirical studies have confirmed the explanatory power of S-O-R theory for user behavior in digital product/service contexts (Kabadayi et al., 2023; Patel et al., 2024). However, existing literature mostly focuses on general digital products/services and still lacks systematic analysis of user behavior mechanisms in generative AI contexts. The unique attributes of AI-generated content platforms, such as human-machine co-creation, algorithmic recommendation, and intelligent interaction, may have differential impacts on S-O-R mechanisms. This study aims to expand the S-O-R framework by incorporating novel user experience dimensions and special mediation mechanisms in AI contexts, deeply exploring the user experience-response chain of AI-generated digital music platforms, with the goal of enriching the application of S-O-R theory in algorithm-driven intelligent service domains.

User Experience Theory

User Experience is an interdisciplinary research topic spanning human-computer interaction, cognitive psychology, and design studies. ISO (2019) defines it as “user’s perceptions and responses that result from the use or anticipated use of a product, system or service,” emphasizing that user experience originates from human-system interaction and encompasses the entire experience chain. Norman’s (2013) three-level emotional design theory provides a framework for understanding user experience, pointing out that user experience includes visceral, behavioral, and reflective levels, featuring multi-level and multi-dimensional characteristics. McCarthy and Wright (2004) further proposed a four-element user experience framework, including compositional, sensual, emotional, and spatio-temporal elements, providing theoretical support for characterizing the multi-dimensionality of user experience. Additionally, Pine and Joseph’s (1998) four-dimensional model of tourist experience and Oh et al.’s (2007) comprehensive tourist experience framework reveal the interaction of multiple elements in experience scenarios, offering insights for understanding the user experience of digital products and services.

However, existing research mainly focuses on general products and services, with less exploration of continuous usage behavior of generative AI products from the perspective of experience dimensions. Based on AI-generated digital music platforms, this study reconstructs the user experience framework from five dimensions: sensory experience, social experience, entertainment experience, content experience, and functional experience, combining the unique attributes of generative AI platforms with classical user experience theories. Specifically, sensory experience stems from Norman’s visceral level and McCarthy’s sensual element (McCarthy & Wright, 2004; Norman, 2013). AI-generated digital music platforms create audiovisual feasts through intelligent algorithms and immersive design, with personalized music recommendations and intelligent timbre matching stimulating users’ emotional responses and influencing their impressions and emotional connections to the platform. Social experience echoes the compositional element of McCarthy’s model (McCarthy & Wright, 2004), aligning with the platform’s social attributes. Specifically, this dimension refers to the interactions between users (user-user interaction), wherein users communicate, share, and collaborate based on common musical interests within the platform. These friendly social functions promote a sense of belonging and emotional attachment, becoming a key factor influencing user retention. Entertainment experience draws on experience economy theory (Pine & Joseph, 1998). The platform satisfies users’ auditory enjoyment and creative pleasure through algorithmic generation and interactive experiences, with quality entertainment experience driving continuous use. Content experience originates from the platform’s core value proposition, namely the massive, diverse, and high-quality AI-generated music works. The innovation, richness, and matching degree of content generation algorithms directly affect users’ willingness to use. AI technology decentralizes content production, making content personalization and user creativity release key issues. Functional experience is based on Norman’s behavioral level theory (Norman, 2013). The platform provides diverse tools for music creation, editing, and sharing, with the usability, practicality, and stability of functions affecting user task efficiency and overall experience. In the context of AI-empowered music creation, platforms need to provide low-threshold, personalized, and intelligent functional support to help users unleash creativity and achieve human-machine collaboration. Thus, the quality of functional experience has a crucial impact on users’ continuous usage intention.

In summary, the five dimensions of sensory experience, social experience, entertainment experience, content experience, and functional experience constitute the core user value proposition of AI-generated digital music platforms, reflecting their digital, intelligent, social, and entertainment characteristics. Although existing user experience theories provide a theoretical basis for this dimensional division, they are mainly limited to general digital service contexts and lack targeted examination of the unique experience mechanisms of generative AI platforms. This study aims to verify the applicability of this multi-dimensional experience framework through empirical analysis, enriching the application of user experience theory in the field of AI-generated creative services.

Emotional Attachment

Emotional attachment theory originates from attachment theory in the field of psychology, first proposed by British psychologist Bowlby (1969) when studying mother-infant relationships. He believed that an infant’s attachment to its mother is a natural response aimed at reducing distance with the attachment figure to enhance a sense of security, with an emotional bond existing between the two. Shultz (2007) extended attachment theory to the marketing field, discovering that this attachment emotion exists not only between people but also between people and objects. Subsequently, scholars have defined the connotation of emotional attachment based on different application contexts. For example, Park et al. (2006) distinguished emotional attachment from satisfaction and usage preference, defining it as the emotional bond connecting an individual to a brand. Maulana and Eckhardt (2007) found that users’ emotional attachment to websites stems from long-term human-computer interaction processes. M. Luo (2022) studied the influence of virtual experience marketing on consumer purchasing behavior, exploring how virtual interactions affect users’ attachment to online environments. Z. Li (2023) examined the impact of tourists’ participation in value co-creation on experience value and revisit intention, emphasizing how user participation in services promotes emotional attachment and drives long-term behavior. Emotional attachment can form between people and between people and objects. It acts like a bond, closely connecting users with products and users with other users and has a subtle influence on users’ continuous usage behavior. Combining existing research, this paper defines emotional attachment as: the long-term emotional experience and connection formed between users and the platform during the use of AI-generated digital music platforms. In essence, emotional attachment is an affective and relational construct, focusing on the user’s emotional bond with the platform, which is conceptually distinct from a user’s cognitive evaluation of their own operational skills.

