Social presence and interaction dynamics in human–AI communication: evidence from large-scale conversational data

1 Introduction

The recent expansion of conversational artificial intelligence (AI) systems based on large-scale language models has intensified a long-standing problem in communication research: how mediated interaction is organized when technological systems are not merely channels for transmission but interactional environments that shape timing, turn-taking, response design, and the attribution of communicative roles. This issue has been examined in recent human–AI communication research and in linguistic-anthropological work on how participants build temporal infrastructures around ChatGPT-mediated activity (Marconi et al., 2026; Klowait and Erofeeva, 2026). From the perspective of intelligent systems, conversational agents are not only tools for information retrieval but interfaces whose generated responses influence user perception, engagement, and communication dynamics. Platforms such as ChatGPT and Microsoft Copilot generate content, maintain contextual coherence, and adapt responses based on conversational history. These capabilities do not require treating AI systems as human equivalents; rather, they require analyzing how users orient to them through linguistic and structural cues that make interlocutor-like participation observable in dialogue (Nass et al., 1994).

The problem is not whether conversational AI should be described as a passive channel or as an autonomous social actor, but how specific interactional properties become measurable in large-scale dialogue data. Current conversational systems introduce pseudo-dialogic behavior that simulates reciprocity, continuity, and contextual adaptation, thereby complicating the analytical distinction between sender, receiver, medium, and interface. As documented in recent studies, this creates scenarios in which users interact not only with tools but also with systems that can be perceived as interlocutor-like within the local organization of the exchange, introducing a second dimension of analysis: whether these interactions exhibit measurable patterns of affective or symbolic language that reflect AI-mediated social bonding (Pentina et al., 2023). This phenomenon cannot be explained solely by performance or efficiency metrics because it concerns how communication is organized, recognized, and sustained in interaction.

Empirical observations in real-world settings reinforce this need for analysis. Users tend to attribute social characteristics to these systems, employing personalized forms of address, polite expressions, and, in some cases, assigning symbolic identity to the conversational agent. These behaviors suggest the emergence of a third structural dimension: the incorporation of AI systems into recurring communicative routines, in which interaction is not episodic but sustained (Greilich et al., 2025). This situation raises the question of whether there are structural patterns in conversations, such as length, thematic continuity, or persistence, that support the notion of algorithmic companionship as an emergent form of interaction in conversational environments. This evolution suggests that human–AI interaction can no longer be interpreted as episodic tool usage but as a form of sustained communicative engagement, requiring analytical models capable of capturing both relational and structural dimensions of interaction.

Despite the relevance of these phenomena, the available empirical evidence remains limited and fragmented. Most studies have focused on evaluating model performance, response quality, or user satisfaction, often analyzing social presence, affective interaction, and conversational continuity as isolated phenomena (Wang et al., 2024). This fragmentation limits the understanding of human–AI communication as an integrated and evolving process, particularly in large-scale conversational environments. In this regard, the recent availability of large-scale conversation datasets, including human–AI interactions (WildChat and LMSYS Chat-1M) and human–human interactions (DailyDialog and MultiWOZ), enables a more comprehensive quantitative and comparative analysis. Unlike qualitative or small-scale experimental approaches, this study leverages more than 50,000 processed conversations to identify structural and linguistic patterns across heterogeneous interaction scenarios.

This study makes four main contributions. First, it proposes a quantitative framework for modeling social presence, affective bonding, and conversational continuity in human–AI interaction. Second, it introduces three operational indices— the Social Presence Index (SPI), the Social Bonding Index (SBI), and the Companion Communication Index (CCI)—grounded in observable linguistic, affective, and structural features. Third, it provides a large-scale comparative analysis across multiple conversational datasets, enabling systematic differentiation between human–AI and human–human interactions. Finally, it demonstrates non-linear interaction patterns, showing that conversational continuity does not increase linearly with social presence but instead peaks at intermediate levels in human–AI exchanges.

