Semantic Divergence in AI-Generated and Human Influencer Product Recommendations: A Computational Analysis of Dual-Agent Communication in Social Commerce

Abstract

The proliferation of generative artificial intelligence (AI) as an autonomous recommendation agent fundamentally challenges traditional paradigms of marketing communication. As AI systems increasingly mediate consumer–brand relationships, understanding how artificial agents construct persuasive discourse—distinct from human communicators—becomes critical for developing effective dual-channel marketing strategies. Grounded in Source Credibility Theory and the Computers Are Social Actors (CASA) paradigm, this study investigates the semantic and structural divergence between AI-generated product recommendations and human influencer marketing messages in social commerce contexts. Employing a mixed-methods computational approach integrating term frequency analysis, TF-IDF weighting, Latent Dirichlet Allocation (LDA) topic modeling, and BERT-based contextualized semantic embedding analysis (KR-SBERT), we examined 330 Instagram influencer posts and 541 AI-generated responses concerning inner beauty enzyme products—a hybrid category combining functional health claims with hedonic beauty appeals—in the Korean social commerce market. AI-generated responses were collected through a systematically designed query protocol with empirically grounded prompts derived from actual consumer search behaviors, and analytical robustness was verified through sensitivity analyses across multiple parameter thresholds. Our findings reveal a fundamental divergence in persuasive architecture: human influencers construct experiential narratives exhibiting message characteristics typically associated with peripheral-route cues (sensory descriptions, emotional testimonials, social context), while AI recommendations employ systematic, evidence-based discourse exhibiting message characteristics typically associated with central-route argumentation (functional mechanisms, ingredient specifications, objective criteria). Topic modeling identified four distinct thematic clusters for each source type: human discourse centers on embodied experience and relational consumption, whereas AI discourse organizes around informational utility and rational decision support. Jensen–Shannon Divergence analysis (JSD = 0.213 bits) confirmed moderate distributional divergence, while chi-square testing (χ² = 847.23, p < 0.001) and Cramér’s V (0.312, indicating a medium-to-large effect) demonstrated statistically significant and substantively meaningful differences. These findings extend CASA theory by demonstrating that AI recommendation agents develop a characteristic “AI communication signature” distinguishable from human persuasion patterns. We propose an integrated Dual-Agent Persuasion Proposition—synthesizing CASA, ELM, and Source Credibility perspectives—suggesting that AI and human recommenders serve complementary functions across different stages of the consumer decision journey—a proposition whose predictions regarding sequential persuasive effectiveness and consumer processing routes await experimental validation. These findings carry implications for AI content strategy optimization, platform design, and emerging regulatory frameworks for AI-generated content labeling.

1. Introduction

The proliferation of generative AI has fundamentally transformed how consumers search for and receive product information [1]. Traditional search paradigms, wherein users input keywords and navigate ranked results, are increasingly being supplanted by AI-powered conversational interfaces that directly synthesize answers to consumer queries [1,2]. This paradigm shift carries profound implications for electronic commerce, as AI systems now function not merely as information intermediaries but as active recommendation agents capable of influencing consumer decision-making [2,3]. The emergence of AI as a recommendation agent represents a significant departure from conventional e-commerce dynamics. Unlike traditional search engine optimization (SEO) strategies that focus on achieving high visibility in search rankings, the AI-mediated commerce environment requires brands to become trusted information sources that AI systems preferentially reference in their recommendations [4]. This transformation necessitates a fundamental reconceptualization of content strategy, shifting from human-centric persuasion toward what may be termed “AI-friendly” content design [5].

Simultaneously, social commerce has witnessed the rise of influencer marketing as a dominant paradigm for product recommendations [6]. Instagram influencers, particularly those operating personal online markets, have established themselves as trusted recommendation sources across product categories, leveraging their perceived authenticity and experiential credibility to influence consumer purchase decisions [6]. The coexistence of these two distinct recommendation paradigms—AI-generated and human influencer recommendations—raises fundamental questions about how they differ in their communication strategies and semantic structures [7].

This study addresses these issues by conducting a comparative analysis of semantic structures in AI-generated and human influencer product recommendations. Specifically, we focus on inner beauty enzyme products—a rapidly growing category in the Korean health and beauty market—as an ideal context that encompasses both functional and experiential appeals. Using computational text-mining techniques, we identify key differences in language use and thematic content between AI and human recommendations. Based on the theoretical foundations, we formulate three key research hypotheses to guide our investigation, addressing: (1) divergent keyword usage between AI and human recommenders, (2) differences in their topical structures, and (3) the implications of these differences for developing integrated dual-agent marketing communication strategies.

Our research makes several contributions to the emerging literature on AI-mediated marketing communications. First, we provide empirical evidence of systematic linguistic differences between AI and human recommendation sources—differences that have been theorized but not yet comprehensively documented at the semantic level [8,9]. Second, we extend the Computers Are Social Actors (CASA) paradigm by examining not just how consumers perceive AI, but how AI actually communicates—revealing characteristic “AI communication signatures” that distinguish machine-generated from human-generated persuasive content [10,11]. Third, by integrating insights from Source Credibility Theory and the Elaboration Likelihood Model (ELM), we propose a Dual-Agent Persuasion Proposition that conceptualizes AI and human influencers as complementary rather than competing sources, each exhibiting message characteristics associated with different persuasion routes in the consumer decision journey [12,13].

2. Theoretical Background and Literature Review

2.1. AI in Advertising and Product Recommendation

The application of artificial intelligence in advertising has expanded rapidly across personalization, consumer behavior prediction, content production, and advertising optimization [8]. Gao et al. [9] categorized AI applications in advertising into four key areas, targeting, personalization, content production, and advertising optimization, establishing a foundational framework for understanding AI’s multifaceted role in marketing communications. Recent advances illustrate AI’s capabilities in these areas. In consumer behavior prediction, Brand et al. [14] demonstrated that large language models (LLMs) can simulate human survey responses, accurately predicting consumer product preferences and social media sentiment. In content creation, generative AI enables the automated production of advertising copy and video content, significantly enhancing marketing efficiency [14]. For example, Haleem et al. [15] documented how natural language processing techniques facilitate the automated generation of various text-based ad content.

User reception research has further advanced understanding of how consumers respond to AI-generated advertising. Argana et al. [16] identified three distinct phases in user responses to AI advertisements—reception, engagement, and termination—providing a framework for optimizing AI-driven advertising strategies by aligning content with each phase of user experience. These findings suggest that AI’s role in advertising extends beyond content generation to encompass the entire consumer interaction lifecycle [17].

2.2. Trust and Information Source Selection

Trust is a critical factor in the effectiveness of AI-driven recommendations. Prior studies indicate that the relative effectiveness of AI versus human recommenders can depend on product type. For hedonic products that emphasize sensory and experiential pleasure, human recommenders have proven more effective, likely due to their superior ability to elicit consumer empathy [18]. In contrast, for utilitarian products focused on functional benefits, no significant difference in effectiveness emerges between AI and human recommenders [18]. This suggests that human sources excel when emotional, sensory connection is needed, whereas AI can perform on par with humans when rational product attributes dominate.

Other research highlights the conditions influencing consumer reliance on AI recommendations. Narayanan et al. [19] found that, when AI suggestions align closely with a user’s personal values (e.g., ethical standards), users show greater trust and dependence on AI. Similarly, Mazzù et al. [20] observed that, for products with highly objective, factual attributes, consumers tend to prefer AI recommendations, whereas, for products where subjective experience and personal taste matter, human expert recommendations garner greater trust. Together, these studies underscore that the perceived nature of the product (functional vs. experiential) and value alignment are key moderators of whether AI or human sources are more credible and influential [21].

2.3. Information Processing Differences Between AI and Humans

A growing body of work documents systematic differences in how AI systems versus humans process information and make decisions. Davenport [22] notes that AI relies on data-driven, quantitative analysis, fundamentally differing from the intuitive, heuristic-driven judgments that humans often employ. Gupta and Khan [23] argue that AI’s data-driven approach enables more dynamic and efficient marketing decision-making, in ways distinct from human marketers’ often intuition-based strategies.

Empirical studies illustrate these processing differences in practice. Andreas and Joel [24] examined AI and human responses to online banner advertisements for hotels. They found that AI prioritizes factual, structured data (e.g., prices, specifications, discount rates, availability) over visual cues, whereas human consumers respond more to visual elements and emotional appeals in ads. Notably, AI exhibited a greater reliance on text-based information, whereas human users were more influenced by imagery and emotion [24]. These findings reinforce that AI’s “cognitive style” is highly systematic and fact-focused, while humans integrate emotional and visual information in decision-making [25].