Existing research confirms that many factors influence the formation of users’ emotional attachment, such as self-realization, positive emotions, social identity, and user experience. Thomson (2006) proposed that when consumers satisfy their needs and achieve self-improvement through products, they develop an attachment to the brand. Bartels and Zeki (2004) pointed out that if a product or brand can bring positive emotions, it will evoke users’ emotional attachment. Stokburger-Sauer et al. (2012) introduced social identity theory, finding that consumers’ perceived brand-self congruity and brand distinctiveness significantly affect their attachment to the brand. Meanwhile, the formation of emotional attachment has a positive impact on user behavior change. Hung (2014) demonstrated that brand attachment in online communities effectively promotes brand loyalty and word-of-mouth behavior, revealing that emotional attachment represents a deep bond formed through long-term use rather than short-term responses. This bond strengthens user-brand stickiness over time. Traditional research examining user experience and behavioral intentions primarily focused on satisfaction and trust as mediating variables, with emotional attachment receiving less attention. However, AI-generated digital music platforms offer unique value propositions—immersive experiences, personalized recommendations, and intelligent creative assistance—that facilitate stronger emotional connections through sustained platform interactions. Therefore, this study introduces emotional attachment as a mediating variable to explore how users’ multi-dimensional experiences influence continuous usage intentions.

Technology Self-Efficacy

Self-efficacy, proposed by American psychologist Bandura in social cognitive theory, refers to an individual’s belief and judgment about successfully executing a task, and is an important factor influencing behavior. Davis et al. further proposed Computer Self-efficacy, which is an individual’s judgment of their ability to use computers to complete specific tasks. With technological development, researchers have derived concepts such as mobile self-efficacy and internet self-efficacy. Technology self-efficacy is an extension of computer self-efficacy in specific technological fields. McDonald and Siegall (1992) first proposed this concept, referring to an individual’s level of confidence in their ability to learn and apply new technologies. Since then, numerous studies have explored its measurement, antecedent variables, and outcome variables. For example, Y. Wang et al. (2024) found in the context of English as a Foreign Language (EFL) teaching that technology self-efficacy significantly predicts students’ acceptance of educational technology. Xu et al. (2024) studied its impact on Information and Communication Technology (ICT) usage behavior in the education field in the post-pandemic era, emphasizing the importance of cultivating technology self-efficacy. In this study, technology self-efficacy refers to users’ evaluation and confidence level in their ability to use AI-generated digital music platforms, reflecting their sense of control over platform functions and ability to deal with problems. Crucially, it is a cognitive construct centered on the user’s internal assessment of their own capabilities, which is theoretically different from the affective connection they may form with the platform itself. In AI-generated digital creative platforms, users are not only content receivers but also creative participants. Faced with massive music element libraries, intelligent creative tools, and diverse templates, users need to possess certain musical literacy and technical abilities, both of which jointly determine their technology self-efficacy.

Technology self-efficacy has a significant impact on users’ adoption of new technologies and platforms. Igbaria (1995) found in the context of PC use that technology self-efficacy has a positive effect on perceived ease of use and usage intention. M.-H. Hsu and Chiu (2004) also pointed out that internet technology self-efficacy is a key factor influencing continuous usage intention of electronic services. The stronger the users’ technology self-efficacy, the more they tend to believe they can obtain the desired experience, thus being more willing to try and continue using related technologies. Recent research further reveals its role in the field of educational technology. Y. Wang et al. (2024) found that technology self-efficacy significantly predicts Chinese EFL learners’ acceptance of educational technology, outperforming other psychological factors such as achievement emotions. Xu et al. (2024) found perceived trust mediates the relationship between technology self-efficacy and ICT usage behavior, with perceived security and electronic word-of-mouth as moderators, emphasizing the importance of cultivating technological confidence. However, existing research focuses mainly on general information systems. In AI creative platforms, users must efficiently collaborate with AI for creation, posing new challenges to technology self-efficacy. AI technology lowers music creation barriers by enabling creation without professional knowledge, enhancing self-efficacy. Conversely, complex AI functions, massive material tags, and frequent recommendations may increase cognitive load, weakening technology self-efficacy. This study explores technology self-efficacy’s unique mechanism in AI-generated digital music platforms, providing theoretical support for optimizing platform design and user retention.

Continuous Usage Intention

Continuous Usage Intention is a key concept in information systems research, referring to users’ subjective willingness to continue using a technology or service after initial use (Bhattacherjee, 2001b). In the field of digital platforms and online services, continuous usage intention is crucial for the long-term success of platforms, as it directly affects user retention rates and platform vitality (Cho, 2016). Its theoretical foundation can be traced back to the Technology Acceptance Model proposed by Davis et al. (1989), which emphasizes perceived usefulness and perceived ease of use as key factors in users’ acceptance of new technologies. However, as research deepened, scholars found that initial acceptance and continuous use are two different behavioral decision processes. Based on this, Bhattacherjee (2001b) proposed the IS Continuance Model, integrating Expectation-Confirmation Theory, emphasizing the impact of user satisfaction and perceived usefulness on continuous usage intention. In recent years, with the iteration of technology and changes in user needs, the research framework of continuous usage intention has been continuously expanding. Kim and Oh (2011) pointed out that hedonic motivation plays an important role in users’ continuous use decisions. Hsiao et al. (2016) further explored the role of social factors in the continuous use of mobile social applications, finding that social interaction and social influence significantly affect user willingness. In the context of artificial intelligence and generative technology, H. Li et al. (2024) revealed the importance of innovation and personalized experiences for users’ continuous use of AI-generated services.