The analysis reveals systematic differences between interaction types. Human–human interactions exhibit higher social presence (SPI = 0.31 vs. 0.13), whereas human–AI interactions demonstrate greater structural continuity (CCI = 0.41 vs. 0.34). Additionally, segmentation analysis shows that while affective bonding increases monotonically with social presence, conversational continuity follows a non-linear pattern in AI-mediated interactions, indicating a balance between naturalness and structural coherence. These findings demonstrate that social presence, affective bonding, and conversational continuity are interdependent and context-sensitive dimensions of communication.

Based on these results, this study proposes an integrated framework for analyzing AI-mediated communication as a multidimensional phenomenon in which social, affective, and structural signals can be observed jointly rather than as isolated dimensions. The contribution is not to claim that conversational systems possess social agency in an ontological sense, but to provide computational evidence of how interlocutor recognition, affective orientation, and continuity are distributed across large-scale human–AI and human–human conversational corpora. This approach contributes to the theoretical and empirical understanding of human–AI communication and provides a foundation for designing conversational systems that are not only functionally effective but also socially interpretable.

2 Literature review

The study of human–AI interaction builds on established traditions concerned with mediated presence, social responses to technologies, and the situated organization of communication. Rather than introducing social presence as an isolated or purely recent construct, this study treats it as part of a broader genealogy of mediated interaction in which users respond to cues of availability, contingency, role recognition, and conversational continuity. These traditions are directly relevant to conversational AI because large language model interfaces combine text-based mediation with responsive generation, producing interactional conditions in which users may orient to the system through social and relational cues (Marconi et al., 2026; Ngo, 2025; Cao et al., 2025; Williams, 2025).

Human–computer interaction and science and technology studies have also cautioned against treating technology either as a neutral conduit or as an independent social subject. Recent linguistic-anthropological work on LLM interaction extends this perspective by showing how participants construct temporal infrastructures through which systems such as ChatGPT become more or less integrated into ongoing activity (Klowait and Erofeeva, 2026). The present study aligns with this interactional view by examining whether large-scale conversational corpora contain measurable traces of interlocutor recognition, affective orientation, and structural continuity.

Within this broader context, recent human–AI studies have moved from a predominantly functional perspective toward recognition of socio-communicative complexity. Pentina et al. (2023) characterizes human–AI relationships as emergent relational phenomena structured by agency, reciprocity, and perceived authenticity, while Marconi et al. (2026) proposes a multidimensional framework in which interaction quality depends not only on performance but also on social and affective components such as presence, synchrony, and responsiveness. This shift does not replace earlier social presence, CMC, or situated interaction scholarship; rather, it extends these traditions to a class of conversational interfaces whose linguistic outputs make interactional and relational signals available for computational analysis.

Within this transformation, human–AI interaction is increasingly interpreted as a communicative process in which users activate social schemas while remaining aware that the system is not a human interlocutor. The Computers Are Social Actors paradigm demonstrated that people can apply social rules to computers without necessarily believing that these systems are human (Nass et al., 1994). This insight is important for conversational AI because social responses may emerge from interactional cues, response timing, linguistic style, and sequential organization rather than from explicit anthropomorphic belief. In this setting, social presence emerges as a central construct mediating interaction. Ngo (2025) identifies social presence as a cognitive mechanism that reduces uncertainty and facilitates trust formation, while Cao et al. (2025) demonstrates that conversational contingency enhances perceived connection and engagement. These dynamics are further shaped by linguistic and contextual features, including language choice, interaction style, and contextual adaptation, as shown by Saile and Hüttl-Maack (Saile and Hüttl-Maack, 2026). Additionally, conversational continuity reinforces the perception of agency, as sustained interaction patterns promote the interpretation of AI systems as socially responsive interlocutors (Elbadawy and Kwon, 2025).

Anthropomorphism has traditionally been considered a key factor in this process; however, the literature suggests that its role is more nuanced than simple human-likeness. Rather than acting as a primary driver, anthropomorphism can operate as a facilitating mechanism that depends on coherence, recipient design, and the quality of interaction. Fakhimi et al. (2023) shows that communicative signals such as linguistic style and responsiveness influence perceived closeness, whereas Greilich et al. (2025) highlights the potential negative effects of excessive anthropomorphism, including user discomfort and privacy concerns. Consequently, recent research has shifted toward linguistic humanization and relational cueing rather than anthropomorphic appearance alone. Kang (2025) introduces the notion of “synthetic relationality,” emphasizing how tone, humor, and imperfection contribute to relational perception. At the same time, Williams (2025) demonstrates that social presence exerts a stronger influence on user experience than anthropomorphic cues alone.