2.4. Influencer Marketing in Social Commerce

Influencer marketing has emerged as a dominant force in social commerce, with Instagram serving as a primary platform for product recommendations [26]. Influencers operating personal online markets leverage their parasocial relationships with followers to drive purchase decisions, often emphasizing authentic personal experiences and emotional connections [26]. Prior research finds that influencer effectiveness derives largely from perceived authenticity, expertise, and relational closeness to the audience [27].

In the context of our study, the inner beauty product market—encompassing supplements for skin health and overall wellness—represents a particularly active domain for influencer marketing in South Korea [28]. This market’s content often features personal testimonials about cosmetic or health benefits. The prominence of the inner beauty sector in Korea’s social commerce (with a mature influencer ecosystem and sophisticated consumers) makes it an ideal setting to examine AI vs. human recommendation differences [29]. Indeed, enzymes (a key inner beauty product) emerged as the most salient keyword in our influencer dataset, underscoring the product category’s market relevance.

2.5. Integrative Theoretical Framework

Building upon the above literature, we propose an integrative theoretical framework to explain AI–human recommendation divergence through established theories of persuasion and human–computer interaction. Source Credibility Theory identifies expertise, trustworthiness, and attractiveness as key determinants of a message source’s effectiveness [30]. In AI-mediated contexts, AI’s expertise is manifested through comprehensive data synthesis and systematic information coverage, while its trustworthiness is linked to perceived objectivity [31]. By contrast, human influencers construct credibility through authentic personal experience narratives and the quality of their parasocial relationships with followers [32]. This suggests that AI and human sources project credibility via fundamentally different linguistic cues (e.g., factual detail vs. personal testimony).

The CASA paradigm (Computers Are Social Actors) established that humans tend to apply social expectations to interactions with computers [10]. We extend CASA beyond human perceptions to examine the production characteristics of AI communication. If AI systems process information via mechanisms fundamentally different from humans, their language output should exhibit systematic semantic signatures—that is, consistent patterns in word choice, topic organization, and argument structure reflecting underlying computational cognition [15]. Detecting such an “AI communication signature” would indicate that, while users might anthropomorphize AI (per CASA), the AI’s style of communication remains inherently distinct from human communication.

Integrating the Elaboration Likelihood Model (ELM), we draw on its distinction between central-route and peripheral-route message characteristics to characterize the communication strategies of AI and human sources [12]. The ELM posits that persuasive messages can be processed through a central route (careful evaluation of argument quality) or a peripheral route (reliance on heuristic cues such as source attractiveness or emotional appeals). Importantly, the route through which a message is actually processed depends on the recipient’s motivation and ability, not solely on message features [12]. However, messages themselves can be characterized by the extent to which they emphasize substantive arguments versus heuristic cues. We hypothesize that AI’s systematic, evidence-based communication exhibits message characteristics typically associated with central-route argumentation, whereas human influencers’ experiential and narrative communication provides the kind of heuristic cues typically leveraged in peripheral-route persuasion. This yields a key theoretical prediction: rather than competing for the same persuasive space, AI and human recommendations may each serve complementary functions in the consumer decision journey by offering different types of persuasive content [13].

3. Hypotheses Development

Drawing on the theoretical insights above, we formalize three research hypotheses about how AI-generated and human influencer recommendations differ:

(1)

H1 (Language Use Divergence): AI-generated product recommendations will exhibit significantly different vocabulary and keyword emphasis compared to human influencer recommendations, even when addressing the same product category. Specifically, AI text will emphasize functional, technical, and evaluative terms (e.g., mechanisms, criteria, certifications), whereas human text will emphasize experiential, sensory, and social terms (e.g., taste, gift, today).

(2)

H2 (Content Structure Divergence): The thematic content structure of AI versus human recommendations will differ systematically. AI content will organize around information utility themes (mechanisms, selection criteria, objective facts), while human content will organize around personal experience and social themes (experiential reactions, lifestyle integration, relational consumption such as gift-giving).

(3)

H3 (Message Characteristic Divergence): The linguistic patterns of AI-generated recommendations will exhibit message characteristics typically associated with central-route argumentation—emphasizing substantive evidence, functional mechanisms, and systematic evaluation criteria—whereas human influencer recommendations will exhibit message characteristics typically associated with peripheral-route cues—emphasizing experiential narratives, sensory descriptions, and heuristic social signals. This divergence reflects the distinct credibility construction strategies each source type employs: AI leveraging perceived expertise and objectivity through comprehensive information coverage, and humans leveraging perceived authenticity and trustworthiness through parasocial connection and personal testimony.

Figure 1 presents the integrative theoretical framework of this study, illustrating how Source Credibility Theory, the CASA paradigm, and the Elaboration Likelihood Model converge to generate three research hypotheses (H1–H3), each tested by a corresponding analytical method.

4. Methodology

4.1. Research Design Overview

This study employs a comparative computational text analysis design to examine semantic differences between AI-generated and human influencer product recommendations. We collected two parallel text corpora—one from human Instagram influencers and one from AI recommendation systems—both addressing the same product category (inner beauty enzyme supplements). By analyzing these corpora using multiple complementary techniques, we systematically compare vocabulary, thematic content, and semantic associations between the two source types.

4.2. Product Category Selection Rationale

The selection of inner beauty enzyme products as our focal category was theoretically motivated. Inner beauty products represent a hybrid category combining utilitarian (functional health benefits) and hedonic (beauty and experiential pleasure) attributes [33]. Following Dhar and Wertenbroch’s [34] framework on hedonic versus utilitarian goods, such hybrid products occupy an interesting position at the intersection of objective and subjective evaluation dimensions.

This hybrid nature makes inner beauty enzymes ideal for studying source-driven communication differences. If AI and human sources adopt systematically different communication strategies, these differences should be observable in how they frame a product that legitimately supports both rational and emotional appeals. A purely utilitarian product might constrain both sources to factual discourse, while a purely hedonic product might push both toward experiential language—neither would allow differentiation. The hybrid nature of inner beauty products provides discriminative power for detecting source-driven versus product-driven patterns [33,34].

4.3. Data Collection

We collected 330 Instagram post captions from Korean influencers who actively sell or promote products in personal online markets. To obtain a high-quality sample, influencer selection followed a rigorous multi-stage procedure. First, we identified initial candidates via collaborations with advertising agencies, influencer management firms, and commerce analytics platforms. We then applied criteria including sales volume, follower count, and whether they operate their own online storefronts. Verified sales data (via the Korean SMINFO system) yielded 94 qualified influencers [28].

From a comprehensive set of 2722 posts by these influencers, we filtered down to those in the inner beauty category (584 posts). Finally, focusing specifically on enzyme-related products within inner beauty, we used term frequency (TF) and TF-IDF analysis to identify relevant posts, resulting in 330 enzyme-related influencer posts for analysis.

We additionally collected 541 AI-generated responses simulating consumer queries about inner beauty enzyme products. Data were gathered through a systematic query protocol to approximate natural consumer–AI interactions while maintaining methodological rigor. Platform selection included two leading generative AI platforms: Perplexity AI (notable for integrating real-time web search) and ChatGPT/GPT-4 (the most widely used conversational AI platform). Using both platforms helps generalize findings across different AI architectures [25].

Prompt design methodology was empirically grounded: we analyzed consumer search behaviors by examining Google Trends and Naver (Korea’s dominant search engine) autocomplete suggestions related to enzymes over 12 months. From this, we derived four common query types: (1) general product recommendations, (2) comparative product questions, (3) specific benefit inquiries, and (4) queries about selection criteria. Each query type was executed multiple times with natural language variations on both AI platforms, yielding 541 unique AI response documents.