This study proposes a multi-dimensional user experience model based on the S-O-R framework for AI-generated digital music platforms’ continuous usage intention. S includes sensory, social, entertainment, content, and functional experiences. O comprises emotional attachment (user-platform connection) and technology self-efficacy (confidence in AI music tools). R is continuous usage intention. The model suggests experience dimensions directly and indirectly affect usage intention through emotional attachment and technology self-efficacy, considering both traditional information system factors and AI’s unique impact on music creation.

Research Model

The integrated model explores influencing mechanisms of users’ continuous usage intentions on AI-generated digital music platforms. It consists of three parts: S with five user experience dimensions covering various platform experiences; O with emotional attachment and technology self-efficacy as mediating variables; and R with continuous usage intentions as the dependent variable. The model hypothesizes that user experience dimensions directly and indirectly affect continuous usage intentions through the mediating variables, as shown in Figure 1.

OPEN IN VIEWER

Figure 1.

The theoretical model.

Research Hypotheses

Participants

This study investigates user behavior on “AI-generated digital music platforms,” specifically defined as platforms centered on AI-assisted creation where users play a dominant role by providing creative inputs (e.g., text prompts, melodies). This focus on active creators excludes “pure AI generation” models where users are passive listeners, as user experiences differ fundamentally between these contexts. Following approval by X University’s Academic Ethics Committee in December 2024, this study employed a questionnaire survey method. To avoid selection bias from traditional snowball sampling, we collaborated with a leading Chinese professional survey company (Wen Juan Xing) using their sample service function. The survey was conducted over 1 week in July 2024, with participants accessing the online questionnaire through WeChat or email links.

Several measures protected participant rights and well-being. The anonymous online questionnaire collected no personally identifiable information. A detailed informed consent form explained the research purpose, voluntary participation, and data confidentiality. Participants were informed of their right to withdraw without penalty and had to formally consent before starting. After eliminating questionnaires with patterned responses and high missing rates, 509 valid questionnaires remained for analysis.

The 509 valid respondents reveal a predominantly young user base, with 82.32% aged 18 to 35. Gender distribution was relatively balanced. Most participants were corporate employees/white-collar workers with college diplomas or bachelor’s degrees. Regarding platform usage, most had been using platforms for 1 to 6 months with high frequency (daily or near-daily). Detailed demographics are in Table 1.

OPEN IN VIEWER

Table 1.

Demographic Profile of Respondents (N = 509).

Category	Options	Frequency	Percentage (%)
Gender	Male	256	50.30
	Female	248	48.72
	Other/Prefer not to say	5	0.98
Age	18–25	218	42.83
	26–35	201	39.49
	36–45	57	11.20
	46–55	23	4.52
	Over 55	10	1.97
Education	Junior high school or below	21	4.13
	High school/Vocational school	81	15.91
	College diploma	193	37.92
	Bachelor’s degree	165	32.42
	Master’s degree	41	8.06
	Doctoral degree or above	8	1.57
Occupation	Student	50	9.82
	Corporate employee/White-collar worker	270	53.05
	Government/Public institution staff	80	15.72
	Freelancer	76	14.93
	Professional	33	6.48
Monthly income (CNY)	3,000 or less	74	14.54
	3,001–5,000	131	25.74
	5,001–8,000	206	40.47
	8,001–12,000	71	13.95
	Over 12000	23	4.52
	Prefer not to say	4	0.79
Duration of use (months)	Less than 1	26	5.11
	1–3	194	38.11
	3–6	148	29.08
	6–12	72	14.15
	Over 12	69	13.56
Usage frequency	Multiple times a day	12	2.36
	Once a day	227	44.60
	2–3 times a week	202	39.69
	Once a week or less	68	13.36

Measures

The scales used in this study have been validated in numerous studies and applied in the Chinese context. The original English scales were translated into Chinese and back-translated to ensure consistency in meaning between the two language versions. Unless otherwise specified, all scales used a 7-point Likert scale (1 = strongly disagree, 7 = strongly agree).

Sensory Experience was measured using a 3-item scale developed by Venkatesh and Davis (2000). The items include “The AI-generated digital music platform page is visually appealing,”“The overall color scheme and layout of the AI-generated digital music platform page are good,” and “The page classification and navigation design of the AI-generated digital music platform are clear, reasonable, and hierarchical.” The composite reliability (CR) and Cronbach’s α of this scale were .917 and .865, respectively.

Social Experience was measured using a 4-item scale developed by Leung (2003). Items include “I can smoothly communicate music creation experiences with other users on this platform,”“The AI-generated music I share can gain recognition and support from other users,”“I have established friendships with other music enthusiasts through the platform,” and “I can gain a sense of belonging and satisfaction when using this AI-generated digital music platform.” The CR and Cronbach’s α of this scale were .918 and .882, respectively.