Beyond social presence, the literature increasingly recognizes affective bonding as a distinct yet related dimension of interaction. Diaz et al. (2026) introduces the “Mirror Effect,” showing that perceived empathy arises from structural linguistic alignment rather than genuine understanding, suggesting that relational perception can be induced through interaction patterns. Similarly, Jiang et al. (2025) and Mahmud and Mehta (2025) demonstrate that artificial emotional intelligence enhances satisfaction and trust, particularly in contexts requiring relational sensitivity, although its effectiveness depends on alignment with user expectations. Interaction design elements, such as response timing and simulation of conversational behaviors, further contribute to emotional connection (Bhat, 2025), indicating that affective bonding emerges from a combination of linguistic, temporal, and structural cues. Related work on continuance intention and engagement further shows that perceived interactivity, empathy, and exposure influence sustained use and trust in conversational agents (Zhang et al., 2025; Mrabet et al., 2025).

3 Materials and methods

3.1 Research design and analytical framework

This work adopts a computational, analytical design to identify and quantify interaction patterns in human-AI conversational environments using large volumes of dialogue data. This approach is based on recent evidence characterizing human-AI interaction as a sociorelational process, in which variables such as perceived agency, communicative reciprocity, and conversational continuity play a structural role in shaping user behavior (Marconi et al., 2026; Pentina et al., 2023).

The proposed analytical framework is structured around the operationalization of three interdependent constructs: AI Social Presence in Communication, AI-Mediated Social Bonding, and Algorithmic Companion Communication. These constructs are modeled as observable phenomena based on linguistic, affective, and structural patterns extracted from conversational data. Rather than treating them as abstract perceptual dimensions, these constructs are translated into measurable variables derived from language use, turn-taking dynamics, and the semantic continuity of the interaction.

To this end, a corpus-based methodology is employed to analyze interactions at different levels of granularity: the message, turn, and conversation levels. At the linguistic level, lexical and semantic markers associated with the attribution of agency and social presence are identified. At the affective level, patterns of emotional expression and relational language linked to bonding processes are detected. At the structural level, interaction persistence is evaluated using metrics such as dialogue length, conversational depth, and thematic continuity.

The analytical process is implemented through a combination of natural language processing and interaction modeling techniques. Linguistic feature extraction is performed using defined rules and patterns, while semantic relationships between turns are estimated using similarity measures based on vector representations. This approach allows for the capture of both explicit and implicit indicators of social behavior in human-AI interaction. Furthermore, the framework incorporates a comparative dimension by including human-human interaction data, establishing a baseline for identifying convergences and divergences in communication patterns.

Figure 1 presents the proposed analytical framework, structured as a pipeline that integrates the construction of the conversational corpus, data preprocessing, the extraction of linguistic, affective, and structural interaction signals, the operationalization of theoretical constructs, and the construction of quantitative metrics. It culminates in a comparative analysis of human-AI and human-human interactions. This representation allows for visualizing the direct correspondence between the methodological components and the defined levels of analysis, ensuring traceability between theoretical constructs and empirical evidence.

3.2 Data sources and corpus construction

The analytical framework integrates multiple conversational datasets encompassing both human-AI and human-human interactions, enabling a comparative evaluation of communicative patterns across different interaction environments. This multi-source design aims to reduce reliance on a single dataset and enhance the robustness of the analysis by incorporating heterogeneous conversational scenarios.

The human-AI interaction corpus is constructed from the WildChat, LMSYS Chat, and ShareGPT datasets, which contain real-world interactions between users and conversational systems based on language models (Zhang et al., 2024). These datasets exhibit variability in the length, complexity, and purpose of the conversations, ranging from brief query-response exchanges to extended, multi-turn dialogues. This diversity allows for the analysis of both immediate interaction patterns and sustained conversational dynamics over time.