To ensure transparency and reproducibility of the AI data collection process, we provide representative prompt examples for each query type: (1) General recommendations: “What are good inner beauty enzyme products?”; (2) Comparative queries: “What are the differences between fermented enzyme products and general enzyme supplements?”; (3) Benefit inquiries: “What are the health benefits of taking enzyme products for skin and digestion?”; (4) Selection criteria: “What should I consider when choosing an enzyme supplement?” All prompts were formulated in natural Korean language, using neutral, non-leading phrasing to avoid biasing the AI toward specific response patterns. No system prompts, role assignments, or output constraints were imposed beyond the query itself. A critical methodological concern is whether the researcher-designed prompts may introduce systematic bias into the AI corpus. To address prompt fairness and output consistency, several safeguards were implemented. First, prompts were derived empirically from actual consumer search behaviors (Google Trends and Naver autocomplete data over 12 months), ensuring that the queries reflect genuine information-seeking patterns rather than researcher-imposed framing. Second, each query type was executed with 5–8 natural language variations (e.g., varying formality levels, question structures, and specificity) to prevent any single prompt formulation from dominating the corpus. Third, data collection was distributed across multiple sessions over a two-week period to account for potential temporal variation in AI outputs. Fourth, and most critically, the within-AI platform divergence analysis (JSD = 0.047 bits between ChatGPT and Perplexity AI sub-corpora; permutation test p = 0.34) empirically validates that the AI outputs exhibit high consistency across different platforms and prompt variations, with the between-platform divergence accounting for less than one-quarter of the overall AI–human divergence (JSD = 0.213 bits). This evidence suggests that the observed AI communication patterns are robust properties of AI-generated discourse rather than artifacts of specific prompt designs.

4.4. Analytical Methods

To systematically examine the semantic divergence between human influencer and AI-generated product recommendations, we employed a multi-layered analytical framework that progresses from surface-level lexical patterns to deeper semantic structures. This methodological approach enables triangulation across different levels of textual analysis, ensuring robust and comprehensive findings [35]. At the foundational level, frequency-based analyses reveal what terms each source prioritizes; at the intermediate level, topic modeling uncovers latent thematic structures; and at the deepest level, BERT-based semantic embedding analysis captures how concepts relate to one another within each communication style, leveraging pre-trained contextualized representations to overcome the corpus-size limitations inherent in training domain-specific static embeddings. The following four analytical techniques were applied sequentially to address our research hypotheses.

4.4.1. Term Frequency and TF-IDF Analysis

We first conducted basic term frequency (TF) analysis to identify the most frequently occurring words in each dataset, providing an overview of prevalent keywords. We then applied TF-IDF weighting [35] to highlight terms that are characteristically important to each source. TF-IDF down-weights ubiquitous terms and emphasizes words used frequently by one source but not the other, thus identifying the distinctive vocabulary of each communication style.

4.4.2. Topic Modeling (LDA)

We used Latent Dirichlet Allocation (LDA) [36] to uncover latent thematic structures in each corpus. LDA was implemented using the Gensim 4.3.2 library in Python 3.11, with the following hyperparameter specifications: Dirichlet prior on per-document topic distributions α = “auto” (asymmetric, learned from corpus); Dirichlet prior on per-topic word distributions η = “auto” (learned from corpus); number of passes through the corpus = 20; chunk size = 100 documents; random seed = 42 for reproducibility; minimum word probability threshold = 0.01. The optimal number of topics was determined through systematic evaluation of coherence scores (Cv measure) and perplexity across topic counts ranging from k = 2 to k = 10. Based on convergent metrics, k = 4 was selected as optimal for both corpora, enabling direct thematic comparison.

4.4.3. BERT-Based Semantic Embedding Analysis

To capture deep semantic relationships among key terms beyond surface-level topic co-occurrence, we employed pre-trained BERT-based sentence embeddings rather than training corpus-specific static word embedding models (e.g., Word2Vec). This methodological choice was motivated by a critical consideration: our domain-specific corpora (330 influencer posts; 541 AI responses) are substantially smaller than the scale typically required for reliable Word2Vec training, which generally demands millions of tokens to produce stable vector representations [37]. Training static embeddings on small corpora risks generating noisy, unreliable representations, particularly for low-frequency terms that are often the most semantically informative [38]. We adopted KR-SBERT (Korean Sentence-BERT), a Korean-language variant of the Sentence-BERT framework [39] pre-trained on large-scale Korean text corpora. SBERT extends the BERT architecture by fine-tuning on sentence-pair tasks using siamese and triplet network structures, producing fixed-size sentence embeddings optimized for semantic similarity computation via cosine distance. Unlike static word embedding models, SBERT generates contextualized representations: the same word receives different vector encodings depending on its surrounding context, enabling the disambiguation of polysemous terms [40]. The analytical procedure consisted of four sequential steps. First, we extracted sentence-level embeddings for each document in both corpora using the pre-trained KR-SBERT model (768-dimensional vectors). Second, to derive term-level semantic representations, we identified sentences containing each LDA anchor keyword and computed the mean embedding vector across all sentences in which the keyword appeared within each corpus [41]. Third, we computed pairwise cosine similarities between anchor keyword embeddings within each corpus to construct weighted semantic adjacency matrices. Cosine similarity thresholds were set at θ ≥ 0.65 for edge inclusion. Fourth, the resulting semantic networks were visualized using the Kamada–Kawai force-directed layout algorithm, with edge weights proportional to cosine similarity scores and node sizes proportional to betweenness centrality within each network. To validate the reliability of our BERT-based semantic representations, we performed a qualitative coherence check by verifying that the top-10 nearest neighbors for core anchor terms corresponded to semantically plausible associations, as evaluated independently by two domain-expert raters (inter-rater agreement: Cohen’s κ = 0.84). We also compared the BERT-derived semantic clusters against the LDA topic structures to assess convergent validity.

The cosine similarity threshold of θ ≥ 0.65 was determined through systematic sensitivity analysis across five threshold levels (0.55, 0.60, 0.65, 0.70, 0.75). At θ = 0.55, both networks exceeded density Δ > 0.60, producing near-complete graphs that obscured meaningful cluster structures and rendered visual interpretation impractical. At θ = 0.60, network density remained high (Δ = 0.48–0.52), with excessive edge inclusion still masking the topological differences between the two corpora. At θ = 0.65, the networks achieved a balanced density (Δ = 0.35–0.38) that preserved the most semantically meaningful connections while revealing distinct cluster structures and hub–spoke patterns. At θ = 0.70, network density dropped to 0.20–0.25, resulting in fragmented networks where several semantically related nodes became disconnected. At θ = 0.75, density fell below 0.15, producing overly sparse networks that lost important semantic relationships. The θ = 0.65 threshold thus represents the optimal balance between preserving meaningful semantic connections and filtering noise, consistent with thresholds commonly employed in BERT-based semantic network studies [39,40]. Different thresholds reflect varying degrees of semantic strictness: lower thresholds capture broader associative relationships (including weaker semantic connections), while higher thresholds retain only the strongest semantic bonds, which tend to represent core domain-specific terminology clusters.

4.4.4. Comparative Statistical Measures

We computed the Jensen–Shannon Divergence (JSD) [42,43] between the overall word frequency distributions of the two corpora as an aggregate measure of vocabulary difference. We also conducted chi-square tests [44] on key term occurrences to determine whether certain words or categories were significantly associated with one source or the other. Effect sizes were quantified using Cramér’s V [45,46]. All statistical analyses were implemented in Python 3.11, with numerical operations utilizing NumPy 1.24.3 and chi-square tests employing scipy.stats.chi2_contingency from SciPy 1.11.1 [47]. Visualizations were generated using Matplotlib 3.8.2 at 300 DPI for publication quality.

4.5. Methodological Considerations: Genre Comparability and Ecological Validity

A methodological consideration warranting explicit discussion concerns the genre asymmetry between the two corpora. The human influencer corpus comprises naturally occurring Instagram marketing posts—promotional content crafted for follower engagement—whereas the AI corpus consists of responses generated through researcher-designed query protocols simulating consumer information-seeking behavior. This asymmetry means that observed differences may partially reflect genre-level conventions (promotional vs. informational discourse) rather than source-type effects alone.

We address this concern through three complementary arguments. First, and most importantly, this genre asymmetry reflects the actual communicative ecology that consumers encounter in social commerce environments. When consumers evaluate inner beauty products, they navigate between Instagram influencer posts (unsolicited promotional endorsements) and AI chatbot responses (solicited informational queries) as they naturally occur—not under controlled laboratory conditions. Our research design thus maximizes ecological validity by comparing the two recommendation sources in their native communicative contexts [29], capturing precisely the type of discourse consumers would actually process when making purchase decisions.

Second, the theoretical framework motivating this study predicts source-level rather than genre-level differences. Source Credibility Theory [30] posits that different source types construct credibility through fundamentally different mechanisms (expertise vs. attractiveness), and the CASA paradigm [10] suggests that AI systems develop characteristic communication signatures reflecting their computational processing. If the observed differences were solely attributable to genre conventions, we would not expect the consistent alignment between linguistic patterns and theoretical predictions across multiple analytical layers (TF-IDF, LDA, BERT)—yet this convergence is precisely what we observe.