Entertainment Experience was measured using a 3-item scale developed by Mehmetoglu and Engen (2011) and Oh et al. (2007). Items include “The process of creating and sharing music on this AI-generated digital music platform makes me feel pleased,”“I enjoy watching music works shared by other users on this AI-generated digital music platform,” and “Exploring the various functions provided by this AI-generated digital music platform is an entertainment experience in itself.” The CR and Cronbach’s α of this scale were .908 and .849, respectively.

Content Experience was measured using a 4-item scale developed by Waqas, Salleh, and Hamzah (2021). Items include “The music content provided by this AI-generated digital music platform is of high quality,”“This AI-generated digital music platform can provide personalized music content recommendations,”“This AI-generated digital music platform provides unique and novel music content,” and “The music content of this AI-generated digital music platform has high artistic value and is worth appreciating.” The CR and Cronbach’s α of this scale were .916 and .877, respectively.

Functional Experience was measured using a 3-item scale developed by Coker (2013). Items include “The songs recommended by the platform’s AI music recommendation are ones I like,”“The platform’s music generation functions are rich and diverse,” and “The platform’s music editing and adjustment function design can meet my needs.” The CR and Cronbach’s α of this scale were .906 and .845, respectively.

Emotional Attachment was measured using a 3-item scale developed by Thomson et al. (2005). Items include “I have an emotional connection (Affection) to this AI-generated digital music platform,”“Using this AI-generated digital music platform makes me feel pleasant (Passion),” and “I feel there is a bond (Connection) between me and this AI-generated digital music platform.” The CR and Cronbach’s α of this scale were .905 and .842, respectively.

Technology Self-efficacy was measured using a 4-item scale developed by Huffman et al. (2013) and Compeau and Higgins (1995). Items include “I can successfully use all functions of this AI-generated digital music platform even without guidance,”“I am confident that I can use this AI-generated digital music platform well even if I have never used a similar platform before,”“I can skillfully use this AI-generated digital music platform to complete the tasks I want,” and “For me, it is easy to learn to use all functions of this AI-generated digital music platform.” The CR and Cronbach’s α of this scale were .912 and .872, respectively.

Continuous Usage Intention was measured using a 3-item scale developed by Bhattacherjee (2001a). Items include “I am willing to continue using this AI-generated digital music platform without terminating its use,”“I am willing to recommend this AI-generated digital music platform to others,” and “I am willing to continue using this platform rather than other alternative AI-generated digital music platforms.” The CR and Cronbach’s α of this scale were .925 and .879, respectively.

Measurement Model

Reliability and Validity

Internal consistency reliability measures result stability and reliability. This study uses Cronbach’s Alpha (CA) and Composite Reliability (CR). CA ranges 0 to 1, with α > .7 indicating reliability and α > .9 high reliability. CR reflects construct indicator consistency, with CR > 0.7 acceptable, though Fornell and Larcker (1981) suggest CR > 0.6 suffices. Average Variance Extracted (AVE) tests convergent validity, showing how well observed variables explain latent variables. AVE ≥ 0.5 indicates good composite validity. Results are in Table 2.

OPEN IN VIEWER

Table 2.

Reliability and Validity.

Measurements	Items	Factor loading	T	p	α	CR	AVE
Sensory experience (SeE)	SeE1←SeE	.922	125.230	.000	.865	0.917	0.787
	SeE2←SeE	.881	77.683	.000
	SeE3←SeE	.858	57.548	.000
Social experience (SoE)	SoE1←SoE	.858	70.700	.000	.882	0.918	0.738
	SoE2←SoE	.870	82.943	.000
	SoE3←SoE	.863	73.836	.000
	SoE4←SoE	.845	66.803	.000
Entertainment experience (EE)	EE1←EE	.882	85.190	.000	.849	0.908	0.768
	EE2←EE	.883	80.258	.000
	EE3←EE	.863	60.136	.000
Content experience (CE)	CE1←CE	.852	70.114	.000	.877	0.916	0.731
	CE2←CE	.859	73.781	.000
	CE3←CE	.851	66.581	.000
	CE4←CE	.858	71.268	.000
Functional experience (FE)	FE1←FE	.852	57.763	.000	.845	0.906	0.763
	FE2←FE	.886	87.025	.000
	FE3←FE	.882	85.714	.000
Emotional attachment (EA)	EA1←EA	.865	66.284	.000	.842	0.905	0.760
	EA2←EA	.873	81.181	.000
	EA3←EA	.876	84.108	.000
Technology self-efficacy (TS)	TS1←TS	.861	73.357	.000	.872	0.912	0.722
	TS2←TS	.835	59.775	.000
	TS3←TS	.858	70.375	.000
	TS4←TS	.846	66.869	.000
Continuous usage intention (CUI)	CUI1←CUI	.910	106.401	.000	.879	0.925	0.805
	CUI2←CUI	.896	95.525	.000
	CUI3←CUI	.887	86.978	.000

Discriminant Validity

Discriminant validity analysis verifies statistical differences between constructs. Items from different constructs should not be highly correlated (correlation coefficient < .85), otherwise they may measure the same concept. This study uses the AVE method following Fornell and Larcker (1981): the square root of AVE for each factor should exceed its correlation coefficients with other factors. Results show diagonal AVE square root values exceed non-diagonal correlation coefficients, confirming good discriminant validity. Table 3 shows detailed results with diagonal representing AVE square roots and lower triangle showing correlation coefficients.

OPEN IN VIEWER

Table 3.

Discriminant Validity.