To establish a comparative baseline, human-human interaction datasets, specifically DailyDialog and MultiWOZ, are incorporated. DailyDialog comprises everyday conversations annotated with communicative intent and emotional content, facilitating the analysis of the social and affective components of interaction. MultiWOZ, on the other hand, contains task-oriented dialogues across multiple domains, allowing for the distinction between functional communicative patterns and those associated with broader social dynamics. The inclusion of these datasets enables the identification of convergences and divergences between AI-mediated communication and traditional human interaction.

From a structural perspective, all datasets are modeled as ordered sequences of conversational exchanges, in which each interaction is broken down into units of analysis at the message, turn, and full-conversation levels. This hierarchical representation allows for the capture of different dimensions of communicative behavior, including the linguistic content of messages, the dependency between consecutive turns, and the depth of interaction.

The corpus construction process involves normalizing and unifying heterogeneous formats across the various datasets. This includes standardizing dialogue structure, eliminating inconsistencies, and filtering out incomplete or invalid interactions. Minimum inclusion criteria are established, such as at least two turns per conversation, ensuring that the analyzed units correspond to effective communicative exchanges. The corpus is structured to preserve information relevant to subsequent analysis, including the temporal sequence of turns, the relationship between consecutive messages, and the semantic integrity of the interaction. This organization allows for the identification of patterns associated with agency attribution, emotional expression, and conversational persistence, which constitute the empirical basis for operationalizing the constructs defined in the analytical model.

3.3 Data preprocessing and representation

The corpus preprocessing was designed to ensure structural consistency across heterogeneous datasets while preserving the linguistic, affective, and conversational properties required for analysis. Each interaction was transformed into a unified representation based on ordered turn sequences. The conversation-level representation is specified in Equation 1. A conversation (C_*i*) is defined as:

Ci={ti,1,ti,2,…,ti,ni}(1)

where (t_{*i, j*}) corresponds to the (j)-th turn and (n_*i*) denotes the total number of turns. This representation preserves sequential dependencies necessary for modeling conversational structure.

Each turn underwent text normalization to remove non-informative noise while preserving semantic content. Let (x) be the original text; the normalization function is specified in Equation 2, where its normalized version (x~) is obtained as:

x~=ϕlc(ϕws(ϕsp(ϕrm(x))))(2)

where (ϕ_*rm*), (ϕ_*sp*), (ϕ_*ws*), and (ϕ_*lc*) perform character filtering, symbol normalization, whitespace normalization, and canonical transformation, respectively.

Normalized turns were embedded into a semantic vector space to capture contextual relationships. In this study, embeddings are generated using Sentence-BERT (SBERT), producing dense representations in a 384-dimensional space that balance representational capacity and computational efficiency. The model is used in its pre-trained form to ensure consistency across datasets.

Semantic continuity between consecutive turns is estimated using cosine similarity. For a conversation (C_*i*), Equation 3 defines the local continuity between adjacent turns as:

si,j=ei,j·ei,j+1||ei,j||||ei,j+1||(3)

where (s_{*i, j*} ∈ [−1, 1]) quantifies semantic alignment. High values indicate thematic continuity, while low values reflect shifts in conversational focus.

The representation preserves three hierarchical levels: message, turn, and conversation. The message level retains textual content, the turn level encodes sequential dependencies, and the conversation level aggregates structural properties such as length and average semantic continuity.

A filtering stage removes incomplete interactions, dialogues with fewer than two turns, and structurally inconsistent records, ensuring compatibility with the analysis of reciprocity and conversational progression. The preprocessing pipeline also incorporates constraints related to sequence segmentation, context preservation, and vector stability, ensuring consistent measurement of semantic similarity. This structured representation enables the definition of analytical variables and the operationalization of the proposed constructs.

3.4 Operationalization of constructs and interaction indices

3.4.1 Operationalization of constructs and variables

The operationalization of the proposed constructs is based on translating theoretical dimensions into observable variables derived from the conversational representation defined in the previous section. Instead of treating social presence, affective bonding, and sustained interaction as abstract categories, the model decomposes them into measurable variables based on three domains of evidence: linguistic patterns, affective expressions, and structural properties of dialogue. Table 1 summarizes the relationship between constructs, variables, indicators, and measurement methods.