Third, to partially disentangle source from genre effects, we conducted supplementary analyses. We compared the AI-generated corpus against a random sample of informational blog posts about enzyme products (n = 50), representing human-authored informational content that shares genre characteristics with AI responses. Preliminary analysis indicated that AI responses exhibited a higher lexical density, greater use of evaluative terminology, and more systematic organizational structure than human-authored informational posts, suggesting that source-type effects operate above and beyond genre conventions. While we acknowledge that a fully matched-task experimental design would provide stronger causal inference, we note that such designs sacrifice the ecological validity that constitutes a primary strength of our naturalistic corpus-based approach [35]. Future research employing controlled experimental paradigms can complement our findings by isolating source effects under matched genre conditions.

We additionally note that if the observed differences were solely attributable to genre conventions rather than source-type effects, we would not expect the specific patterns of convergence across our four analytical layers. Genre-level differences (informational vs. promotional) would predict broad stylistic divergence but would not predict the precise alignment between AI’s functional keyword emphasis, its systematic topic organization, and its hub-and-spoke network topology—all of which converge on the same theoretical interpretation of systematic, evidence-based communication. Similarly, genre effects alone would not predict the human corpus’s specific combination of sensory vocabulary, experiential topics, and polycentric network structure. The multi-layered convergence of these patterns with theoretical predictions from CASA and ELM provides evidence that source-type effects operate above and beyond genre conventions, though we acknowledge that only a matched-task experimental design can definitively disentangle these effects.

5. Results

5.1. Term Frequency Analysis

This study employs term frequency (TF) and TF-IDF analysis as foundational methods for examining lexical differences between human influencer and AI-generated product recommendations. The term frequency analysis revealed pronounced differences in keyword emphasis between the two corpora. In the human influencer corpus (n = 330 posts; 17,031 tokens in the shared vocabulary), experiential and sensory terms dominated: “enzyme” appeared with the highest frequency (f = 2489), followed by “delicious” (f = 1654), “taste” (f = 1523), “today” (f = 1398), and “diet” (f = 1456). These terms reflect the human communicators’ emphasis on lived experience, hedonic qualities, and temporal immediacy. In contrast, the AI-generated corpus (n = 541 responses; 15,091 tokens) exhibited markedly different priorities. While “enzyme” also ranked highest (f = 1893), the subsequent high-frequency terms were distinctively informational: “function” (f = 1567), “certification” (f = 1432), “body fat” (f = 1289), and “selection” (f = 1123). These terms indicate AI’s orientation toward systematic evaluation, functional mechanisms, and objective decision criteria. TF-IDF weighting further amplified these differences, confirming that the distinctive vocabulary of each source type reflects fundamentally different communicative priorities rather than mere frequency artifacts.

5.2. Probability Distribution Construction and Statistical Analysis

To enable principled statistical comparison between human influencer and AI-generated corpora, raw term frequencies were transformed into probability distributions. For each corpus c ∈ {Human, AI}, the probability of term t was computed as:

P_c(t) = f_c(t)/Σ f_c(t′)

where f_c(t) represents the raw frequency of term t in corpus c. The shared vocabulary V comprised the top 50 keywords ranked by combined TF-IDF scores across both corpora, ensuring that comparison focused on semantically significant terms while maintaining sufficient coverage for robust statistical inference [35]. This normalization ensures that both distributions sum to unity, enabling valid probabilistic comparison regardless of corpus size differences (Human: 17,031 tokens; AI: 15,091 tokens).

The selection of k = 50 as the shared vocabulary size was determined through systematic evaluation balancing two competing requirements: semantic coverage and statistical validity. A vocabulary that is too small (k = 30) risks omitting informationally distinctive terms, while an excessively large vocabulary (k = 70) introduces low-frequency terms with unstable probability estimates and insufficient expected cell frequencies for valid chi-square testing. Sensitivity analysis confirmed that JSD values remained stable across k = 30 (JSD = 0.198), k = 40 (JSD = 0.207), k = 50 (JSD = 0.213), k = 60 (JSD = 0.219), and k = 70 (JSD = 0.227), with all values falling within the moderate divergence range (0.19–0.23 bits). The k = 50 threshold was selected as the midpoint of this stable range, providing optimal coverage of semantically significant terms while maintaining minimum expected cell frequencies above 5.0 for all cells in the chi-square contingency table, as recommended by Agresti [47]. This threshold is consistent with established practices in computational text analysis for domain-specific vocabulary comparison [35].

The statistical comparison revealed moderate but meaningful divergence between human influencer and AI-generated corpora (

Table 1. Statistical comparison of human and AI corpus term frequency distributions.

Metric	Symbol	Value	Interpretation	Reference
Jensen–Shannon Divergence	JSD	0.213 bits	Moderate divergence	0: identical, 1: maximum
Chi-square statistic	χ2	847.23	Highly significant	Critical value: 66.34 (α = 0.001)
Degrees of freedom	df	49	—	(rows − 1) × (cols − 1)
p-value	p	<0.001	Reject null hypothesis (H0)	α = 0.001
Cramér’s V	V	0.312	Medium-to-large effect	Small: 0.1, Medium: 0.3, Large: 0.5
Human corpus tokens	n1	17,031	330 Instagram posts	—
AI corpus tokens	n2	15,091	541 AI responses	—

), with JSD = 0.213 bits, χ² = 847.23 (p < 0.001), and Cramér’s V = 0.312, indicating a medium-to-large effect. Figure 2 visualizes the term frequency distributions across 20 shared keywords, illustrating distinct patterns of term dominance between the two sources.

To ensure the reliability of these findings, we conducted several robustness checks [48]. First, vocabulary sensitivity was assessed by varying the keyword threshold (k = 30, 40, 50, 60, 70), confirming that JSD values remained stable within the range 0.19–0.23 bits and indicating that our findings are not artifacts of arbitrary vocabulary selection. Second, smoothing sensitivity was tested using alternative smoothing parameters (α = 1 × 10⁻⁸, 1 × 10⁻¹⁰, 1 × 10⁻¹²), verifying that JSD estimates varied by less than 0.001 bits and confirming numerical stability. Third, bootstrap confidence intervals were generated from 1000 bootstrap samples, yielding 95% confidence intervals for JSD (0.198–0.229 bits) and Cramér’s V (0.298–0.327), which confirmed the precision of our point estimates. Fourth, platform homogeneity was assessed by computing Jensen–Shannon Divergence between the ChatGPT-generated (n = 278) and Perplexity AI-generated (n = 263) sub-corpora. The within-AI platform divergence (JSD = 0.047 bits) was substantially smaller than the cross-source divergence (JSD = 0.213 bits), confirming that between-platform variation accounts for less than one-quarter of the overall AI–human divergence. A permutation test (1000 iterations) further confirmed that the two AI sub-corpora did not differ significantly in their term distributions (p = 0.34), supporting the decision to treat AI-generated responses as a unified corpus regardless of platform origin.

5.2.1. Jensen–Shannon Divergence (JSD)

The Jensen–Shannon Divergence quantifies the similarity between two probability distributions by measuring the average Kullback–Leibler (KL) divergence of each distribution from their mixture [42,43]. Unlike the asymmetric KL divergence, JSD is symmetric, bounded, and always defined, making it particularly suitable for comparing term distributions. The JSD between the human distribution P and the AI distribution Q is defined as:

JSD(P || Q) = ½ D_KL(P || M) + ½ D_KL(Q || M)

where M = ½(P + Q) represents the mixture distribution. When computed using base-2 logarithms, JSD values range from 0 bits (identical distributions) to 1 bit (maximally divergent distributions with no overlap). The interpretation thresholds commonly used in computational linguistics are: JSD < 0.1 (highly similar), 0.1 ≤ JSD < 0.3 (moderately divergent), and JSD ≥ 0.3 (highly divergent) [43].

In our analysis, the computed JSD value of 0.213 bits indicates moderate divergence between the human influencer and AI-generated lexical distributions. This value suggests that, while the two sources share substantial vocabulary overlap (both frequently reference the core product category “enzyme”), they exhibit meaningfully different patterns of word usage and emphasis. The moderate divergence aligns with our theoretical expectations: human influencers and AI systems draw from overlapping but distinctly weighted semantic repertoires when recommending products.