Constructs	SeE	SoE	EE	CE	FE	EA	TS	CUI
SeE	0.887
SoE	0.399	0.859
EE	0.159	0.184	0.876
CE	0.154	0.253	0.305	0.855
FE	0.138	0.209	0.126	0.267	0.874
EA	0.404	0.509	0.399	0.493	0.360	0.872
TS	0.433	0.510	0.457	0.455	0.409	0.501	0.850
CUI	0.312	0.355	0.381	0.370	0.300	0.650	0.378	0.897

Next, we used the Heterotrait-Monotrait (HTMT) ratio, which compares between-trait correlations to within-trait correlations. Results in Table 4 show all HTMT values below 0.85, indicating good discriminant validity between variables.

OPEN IN VIEWER

Table 4.

HTMT Discriminant Validity.

Constructs	SeE	SoE	EE	CE	FE	EA	TS	CUI
SeE
SoE	0.455
EE	0.184	0.212
CE	0.174	0.286	0.352
FE	0.155	0.239	0.146	0.308
EA	0.471	0.590	0.472	0.574	0.424
TS	0.495	0.581	0.530	0.519	0.476	0.585
CUI	0.356	0.404	0.441	0.422	0.346	0.756	0.431

Structural Equation Model

Predictive Relevance Indicator Q²

Q² measures predictive relevance of exogenous variables for endogenous variables (0–1 range). Values > 0 indicate predictive capability; negative values show no predictive power. Thresholds: 0.02 to 0.13 (small), 0.13 to 0.26 (medium), >0.26 (large). Q² values for EA, TS, and CUI are 0.380, 0.390, and 0.339 respectively, all exceeding 0.26, demonstrating strong predictive capability, as shown in Table 5.

OPEN IN VIEWER

Table 5.

Predictive Relevance Indicator Q².

Constructs	SSO	SSE	Q 2 ( =1-SSE/SSO)
EA	1,527	945.993	0.380
TS	2,036	1,242.648	0.390
CUI	1,527	1,009.668	0.339

Collinearity Diagnosis VIF

A collinearity diagnostic analysis was performed on the model. The results are shown in Table 6. It can be concluded that the VIF values between the measurement variables are all below 8, and the VIF values between the latent variable factors are also below 3, indicating that there is no collinearity in the model, as shown in Table 7.

OPEN IN VIEWER

Table 6.

Collinearity of Measurement Variable Indicators.

Items	VIF
CE1	2.119
CE2	2.192
CE3	2.177
CE4	2.235
CUI1	2.685
CUI2	2.444
CUI3	2.248
EA1	1.934
EA2	1.997
EA3	2.069
EE1	2.088
EE2	2.122
EE3	1.985
FE1	1.961
FE2	2.043
FE3	2.085
SeE1	2.747
SeE2	2.163
SeE3	2.116
SoE1	2.231
SoE2	2.307
SoE3	2.252
SoE4	2.156
TS1	2.269
TS2	2.000
TS3	2.200
TS4	2.108

OPEN IN VIEWER

Table 7.

Collinearity of Latent Variable Indicators.

Constructs	EA	TS	CUI
SeE	1.205	1.205
SoE	1.271	1.271
EE	1.126	1.126
CE	1.207	1.207
FE	1.106	1.106
EA			1.336
TS			1.336

Significance of Path Coefficients

Path coefficients assess hypothesis relationships. Standardized coefficients (−1 to 1) indicate correlation strength and direction. T-values, derived from path coefficients and standard deviations, determine significance using critical values: 1.96 (5%), 2.57 (1%), and 3.29 (0.1%). This study uses 5,000-sample bootstrapping to calculate path coefficients and T-values. Figure 2 shows model path coefficients, with results in Table 8.

OPEN IN VIEWER

Figure 2.

Results of the hypothesized structural equation model.

OPEN IN VIEWER

Table 8.

Path Coefficients of PLS Structural Equation Model.

Path	O	M	STDEV	T	p
SeE→EA	0.189	0.188	0.035	5.425	.000
SeE→TS	0.216	0.216	0.033	6.475	.000
SoE→EA	0.288	0.289	0.036	8.068	.000
SoE→TS	0.272	0.271	0.035	7.784	.000
EE→EA	0.208	0.208	0.033	6.251	.000
EE→TS	0.280	0.281	0.031	9.087	.000
CE→EA	0.282	0.281	0.036	7.866	.000
CE→TS	0.205	0.205	0.035	5.888	.000
FE→EA	0.173	0.174	0.034	5.031	.000
FE→TS	0.233	0.233	0.033	7.156	.000
EA→CUI	0.615	0.616	0.031	19.536	.000
TS→CUI	0.070	0.070	0.039	1.778	.076

Results show five user experience dimensions (SeE, SoE, EE, CE, FE) significantly positively affect both EA and TS (p < .001): SeE→EA (β = .189), SeE→TS (β = .126), SoE→EA (β = .289), SoE→TS (β = .272), EE→EA (β = .208), EE→TS (β = .280), CE→EA (β = .282), CE→TS (β = .205), FE→EA (β = .173), FE→TS (β = .233). EA→CUI shows significant positive effect (β = .615, p < .001), while TS→CUI shows no significant effect (β = .070, p > .05). H1-H11 are supported; H12 is not supported.