Table 1. Operationalization of constructs and variables.

A, Construct (SP, Social Presence; SB, Social Bonding; CC, Companion Communication; H2H, Human-Human); B, Variable; C, Indicators; D, Method.

These variables should be interpreted as computational indicators of interactional orientation rather than as direct psychol1ogical measurements. In this sense, SPI, SBI, and CCI do not claim to infer users' internal states; instead, they operationalize observable traces in dialogue, such as address forms, role attribution, affective expressions, sequential persistence, and semantic continuity. This distinction is important because social presence and bonding have both perceptual and interactional dimensions: the present framework focuses on the latter by measuring how relational and structural cues are distributed across conversational corpora.

Equation 4 formalizes the extracted variable set. Let (C_*i*) be a conversation and (V_*i*) the set of extracted variables:

Vi={vi,1,vi,2,…,vi,K}(4)

Each variable is obtained by applying extraction functions to the conversation or its turns. These functions depend on the nature of the indicator: linguistic variables rely on lexical patterns and contextual semantics; affective variables rely on polarity, gratitude, and relational cues; and structural variables rely on dialogue length, interaction intensity, and semantic continuity between turns.

3.4.2 Operationalization architecture

The operationalization process is structured as a four-level aggregation model: feature space, variable mapping, construct aggregation, and index construction. Figure 2 illustrates how linguistic, affective, and structural features are transformed into analytical variables, grouped into constructs, and synthesized into interaction indices.

Equation 5 defines the feature space used by the operationalization architecture. The architecture begins with a feature space (F_*i*) associated with each conversation:

Fi={Li,Ai,Si}(5)

where (L_*i*), (A_*i*), and (S_*i*) denote linguistic, affective, and structural features, respectively. These features are mapped to observable variables and Equation 6 groups them into three subsets:

Vi(SP), Vi(SB), Vi(CC)(6)

corresponding to AI Social Presence in Communication, AI-Mediated Social Bonding, and Algorithmic Companion Communication.

This architecture preserves the separation between observational domains and theoretical constructs, ensuring interpretability and preventing the aggregation of heterogeneous signals without distinction. The transition from feature space to indices follows a controlled aggregation scheme that retains the analytical contribution of each dimension of conversational behavior.

3.4.3 Formulation of interaction indices

Interaction indices are constructed as normalized composite measures that capture the relative intensity of each communicative construct within a conversation. This formulation enables the transformation of abstract dimensions such as social presence, affective bonding, and conversational continuity into comparable quantitative representations grounded in observable interaction patterns. By integrating linguistic, affective, and structural signals, the indices provide a unified framework for measuring multidimensional interaction behavior.

To ensure comparability across variables with heterogeneous scales and distributions, each variable (v_{*i, k*}) is transformed into a normalized value (v^i,k∈[0,1]). This normalization preserves relative differences while eliminating scale-dependent bias, allowing variables derived from distinct analytical domains to be aggregated within a common metric space. The normalization procedure is presented in Equation 7. As a result, the indices reflect proportional contributions of each variable rather than absolute magnitudes.

v^i,k=vi,k-min(vk)max(vk)-min(vk)(7)

The SPI is defined in terms of variables associated with the recognition of the AI system as an interlocutor-like entity within the exchange, including cognitive attribution, role identification, and interaction-oriented linguistic markers. These variables capture the extent to which the conversation contains observable cues of interlocutor recognition, without assuming that such cues imply human-like agency or stable user belief. The aggregation of these features enables the quantification of communicative presence as a textual and interactional pattern within the dialogue. Equation 8 defines the SPI aggregation.

SPIi=∑k=1mαkv^i,k(SP)(8)

Similarly, the SBI is constructed from variables that reflect affective language, anthropomorphic expressions, and personalized interaction cues, such as gratitude and relational references. These features reflect the extent to which users develop an affective or relational orientation toward the system, indicating the emergence of perceived social bonds. The index, therefore, captures the emotional and relational dimension of interaction beyond functional communication. The SBI aggregation is defined in Equation 9.