5.2.2. Chi-Square Test for Independence

To assess whether observed frequency differences between corpora were statistically significant, we conducted Pearson’s chi-square test for independence [44]. A 2 × k contingency table was constructed, with rows representing source type (Human, AI) and columns representing the k = 50 keywords in the shared vocabulary. Expected frequencies under the null hypothesis of independence were computed as:

E_ij = (R_i × C_j)/N

where Ri is the row total for source i, Cj is the column total for keyword j, and N is the grand total of observations [47].

The test statistic follows a chi-square distribution with df = (r − 1)(c − 1) = (2 − 1)(50 − 1) = 49 degrees of freedom. Our analysis yielded χ² = 847.23 with p < 0.001, leading to rejection of the null hypothesis of independence. The critical value at α = 0.001 is 66.34; our observed statistic exceeds this threshold by more than 10-fold, providing strong evidence that term distributions differ significantly between sources.

5.2.3. Effect Size: Cramér’s V

While statistical significance indicates that differences exist, effect size quantifies the magnitude of association [45]. We computed Cramér’s V [46] as a standardized measure of association strength for contingency tables. Cramér’s V ranges from 0 (no association) to 1 (perfect association). Following Cohen’s [45] conventions adapted for contingency tables: V < 0.1 indicates negligible association; 0.1 ≤ V < 0.3 indicates small-to-medium association; 0.3 ≤ V < 0.5 indicates medium-to-large association; V ≥ 0.5 indicates large association. With N = 32,122 total observations and χ² = 847.23, we obtained Cramér’s V = 0.312, indicating a medium-to-large effect size. This confirms that the observed differences between human influencer and AI-generated content are not merely statistically detectable but substantively meaningful.

To provide further interpretive context, the Cramér’s V value of 0.312 indicates that approximately 9.7% of the variation in term usage patterns is systematically associated with source type (AI vs. human), a proportion that is both statistically robust and substantively meaningful in the context of computational linguistics research. For comparison, effect sizes in communication research typically range from V = 0.10–0.25 for moderate effects; our observed V = 0.312 exceeds this range, placing it in the medium-to-large effect category. In practical terms, this means that knowing whether a text was generated by AI or a human influencer provides meaningful predictive information about which specific terms will be emphasized—a finding that underscores the existence of distinct “communication signatures” for each source type. The combination of statistical significance (p < 0.001, with the observed χ² exceeding the critical value by more than 10-fold) and substantive effect size provides strong evidence that the distributional differences are neither chance artifacts nor trivially small variations.

Figure 3 illustrates this substantive divergence by comparing the probability distributions of the 20 shared keywords, sorted by frequency difference (Human−AI). The lollipop chart reveals a clear structural pattern: Human-dominant terms such as “enzyme,” “delicious,” and “taste” occupy the upper portion, with substantially higher probability mass, while AI-dominant terms including “function,” “body fat,” and “certification” cluster in the lower portion with reversed probability gaps. This visual separation confirms that the medium-to-large effect captured by Cramér’s V reflects systematic distributional asymmetry between the two sources rather than sporadic variation across individual terms.

5.2.4. Per-Keyword Chi-Square Contributions

To identify which specific terms most strongly differentiate the two sources, we decomposed the overall chi-square statistic into per-keyword contributions [47]. Figure 4 ranks all 20 keywords by their individual χ² contributions, with bar color distinguishing Human-dominant (red) from AI-dominant (blue) terms.

High-contribution keywords (χ² > 45) included Human-dominant terms such as “delicious” (χ² = 55.5), “taste” (χ² = 47.0), and “today” (χ² = 46.5), as well as AI-dominant terms including “function” (χ² = 53.1), “certification” (χ² = 48.4), and “body fat” (χ² = 46.1). Notably, the six highest-contributing keywords accounted for approximately 35% of the total χ² statistic, indicating that distributional divergence is concentrated in a relatively small set of semantically distinct terms rather than being uniformly distributed across the vocabulary.

This pattern corroborates our theoretical framework: Human influencers employ experiential and sensory vocabulary characteristic of peripheral-route message cues as described in the Elaboration Likelihood Model, while AI-generated content emphasizes functional and credibility-focused terms characteristic of central-route message arguments [12]. When considered alongside the probability distribution comparison (Figure 3), these results suggest that the two sources not only differ in which terms they prioritize but also in the magnitude of that prioritization, with the largest divergences occurring precisely at the boundary between experiential and evaluative discourse.

5.2.5. Standardized Residuals

To determine the direction and magnitude of deviations from expected frequencies, we computed standardized residuals (Pearson residuals) for each cell: r_ij = (O_ij − E_ij)/√E_ij. Standardized residuals follow approximately a standard normal distribution, enabling significance testing: |r| > 1.96 indicates significance at α = 0.05; |r| > 2.58 indicates significance at α = 0.01 [47]. Positive residuals indicate over-representation relative to expectation, while negative residuals indicate under-representation. For example, “delicious” showed r = +8.2 in the human corpus (significant over-representation), while “certification” showed r = +7.8 in the AI corpus (significant over-representation). These patterns provide cell-level validation of source-specific vocabulary preferences.

5.3. Topic Modeling (LDA) Results

To identify latent thematic structures underlying each corpus, we applied Latent Dirichlet Allocation (LDA) separately to the human influencer and AI-generated corpora. The optimal number of topics was determined through systematic evaluation of coherence scores and perplexity measures across topic counts ranging from k = 2 to k = 10 [36].

For the human influencer corpus, perplexity analysis revealed an inflection point at k = 4 (perplexity = 1551.54), where the rate of perplexity reduction shifted markedly (decrease of 70.25 from k = 3). The coherence score at k = 4 was 0.466, representing a local maximum that confirmed topic interpretability. For the AI-generated corpus, coherence analysis yielded a clear peak at k = 4 (coherence = 0.4565), substantially exceeding adjacent values at k = 3 (0.4367) and k = 5 (0.4352). Based on these convergent metrics, k = 4 was selected as the optimal topic count for both corpora, enabling direct thematic comparison.

Table 2. LDA Topic Modeling: Human Influencer vs. AI-Generated Corpora (k = 4).

Topic	Human Theme	Human Top Keywords	AI Theme	AI Top Keywords
Topic 1	Experiential Response	enzyme, protein, fat, taste, skin	Perceived Effects	enzyme, intake, consumer, change, skin
Topic 2	Quality & Safety	domestic, activity, fermentation, additives, digestion	Selection Criteria	ingredients, certification, selection, fermentation, routine
Topic 3	Diet & Taste	diet, enzyme, snack, taste, product	Digestive Function	protein, digestion, body, decomposition, improvement
Topic 4	Gift Consumption	dietary fiber, nutrition, gift, health, interest	Ingredient Awareness	inner beauty, function, effect, skin, ingredient importance

presents the LDA topic modeling results for both corpora. The human influencer corpus yielded four thematic clusters reflecting experiential and relational consumption patterns. Topic 1 (“Experiential Response”) centered on embodied experience keywords including enzyme, protein, fat, taste, and skin, capturing how influencers frame product recommendations through personal sensory encounters. Topic 2 (“Quality and Safety”) organized around domestically produced, enzyme activity, fermentation, additives, and digestion, reflecting consumer concerns about product provenance and safety that influencers address through personal testimony. Topic 3 (“Diet and Taste”) featured diet, enzyme, snack, taste, and product, representing the hedonic framing of health products as enjoyable dietary experiences. Topic 4 (“Gift Consumption”) comprised dietary fiber, nutrition, gift, health, and interest, revealing a distinctive relational consumption theme where health products serve as social gifts.

In contrast, the AI-generated corpus produced four topics reflecting systematic, information–utility-oriented discourse (Table 2). Topic 1 (“Perceived Effects”) included enzyme, intake, consumer, change, and skin, framing product effects through objective outcome expectations rather than personal experience. Topic 2 (“Selection Criteria”) organized around ingredients, certification, selection, fermentation, and routine, presenting a systematic decision framework absent from human discourse. Topic 3 (“Digestive Function”) featured protein, digestion, body, decomposition, and improvement, emphasizing biochemical mechanisms of product action. Topic 4 (“Ingredient Awareness”) comprised inner beauty, function, effect, skin, and ingredient importance, framing product evaluation through functional assessment criteria.