Effect Size (f²)

Effect size (f²) measures an independent variable’s impact magnitude by assessing how removing exogenous variables affects R² and endogenous variables. Using PLS, effect size is calculated as the difference between original and adjusted R² divided by (1 − R²). Cohen (1988) established thresholds: 0.35 (large), 0.15 (medium), and 0.02 (small). Table 9 shows the study’s effect sizes.

OPEN IN VIEWER

Table 9.

Effect Size Assessment f².

Constructs	EA	TS	CUI
SeE	0.060	0.085
SoE	0.132	0.129
EE	0.078	0.155
CE	0.133	0.077
FE	0.055	0.108
EA			0.494
TS			0.006

Bootstrap Mediation Effect Test

To demonstrate the mediating effect, this study used the Bootstrap mediation effect test method to examine whether the mediating effect is significant. The confidence interval was set to Bias Corrected (95%), with 5,000 resampling iterations. The results of the mediation effect test are shown in the Table 10 below.

OPEN IN VIEWER

Table 10.

Bootstrap Mediation Effect Test.

Path	O	M	STDEV	T	p	2.50%	97.50%
SeE→TS→CUI	0.015	0.015	0.009	1.673	.094	0.000	0.035
SeE→EA→CUI	0.116	0.116	0.022	5.161	.000	0.073	0.161
SoE→TS→CUI	0.019	0.019	0.011	1.735	.083	−0.001	0.042
SoE→EA→CUI	0.177	0.178	0.023	7.595	.000	0.133	0.225
EE→TS→CUI	0.019	0.020	0.011	1.702	.089	−0.001	0.043
EE→EA→CUI	0.128	0.128	0.022	5.795	.000	0.087	0.172
CE→TS→CUI	0.014	0.014	0.008	1.693	.090	0.000	0.033
CE→EA→CUI	0.173	0.173	0.024	7.176	.000	0.128	0.223
FE→TS→CUI	0.016	0.016	0.010	1.696	.090	−0.001	0.037
FE→EA→CUI	0.106	0.107	0.022	4.823	.000	0.063	0.150

Mediation analysis shows 95% confidence intervals for technology self-efficacy mediated paths (SeE/SoE/EE/CE/FE→TS→CUI) include 0 (non-significant), while emotional attachment mediated paths (SeE/SoE/EE/CE/FE→EA→CUI) exclude 0 (significant). Therefore, H13a-e are supported, H14a-e are not supported.

Discussion of Key Findings

We obtained several important findings from the current study. Firstly, the empirical results of this study support H1–H10, verifying that the five dimensions of user experience (sensory experience, social experience, entertainment experience, content experience, functional experience) have significant positive effects on both emotional attachment and technology self-efficacy. This finding is generally consistent with the existing user experience theoretical framework in the literature (Oh et al., 2007; Waqas, Hamzah, & Salleh, 2021), but presents a unique influence pattern in the specific context of AI-generated digital music platforms. Specifically, social experience (β = .288, p < .001) and content experience (β = .282, p < .001) have the most significant impact on emotional attachment, while entertainment experience (β = .280, p < .001) and social experience (β = .272, p < .001) have the most prominent impact on technology self-efficacy. This differentiated influence pattern reflects the unique attributes of AI-generated digital music platforms, namely that social interaction and content quality play key roles in cultivating users’ emotional connections, while entertainment and social aspects enhance users’ sense of technological control. This finding echoes the research of Xiang et al. (2022), providing theoretical support for understanding the differentiated impact of user experience dimensions in AI-generated digital music platforms.

Secondly, emotional attachment demonstrates a significant mediating role between user experience and continuous usage intention, validating hypotheses H11 and H13a–e. This study provides deeper insights into how this attachment is formed and how it functions as a driver of long-term engagement. The formation of emotional attachment is primarily channeled through social and content experiences. Specifically, social features that allow users to share their creations and receive recognition and support from a community foster a strong sense of belonging and validation, transforming the platform from a solitary tool into a shared creative space. Similarly, a high-quality content experience, characterized by personalized, novel, and artistically valuable AI-generated music, makes users feel understood and creatively empowered by the platform. Once this emotional bond is established, it becomes the most powerful direct driver of continuous usage intention, as evidenced by its substantial effect size (β = .615, p < .001). This attachment functions by creating a psychological anchor, shifting the user’s perception of the platform from a mere utility to a valued creative partner. It is this deep-seated, affective connection—rather than simple satisfaction—that motivates users to return, recommend the platform, and remain loyal. This finding not only extends the applicability of emotional attachment theory (Thomson et al., 2005) to emerging technology environments but also specifies the precise experiential pathways through which emotional connections are forged and sustained on AI-generated content platforms, a conclusion supported by recent studies highlighting the importance of emotional factors in complex technological environments (Mamun et al., 2023; Nafees & Sujood, 2025; Xiang et al., 2022).

Thirdly, a key finding of this study is that the direct effect of technology self-efficacy on continuous usage intention is not significant (β = .070, p > .05), and its mediating role is not supported, contradicting hypotheses H12 and H14a–e. This result contrasts sharply with conclusions from researchers like Compeau and Higgins (1995) in traditional technology acceptance studies, revealing unique user behavior patterns on AI-generated digital music platforms and challenging existing theoretical paradigms. We identify three fundamental reasons behind this “theoretical anomaly.” First, generative AI significantly lowers technical barriers to music creation, diminishing the importance of technical confidence. Unlike professional music software such as digital audio workstations, generative AI platforms offer “democratized creativity”—users transform ideas into musical works through simple text prompts or parameter adjustments without requiring complex music theory or operational skills. The AI handles most complex technical execution. In this “low-barrier creation” model, users function as “conductors” or “creative directors,” proposing ideas rather than implementing them technically. Consequently, users’ confidence in mastering platform functions no longer serves as a critical cognitive barrier or key driver for continued usage.