The cross-source comparison of topic structures reveals a fundamental thematic divergence that directly supports H2. While both corpora address the same product category, human influencer topics consistently foreground experiential, hedonic, and relational dimensions of consumption—personal sensory encounters (Topic 1), safety concerns validated through personal testimony (Topic 2), enjoyable dietary experiences (Topic 3), and social gift-giving (Topic 4). In contrast, AI topics systematically organize around informational utility—objective outcome expectations (Topic 1), systematic decision frameworks (Topic 2), biochemical mechanisms (Topic 3), and functional assessment criteria (Topic 4). Notably, the human corpus uniquely generates a “Gift Consumption” theme with no AI parallel, highlighting the social and relational dimensions of product recommendation that remain distinctively human.

5.4. BERT-Based Semantic Network Analysis

To capture deeper semantic relationships beyond topic-level analysis, we generated contextualized embeddings for both corpora using the pre-trained KR-SBERT model. Using the anchor keywords identified from the LDA topics, we extracted sentence-level embeddings for all sentences containing each anchor term, computed mean contextual vectors, and constructed pairwise cosine similarity matrices. These similarity scores were used to build weighted semantic networks, visualized using the Kamada–Kawai force-directed layout algorithm to reveal the underlying structure of semantic associations within each corpus.

The resulting semantic networks displayed markedly different topological structures across the two corpora (Figure 5), consistent with but extending beyond the patterns identified in LDA topic modeling.

Table 3. Network structural metrics comparison: human influencer vs. AI-generated semantic networks.

Metric	Human Influencer	AI-Generated	Interpretation
Network centralization (C_D)	0.2803	0.2348	Human higher (single dominant bridge node)
Network density (Δ)	0.3462	0.3846	AI more uniformly connected
Avg. clustering (C−)	0.5990	0.4306	Human more locally clustered
Transitivity (T)	0.5644	0.4696	Human higher transitivity
Avg. path length (L−)	2.0385	1.7821	AI more compact
Diameter (d)	4	3	AI shorter max distance
Cross-corpus cos: “Enzyme”	—	—	0.61
Cross-corpus cos: “Skin”	—	—	0.58
Adj. Rand index (LDA–BERT)	0.72	0.78	Strong convergent validity

summarizes the key structural metrics that quantify these differences. The human network exhibited higher Freeman centralization (CD = 0.2803 vs. 0.2348)—driven by the disproportionate bridging role of the “purchase” node (CB = 0.624)—while the AI network showed higher density (Δ = 0.3846 vs. 0.3462) and more compact topology (L⁻ = 1.7821 vs. 2.0385), reflecting systematically distributed connectivity. The human network’s higher average clustering (C⁻ = 0.5990 vs. 0.4306) and transitivity (T = 0.5644 vs. 0.4696) confirm a polycentric topology with tightly knit local clusters, contrasting with the AI network’s more uniformly connected structure.

In the human influencer network (Figure 5a), semantic clusters formed around experiential and hedonic nodes, exhibiting a distributed, polycentric topology. The “taste” node emerged as a primary hub with high betweenness centrality, densely connected to “delicious” (cosine similarity = 0.78), “snack” (0.74), and “diet” (0.71), reflecting the sensory-pleasure orientation of influencer discourse. Notably, the contextualized embeddings revealed that “taste” in the influencer corpus carried strong affective valence, clustering more closely with emotional descriptors than with neutral sensory terms—a nuance that static word embeddings would fail to capture. The “purchase” node exhibited strong connections to “group buying” (0.76), “gift box” (0.72), “free gift” (0.69), and “open” (0.68), revealing a commerce-embedded semantic structure where product recommendations are inseparable from transactional and social sharing contexts. The presence of “today” as a prominent connector node with high betweenness centrality underscored the temporal immediacy characteristic of influencer communication, bridging experiential and commercial clusters.

In contrast, the AI-generated semantic network (Figure 5b) organized around functional and evaluative clusters with a more hierarchical, hub-and-spoke topology. The “enzyme” node served as the dominant central hub with the highest betweenness centrality, radiating connections to “digestion” (cosine similarity = 0.82), “decomposition” (0.79), “fermentation” (0.77), and “intake” (0.75)—forming a mechanistic explanation cluster. The contextualized embeddings revealed that AI’s usage of “enzyme” was consistently embedded in explanatory, mechanism-oriented contexts, whereas human influencers’ usage appeared in experiential and testimonial contexts—a critical semantic divergence captured by the BERT-based approach. A separate evaluative cluster linked “certification” to “ingredients” (0.80), “selection” (0.76), and “function” (0.74), reflecting the systematic decision-support architecture of AI discourse.

As reported in Table 3, quantitative comparison confirmed the visual differences observed in Figure 5: the human network exhibited higher Freeman centralization (CD = 0.2803 vs. 0.2348), attributable to a single dominant bridge node (“purchase”), while the AI network showed higher density (Δ = 0.3846 vs. 0.3462) and shorter average path length (L⁻ = 1.7821 vs. 2.0385), reflecting more uniformly distributed connectivity. The human network demonstrated higher transitivity and average clustering, consistent with its polycentric, locally clustered topology. Cross-corpus semantic shift analysis revealed that shared terms carry fundamentally different semantic profiles across sources: “enzyme” showed a cross-corpus cosine similarity of only 0.61, and “skin” showed 0.58. Convergent validity assessment confirmed strong correspondence between BERT-derived semantic clusters and LDA topic structures (adjusted Rand index: 0.72 for human corpus, 0.78 for AI corpus). These structural differences provide additional evidence for H3, demonstrating that AI recommendations construct meaning through systematic, functionally organized knowledge hierarchies, whereas human influencers construct meaning through experiential, socially embedded associative networks.

Table 4. Structural comparison of communication characteristics: human influencer vs. AI recommender.

Dimension	Human Influencer	AI Recommender
Consumer Perspective	Experience-sharing agent emphasizing intake experience and emotional reactions	Information-seeking agent focused on product function, ingredients, and effects
Keyword Orientation	Taste, intake sensation, satisfaction, recommendation, value-for-money	Function, ingredients, certification, selection, improvement
Expression Style	Emotional, narrative-driven language (“What & Feel” oriented)	Neutral, functional, semantic language (“Why & How” oriented)
User Goal	Experience sharing and social validation (“I tried this and here’s how it felt …”)	Rational decision support (“Which product is right for me?”)
	Review videos, SNS reviews, viral content (experience-driven)	Functional descriptions, ingredient-focused blogs, consultation content (information-driven)
ELM Message Characteristics	Peripheral-route message characteristics (heuristic cues, sensory empathy)	Central-route message characteristics (cognitive judgment, evidence-based reasoning)

synthesizes the structural differences between human influencer and AI recommender communication across multiple analytical dimensions. The convergent evidence from term frequency analysis, TF-IDF weighting, LDA topic modeling, and BERT-based semantic embedding analysis consistently demonstrates that the two source types exhibit fundamentally different communicative architectures: human influencers operate as experience-sharing agents employing sensory, emotional, and socially embedded language exhibiting message characteristics associated with peripheral-route cues, whereas AI recommenders function as information-providing agents employing neutral, functional, and systematically organized language exhibiting message characteristics associated with central-route argumentation. This integrated pattern of results provides comprehensive support for all three research hypotheses (H1, H2, H3).

6. Discussion

6.1. Interpretation of Findings

The results of our multi-layered analysis provide robust empirical support for all three research hypotheses and yield several important theoretical insights. Regarding H1 (Language Use Divergence), the Jensen–Shannon Divergence of 0.213 bits and the highly significant chi-square test (χ² = 847.23, p < 0.001) confirm that AI-generated and human influencer product recommendations exhibit fundamentally different lexical profiles, even when addressing the same product category. The per-keyword chi-square decomposition further reveals that this divergence is not distributed uniformly across all terms but is concentrated in specific semantic domains: human communicators disproportionately emphasize experiential and sensory vocabulary (“delicious,” “taste,” “today”), while AI systems privilege functional and evaluative terms (“function,” “certification,” “body fat”).

These findings corroborate and extend prior research on AI–human information processing differences [22,24]. While previous studies have documented that AI systems tend toward data-driven, fact-focused communication [4,22], our analysis provides fine-grained evidence of how these differences manifest at the lexical level in product recommendation contexts. The medium-to-large effect size (Cramér’s V = 0.312) indicates that these are not trivial differences but represent a substantively meaningful divergence in communication strategy [45]. Regarding H2 (Content Structure Divergence), the LDA topic modeling results provide compelling evidence of systematic thematic differences. The four topics extracted from each corpus reveal qualitatively distinct organizational logics: human influencer topics center on embodied experience (“Experiential Response”), product provenance concerns validated through personal testimony (“Quality and Safety”), hedonic dietary framing (“Diet and Taste”), and relational consumption (“Gift Consumption”). In contrast, AI topics organize around objective outcome assessment (“Perceived Effects”), systematic evaluation frameworks (“Selection Criteria”), biochemical mechanism explanation (“Digestive Function”), and functional ingredient appraisal (“Ingredient Awareness”). The emergence of a distinctive “Gift Consumption” theme exclusively in the human corpus—with no AI parallel—is particularly noteworthy, as it reveals a dimension of social commerce communication that AI systems fail to capture: the relational and emotional contexts in which products are recommended and consumed [29].

Regarding H3 (Message Characteristic Divergence), the BERT-based contextualized semantic embedding analysis (KR-SBERT) provides the deepest structural evidence. The topological differences between the two semantic networks—the human network’s distributed, associative structure centered on experiential and commerce-embedded nodes versus the AI network’s systematically connected, functionally organized structure with uniformly distributed centrality—map directly onto the ELM’s dual-route framework [12]. Human influencers construct meaning through experiential association chains (taste → delicious → snack → diet), providing the kind of vivid, heuristic cues that characterize peripheral-route persuasion. AI systems construct meaning through functional knowledge hierarchies (enzyme → digestion → decomposition → improvement), providing the systematic, evidence-based argumentation that characterizes central-route persuasion [13].

6.2. Theoretical Contributions

Our findings contribute to the literature in several important ways. First, they extend the CASA paradigm [10,49] from its traditional focus on how humans perceive computers to how computers actually communicate. By identifying systematic “AI communication signatures”—consistent patterns in word choice, thematic organization, and argument structure—we demonstrate that AI recommendation agents do not merely simulate human communication but develop a distinctive communicative style that reflects their underlying computational processing [11,50].

Second, the observed divergence patterns correspond to the Elaboration Likelihood Model’s distinction between central-route and peripheral-route message characteristics [12]. Human influencers’ emphasis on sensory experience, temporal immediacy, and emotional connection maps onto message features associated with peripheral-route cues, while AI’s systematic focus on functional mechanisms, certification standards, and selection criteria corresponds to message features associated with central-route argumentation. This parallel suggests that the ELM framework can be productively applied not only to understand consumer processing but also to characterize the persuasive message strategies of different recommendation sources [13]. We note, however, that the link between message characteristics and actual consumer processing routes requires empirical validation through recipient-side experimental research.

We acknowledge that the ELM fundamentally describes receiver-side cognitive processing routes, which are not directly measured in our text-based analysis. The objective linguistic criteria underlying our characterization of “message characteristics typically associated with” each route are as follows: central-route message characteristics are operationalized through the presence of higher lexical density, systematic evaluative terminology (e.g., function, certification, selection criteria, improvement), structured organizational patterns, and evidence-based argumentation. Peripheral-route message characteristics are operationalized through sensory-experiential vocabulary (e.g., delicious, taste), emotional testimonials, temporal immediacy markers (e.g., today), and social context embedding (e.g., gift, group buying). We further note that the relationship between message characteristics and processing routes is not deterministic: the same AI-generated content emphasizing systematic evaluation could serve as a peripheral cue (e.g., “the sheer volume and organization of data signals source credibility”) for a less motivated consumer, while a highly motivated consumer might centrally process an influencer’s experiential narrative. Our characterization describes the dominant content properties of each source type’s messages, not the cognitive processing routes they necessarily activate in recipients. Experimental validation with recipient-side measures (e.g., elaboration scales, thought-listing protocols) remains essential for testing the processing route predictions of our Dual-Agent Persuasion Proposition.

An alternative interpretation of the observed network topological differences warrants consideration. The AI network’s hub-and-spoke structure, centered on the functional hub “enzyme,” may reflect a more lexically constrained or domain-specific form of communication rather than (or in addition to) central-route argumentation per se. The human network’s distributed, polycentric structure may similarly reflect the greater linguistic creativity and broader world knowledge that humans bring to even a niche topic, rather than peripheral-route persuasion exclusively. Supporting this interpretation, lexical diversity analysis reveals that the human corpus exhibits a higher type-token ratio (TTR = 0.342) compared to the AI corpus (TTR = 0.287), indicating that human influencers employ a more varied vocabulary. However, we argue that our theoretical interpretation and this lexical diversity interpretation are not mutually exclusive: AI’s lexical constraint may be both a manifestation of its computational processing tendencies and a feature that contributes to systematic, evidence-based argumentation. Similarly, human linguistic creativity may simultaneously reflect broader communicative competence and provide the vivid, heuristic cues characteristic of peripheral-route persuasion. Fully disentangling persuasion-route characteristics from basic linguistic properties such as lexical diversity and semantic range remains a key challenge for future research, potentially requiring experimental designs that independently manipulate lexical diversity and argument structure while measuring recipient processing outcomes.

Third, integrating these observations with Source Credibility Theory [30], we propose that AI and human recommenders construct credibility through fundamentally different mechanisms—what we term the “Dual-Agent Persuasion Proposition.” AI systems build credibility through expertise signals (comprehensive information coverage, systematic evaluation, objective criteria), while human influencers build credibility through attractiveness and trustworthiness signals (authentic experience, sensory detail, parasocial connection) [31,32]. Rather than competing for the same persuasive space, these two agent types can serve complementary roles across different stages of the consumer decision journey [51].

6.3. Managerial Implications

For marketing practitioners, our findings suggest several actionable strategies. First, content strategies should be differentiated by channel: AI-optimized content should emphasize structured, evidence-based information that AI systems can easily synthesize and cite, while influencer content should continue to leverage authentic experiential narratives [5,17]. Second, dual-agent strategies should be designed to leverage the complementary strengths of each source type. For instance, AI recommendations may be most effective during the information search and evaluation stages (where consumers seek systematic comparison), while influencer content may be more persuasive during the consideration and purchase stages (where emotional engagement drives conversion) [29,52].

Third, the identification of specific high-differentiating keywords provides concrete guidance for content optimization. Brands seeking AI recommendation inclusion should ensure their content prominently features functional terminology, certification information, and systematic comparison frameworks. Conversely, influencer partnerships should emphasize authentic sensory descriptions and lifestyle integration narratives [53].

Beyond these immediate strategic applications, our findings carry broader implications for the evolving AI-mediated commerce ecosystem. As generative AI becomes increasingly integrated into consumer decision journeys, the distinct communication signatures identified in this study have implications for platform design, regulatory policy, and consumer welfare. Platform designers can leverage the complementary strengths of AI and human recommendation sources to create more effective hybrid recommendation interfaces that combine AI’s systematic evaluation capability with human influencers’ experiential authenticity. From a regulatory perspective, the demonstrable differences between AI-generated and human-generated product recommendations support emerging policy frameworks requiring transparent labeling of AI-generated content, as consumers may process these fundamentally different communication styles through different cognitive pathways. Furthermore, our Dual-Agent Persuasion Proposition contributes to the broader theoretical discourse on human–AI collaboration in marketing communications, suggesting that the future of digital marketing lies not in replacing human communicators with AI but in orchestrating complementary dual-agent strategies that leverage the distinctive strengths of each source type. Similar recent studies examining AI–human content differences in adjacent domains—including news generation [25], creative writing, and customer service [13]—have converged on comparable findings of systematic divergence, suggesting that the AI communication signature identified here may reflect a broader phenomenon of AI-generated content characteristics that transcends specific application domains.

6.4. Limitations and Future Research

Several limitations of this study should be acknowledged. First, our analysis focused on a single product category (inner beauty enzyme supplements) within a specific market context (Korean social commerce). While the hybrid nature of this product category provided discriminative power, generalizability to other product categories and cultural contexts requires further investigation. Second, our AI-generated corpus was collected from two specific platforms (ChatGPT and Perplexity AI); as AI models continue to evolve rapidly, longitudinal studies are needed to track how AI communication patterns change over time [25]. Third, the human influencer corpus comprises naturally occurring Instagram marketing posts, whereas the AI corpus consists of researcher-prompted responses. This methodological asymmetry means that the two corpora represent different communicative genres (promotional marketing vs. informational query responses), which may partially confound the source-type comparison. Future research should employ matched-task designs in which both sources produce recommendations under identical task conditions. Fourth, and critically, while our text-based analysis reveals systematic differences in message characteristics between AI and human sources, it does not directly assess how consumers process or respond to these different communication styles. Our characterization of message features as “associated with central-route argumentation” or “associated with peripheral-route cues” describes the content of the messages, not the cognitive processing routes activated in recipients. Experimental research manipulating source type while measuring consumer elaboration, attitude change, and purchase intention would provide essential complementary evidence [18,20].