Second, AI-generated digital music platforms are inherently hedonistic, experience-oriented systems where emotional factors dominate decision-making (Fišer et al., 2025; Lerch et al., 2025). Users’ primary motivation involves experiencing creative joy, sensory pleasure, and emotional expression rather than completing utilitarian tasks. In such “sensory consumption” and “emotional experience” scenarios, user behavior follows emotional logic over instrumental rationality. Consequently, emotional attachment (β = .615, p < .001), reflecting the user-platform emotional bond, naturally outweighs technology self-efficacy—which reflects rational technical assessment—becoming the core psychological mechanism driving retention. This phenomenon reflects the unique human-computer collaboration model in AI music creation. Unlike other AI creative fields such as AI-assisted programing or scientific computing, music creation connects more intimately with human emotion. Users view AI as a “creative partner” that understands and executes vague emotional intentions rather than a “tool” requiring precise control. This relationship shift moves user evaluation focus from “Can I use it well?” (technology self-efficacy) to “Does it understand me and provide good experience?” (emotional attachment). This collaborative model specifically explains why technology self-efficacy effects diminish on AI-generated digital music platforms, distinguishing them from other creative platforms like AI painting or writing. The reduced impact of technology self-efficacy is not accidental but stems from generative AI’s “de-skilling” nature, platforms’“emotion-centric” attributes, and the emerging “human-computer collaboration” paradigm. This finding provides crucial insight for understanding AI-era user behavior: as AI capabilities advance, traditional model variables emphasizing human control and technical skill may gradually yield to variables capturing human-computer emotional interaction and experiential value.

Fourthly, through a systematic analysis of the total effects of various user experience dimensions, this study finds that social experience (total effect = 0.196) and content experience (total effect = 0.187) show the most significant impact. The prominence of these two dimensions can be understood by their fundamental roles in fulfilling users’ higher-order needs. Content experience represents the platform’s core value proposition, as high-quality, personalized, and novel AI-generated music directly satisfies the user’s primary creative and esthetic goals. Social experience, in turn, addresses a deeper, intrinsic human need for connection and recognition, transforming the solitary act of creation into a powerful social ritual. This hierarchy of importance aligns well with the perspective of Conservation of Resources Theory (Hobfoll et al., 2018). In the unique environment of AI-generated digital music platforms, the most valued resources users seek to gain are not merely functional but are centered on cognitive pleasure (derived from high-quality content) and emotional support (derived from social interaction). The other dimensions—functional, sensory, and entertainment—play crucial, albeit supportive, roles, highlighting their interplay. They function as foundational enablers. A seamless functional experience and an appealing sensory design are necessary prerequisites for users to create high-quality content efficiently and pleasantly. Likewise, an entertaining process enhances both content creation and social interaction. In essence, while functional, sensory, and entertainment experiences ensure the platform is usable and enjoyable, it is the quality of the content and the richness of the social interactions that ultimately provide the sustained, higher-order value that keeps users coming back.

Theoretical Implications

This research offers several significant theoretical implications for understanding user behavior in the era of generative AI. Firstly, our findings challenge the universality of traditional technology acceptance paradigms. As established in our discussion, the non-significant influence of technology self-efficacy—long considered a core antecedent of usage behavior (Bandura, 1997; Compeau & Higgins, 1995)—is a critical theoretical anomaly. This suggests that the boundary conditions of established models like the Technology Acceptance Model must be re-examined. Our study contributes by identifying a specific, emerging context where a core psychological antecedent—the sense of technological control—loses its predictive power, thus calling for a more nuanced application of these classic theories in the AI era.

Secondly, this study refines the application of the S-O-R theory in technologically advanced environments. The clear asymmetry in the mediating effects—where the emotional pathway (via emotional attachment) decisively outweighs the cognitive pathway (via technology self-efficacy)—enriches the S-O-R framework originally proposed by Mehrabian and Russell (1974). It demonstrates that the “Organism” state is not always a balanced interplay of parallel cognitive and emotional responses. In AI-driven creative contexts, the emotional pathway can become the dominant mechanism, a finding that deepens our understanding of user decision-making processes.

Lastly, this study extends user experience theory by revealing the shifting hierarchy of experiential dimensions. The prominence of social and content experiences over functional aspects suggests a re-weighting of the user experience construct is necessary for AI-native platforms (Hassenzahl & Tractinsky, 2006). In these environments, user value appears to be migrating from traditional usability and functionality (Norman, 2007) toward the quality of AI-generated content and the social interactions it fosters. This echoes the principle of context-dependency in user experience (Forlizzi & Battarbee, 2004) and provides an empirical foundation for developing new UX models tailored to the unique characteristics of generative AI.