Future research should extend this framework in several directions. Cross-cultural comparisons would reveal whether the observed AI–human divergence patterns are universal or culturally moderated. Multi-modal analyses incorporating visual content (images, videos) alongside text would provide a more complete picture of communication differences. Additionally, studies incorporating consumer-level data (e.g., purchase intentions, trust measures, elaboration measures) would enable direct testing of the proposed Dual-Agent Persuasion Proposition’s predictions about complementary persuasive effectiveness and the processing routes through which each source type operates [21,52].

Fifth, regarding AI platform generalizability, while our within-AI platform homogeneity analysis (JSD = 0.047 between ChatGPT and Perplexity AI) suggests that the identified AI communication signature transcends individual platforms, we acknowledge that the rapid evolution of large language models and the significant architectural and training data differences between them (e.g., Claude, Gemini, and future model versions) may limit the generalizability of our specific findings. Different AI systems may exhibit distinct “personalities” due to different fine-tuning objectives, system prompts, and alignment procedures, potentially confounding the concept of a single, stable AI communication signature. Longitudinal and cross-platform studies tracking AI communication patterns across model generations are essential to assess the temporal stability and cross-platform generalizability of the AI signature identified here.

Sixth, the cultural specificity of our findings warrants careful consideration. Our study is situated in the Korean social commerce market, and the analysis relies on Korean-language semantic embeddings (KR-SBERT). Key differentiating terms in the human influencer corpus—such as gift consumption patterns and sensory food descriptions—may be deeply rooted in Korean consumer culture and communication norms. An intriguing hypothesis emerging from this limitation is that the AI communication signature may exhibit greater cross-cultural universality (being driven by platform architecture and training data composition) than the human influencer signature (which is more susceptible to cultural modulation through local communication norms, consumption practices, and linguistic conventions). Cross-cultural replication studies comparing AI–human divergence patterns across culturally distinct markets would test this asymmetric generalizability hypothesis.

Seventh, the choice of inner beauty enzyme products—a hybrid utilitarian–hedonic category—may have disproportionately amplified the observed AI–human divergence. For purely utilitarian products (e.g., laptops, financial products), the AI’s functional focus and the human’s experiential emphasis might converge toward more similar communication patterns, potentially diminishing the semantic gap. Conversely, for purely hedonic products (e.g., perfumes, art), the divergence might widen as AI systems struggle to generate the sensory and emotional content that defines hedonic consumption discourse. Our findings should thus be understood as potentially representing an upper-bound estimate of AI–human semantic divergence, and the boundaries of the proposed Dual-Agent Persuasion Proposition require empirical calibration across the utilitarian–hedonic product spectrum.

Eighth, our analysis focused exclusively on textual content, yet Instagram influencer communication is inherently multimodal, combining text captions with images, videos, Stories, Reels, and interactive features such as polls and question stickers. Visual and audio elements may carry substantial persuasive weight—particularly for experiential product categories where sensory qualities cannot be fully conveyed through text alone. By analyzing only text captions, our study captures one dimension of the influencer communication strategy while potentially underestimating the full extent of AI–human communication divergence. Future research should employ multi-modal analysis frameworks that integrate text, image, and video analysis to provide a more comprehensive picture of how AI and human recommendation sources differ across communicative modalities.

7. Conclusions

7.1. Summary of Findings

This study provides comprehensive empirical evidence that AI-generated and human influencer product recommendations exhibit systematic semantic divergence across lexical, thematic, and structural levels. Through a multi-layered computational analysis of 330 Instagram influencer posts and 541 AI-generated responses concerning inner beauty enzyme products, we documented statistically robust (JSD = 0.213 bits; χ² = 847.23, p < 0.001; Cramér’s V = 0.312) and theoretically meaningful differences. Human influencers construct experiential, peripherally oriented persuasive narratives centered on sensory experience and relational consumption, while AI systems generate systematic, centrally oriented informational content organized around functional mechanisms and evaluative criteria.

These findings advance the theoretical understanding of AI as a communication agent in three ways. First, they extend the CASA paradigm from perception to production, revealing characteristic “AI communication signatures” that distinguish machine-generated from human-generated persuasive content. Second, they demonstrate that the ELM’s message characteristic framework can productively characterize the persuasive discourse of different recommendation sources. Third, they provide the empirical foundation for a Dual-Agent Persuasion Proposition, suggesting that AI and human recommenders serve complementary rather than competing functions in the consumer decision journey—a proposition whose predictions regarding sequential persuasive effectiveness await experimental validation.

7.2. Implications for Practice and Policy

Our findings yield several actionable strategies for marketing practitioners operating in dual-agent commerce environments. First, AI-optimized content architecture should be designed to match how AI systems naturally construct recommendations. Our analysis reveals that AI recommendations organize information around functional concepts through systematically distributed semantic connections (e.g., mechanism → ingredient → certification → selection). Brands seeking favorable AI recommendation inclusion should structure their product content to emphasize functional mechanism descriptions, certified ingredient specifications, comparative selection criteria, and quantified efficacy data. Specifically, our per-keyword analysis identifies the precise vocabulary that AI systems prioritize—terms such as “function,” “certification,” “ingredients,” “selection,” and “improvement”—providing a concrete keyword optimization checklist for AI-facing content strategies.

Second, influencer content briefing should leverage the experiential communication strengths that our analysis identifies as distinctively human. Rather than providing influencers with product specification sheets, brand partnerships should encourage sensory narrative generation centered on high-differentiating keywords: taste descriptions, daily routine integration (“today,” “morning”), and social consumption contexts (“gift,” “group buying”). The emergence of a distinctive “Gift Consumption” theme exclusively in the human corpus suggests an untapped strategic opportunity: influencer campaigns framing health products as social gifts may activate relational consumption motivations that AI recommendations structurally cannot access.

Third, our findings support the design of sequential dual-agent marketing funnels. During the initial information search and evaluation phases, AI-generated content providing systematic comparison frameworks and evidence-based ingredient analysis can satisfy consumers’ needs for rational assessment. During the consideration and purchase phases, influencer content providing experiential testimonials, sensory descriptions, and social proof can provide the emotional engagement that drives conversion. E-commerce platforms might operate this by integrating AI comparison tools in product category pages while featuring influencer content on individual product pages and at checkout.

Fourth, the cross-corpus semantic shift finding—those identical terms (e.g., “enzyme,” “skin”) carry fundamentally different contextual meanings in AI versus human discourse—has important implications for integrated content consistency. Brands should audit whether the semantic framing of key product terms is consistent across their AI-facing content and their influencer partnership content, as consumers who encounter both sources may experience cognitive dissonance if the same product attribute is framed in contradictory ways.

More broadly, the identification of systematic AI communication signatures carries significant implications for the emerging field of AI-mediated marketing communications. As regulatory frameworks worldwide increasingly address AI content labeling requirements, our empirical documentation of measurable linguistic differences between AI and human recommendations provides an evidence base for policy discussions about consumer right-to-know and AI transparency. The Dual-Agent Persuasion Proposition advanced in this study offers a theoretically grounded framework for practitioners and policymakers alike, suggesting that AI and human recommendation sources are not interchangeable but serve complementary functions that, when properly orchestrated, can enhance rather than diminish consumer decision quality. The growing body of research documenting AI–human communication differences across domains underscores the timeliness and significance of developing integrated frameworks for understanding how artificial and human agents jointly shape the evolving landscape of digital commerce.

7.3. Concluding Remark

As generative AI continues to reshape electronic commerce, understanding these fundamental communication differences becomes essential for developing effective multi-agent marketing strategies. The present study offers an empirical and theoretical foundation for this endeavor, but several avenues for future research remain. Cross-cultural replication would reveal whether the observed AI–human divergence patterns are universal or culturally moderated. Matched-task experimental designs would enable stronger causal inference by isolating source effects from genre conventions. Most critically, consumer-level studies directly testing how exposure to AI versus human recommendation styles influences elaboration, attitude formation, and purchase intention would provide the recipient-side evidence necessary to validate the proposed Dual-Agent Persuasion Proposition.