Implications to Practice

The findings of this study provide profound and actionable practical guidance for AI-driven creative platforms. First, the research clearly indicates that social experience and content experience are the two strongest drivers with the greatest total effect on continuous usage intention. Therefore, platforms should invest core resources in these two dimensions to establish and strengthen users’ emotional attachment. Specifically, platforms should not merely stop at providing basic social functions such as likes and comments. To maximize the promotional effect of social experience on emotional attachment, deeper interactive mechanisms should be introduced. Currently, platforms like Suno AI already possess the “Continue this song” feature, allowing users to create derivative works based on others’ compositions, which represents the embryonic form of social co-creation. Platforms can further deepen this model by, for example, developing “AI-assisted collaborative arrangement” functions that allow User A to generate a melody through text input, then invite User B to add vocals or harmonies through humming, ultimately completing a collaborative work. Additionally, platforms can design social sharing incentive mechanisms. For instance, drawing inspiration from “Tianpuyue’s” AI music collection campaign in collaboration with Guangzhou cultural tourism, platforms can regularly host themed creative challenges (such as “Summer Cyberpunk” or “Graduation Songs”) or establish “This Week’s Popular AI Works” rankings, providing honor badges like “Certified Musician” for highly active creators. These features not only enhance users’ sense of participation and community belonging but also establish strong emotional bonds between users and between users and the platform through collaboration and recognition.

Furthermore, content experience optimization should not be limited to precise algorithmic recommendations. To enhance users’ emotional connection with content, platforms can introduce an “emotional tagging” system. Beyond traditional genre classifications, users or AI can tag music with emotional or scenario labels such as “healing,”“immersive,”“epic,” or “cyberpunk rainy night.” For example, when users input “write a jazz piece for listening to while driving alone on a rainy night” in Suno or Udio, platforms not only generate music but can also automatically attach tags like “rainy night,”“driving,”“jazz,” and “contemplative.” This not only enables emotion-based music discovery but also makes users feel that the platform “understands me,” thereby deepening emotional attachment. Additionally, platforms can enhance the artistic value perception by emphasizing the uniqueness and narrative quality of AI music. For instance, they can draw inspiration from “Tianpuyue’s”“Dream Composer” program, showcasing the collaborative process between AI and human artists, or display the creative prompts that generated the music alongside user works, allowing other users to see the creativity behind them and imbue algorithmic outputs with warm humanistic emotions.

Second, the findings of this study are not limited to AI-generated digital music platforms but also have important implications for other AI-driven creative technology domains. The core conclusion of this study—that the importance of emotional attachment exceeds technology self-efficacy—is likely applicable to AI creative platforms that share two core characteristics: significantly lowering technical entry barriers and having hedonic and esthetic experiences as primary user goals. For AI painting platforms (such as Midjourney, Kling AI, etc.), users similarly play the role of “art director” or “creative director” rather than “digital painter” requiring mastery of advanced painting techniques. The core experience is equally esthetic and emotional. Therefore, this study’s model is likely equally applicable. Platform designers should also focus on strengthening the artistic exchange atmosphere within communities (such as hosting themed exhibitions and establishing style challenges) and enhancing emotional resonance with content, rather than merely pursuing technical indicators of generated images. For AI writing platforms (such as iyunbi, ima.copilot, etc.), the applicability of this study’s model may depend on users’ specific writing objectives. For creative writing (such as poetry, novels, scripts), emotional connection (users’ emotional investment in stories and characters) may remain key. However, for utilitarian writing (such as technical reports, marketing copy, code assistance), factors from traditional technology acceptance theory, such as perceived usefulness, may significantly regain importance, and the role of technology self-efficacy may also become important again. Based on this, for the entire AI creative field, platform design paradigms are undergoing profound transformation. For platforms aimed at inspiring user creativity and providing esthetic experiences, the key to success is no longer merely providing a powerful “tool” but creating a “community” and “creative partner” where users can immerse themselves, establish emotional connections, and gain recognition.

Limitations and Future Research Directions

Although this study reveals the key mechanisms underlying users’ continuous usage intention for AI-generated digital music platforms, several limitations remain. First, this study primarily employs cross-sectional data for analysis, which cannot eliminate potential endogeneity issues between variables. Second, the scope of this study focuses on exploring the driving factors of continuous usage while not addressing mechanisms leading to user churn. Finally, the sample of this study consists entirely of Chinese users, and this geographical and cultural specificity poses significant limitations to the generalizability of our research findings. Cultural factors can profoundly influence user behavior, particularly in terms of emotional attachment and technology adoption. For instance, Chinese culture, typically characterized by collectivist values, may place greater emphasis on social harmony and community integration, which might amplify the strong influence of “social experience” on emotional attachment observed in this study. In contrast, users from more individualistic cultural backgrounds may form emotional attachment based on different experiential drivers (such as personal achievement or self-expression). Therefore, the relationships identified in this research model may not be universally applicable.

Based on these limitations, this study opens multiple directions for future theoretical exploration. The most important finding of this study—the non-significant mediation effect of technology self-efficacy—represents a noteworthy null result that warrants deeper theoretical explanation in future research. For example, researchers could examine the contextual specificity of this phenomenon by comparing whether technology self-efficacy similarly fails in other AI-generated creative platforms such as digital painting and writing, or by introducing new moderating variables (such as users’ familiarity with AI technology). Furthermore, future research should also consider the impact of AI technology’s “black box effect” on user trust and construct a “bidirectional behavioral model” that integrates user retention and churn. Additionally, future research should strive to validate this research model across different cultural contexts. A cross-cultural comparative study would be invaluable for testing the robustness of this study’s findings and revealing how cultural dimensions moderate the pathways between user experience, emotional attachment, and continuous usage intention.