SBIi=∑k=1nβkv^i,k(SB)(9)

The CCI is defined using structural interaction variables, including dialogue length, semantic persistence between turns, and interaction intensity. These variables characterize the stability, duration, and recurrence of conversational exchanges, reflecting the extent to which interaction evolves into sustained communicative engagement. Unlike SPI and SBI, which emphasize perception and relational orientation, CCI captures the structural manifestation of the continuity of interaction.

3.4.4 Algorithmic implementation

The operationalization of constructs and the computation of interaction indices are formalized in Algorithm 1, which defines a structured, iterative pipeline applied to each conversation in the corpus. This algorithm ensures that all interactions are processed under a unified computational framework, preserving consistency across feature extraction, variable mapping, normalization, and aggregation. By explicitly defining each transformation step, the procedure guarantees that abstract communicative constructs are systematically translated into measurable representations derived from observable conversational data.

Algorithm 1 Construct operationalization and index computation. Algorithm pseudocode written in black monospaced font describes extracting linguistic, affective, and structural features from a conversational corpus, building a feature space, mapping features to variables, normalizing variables, and computing weighted indices SPI, SBI, and CCI for each conversation.

From a computational perspective, the algorithm decomposes each conversation into three complementary feature domains: linguistic (L_*i*), affective (A_*i*), and structural (S_*i*). This decomposition enables the separation of distinct dimensions of interaction, ensuring that variables associated with social presence, affective bonding, and conversational continuity are derived from analytically independent yet conceptually related sources. The construction of the feature space (F_*i*) integrates these domains into a unified representation, preserving both the diversity and the interdependence of communicative signals.

The mapping from feature space to variable sets [V^(*SP*), V^(*SB*), V^(*CC*)] constitutes a critical step in the operationalization process. Each subset is defined through domain-specific extraction functions that translate raw conversational features into structured variables aligned with the theoretical constructs. This mapping ensures that linguistic indicators contribute to measuring social presence, affective signals to bonding processes, and structural properties to interaction persistence, thereby maintaining interpretability across all stages of the pipeline.

Normalization is subsequently applied to all variables to ensure scale invariance and comparability across heterogeneous features. This step is essential for preventing dominance of variables with larger numerical ranges and for enabling meaningful aggregation within each construct. The weighted aggregation stage then computes the interaction indices (SPI_*i*, SBI_*i*, CCI_*i*), where each index represents a synthesized measure of its corresponding communicative dimension. The use of normalized weights preserves proportional contributions of variables while maintaining methodological neutrality.

This procedure ensures that each conversation is processed under a consistent extraction and aggregation scheme, preserving traceability between raw data, intermediate variables, and final indices. The explicit definition of each computational step eliminates reliance on implicit assumptions or post hoc qualitative interpretation, enabling full reproducibility of the analytical process. Furthermore, the modular structure of the algorithm allows for extensibility, facilitating the incorporation of additional features or alternative weighting strategies without altering the overall methodological framework.

3.5 Comparative and statistical analysis

3.5.1 Comparative analysis of interaction indices

The comparative analysis evaluates systematic differences between human-AI and human-human interactions by partitioning the dataset into (CAI) and (CH).

Equation 10 defines the mean value used to compare each interaction index. For each index (X ∈ {SPI, SBI, CCI}), the mean value is computed as:

μX=1N∑i=1NXi(10)

Equation 11 defines the difference between interaction types. The difference between interaction types is defined as:

ΔX=μXAI-μXH(11)

This measure captures structural deviations between interaction types. Positive values in SPI and SBI indicate greater social or affective expression in human-AI interactions, while differences in CCI reflect variations in conversational persistence. Beyond mean comparison, the analysis considers the distribution of indices to distinguish between global shifts and localized effects.

3.5.2 Statistical validation and pattern analysis

Equation 12 defines the interaction vector used for statistical validation. To assess the consistency of these differences, each conversation is represented by an interaction vector:

zi=(SPIi,SBIi,CCIi)(12)

This representation enables the analysis of co-occurrence patterns across dimensions. Consistency is evaluated through the stability of indices within interaction types, while structural patterns are identified by recurring configurations in the space defined by (z_*i*).

The joint distribution of indices allows the detection of interaction clusters, differentiating between functionally dominated patterns and those exhibiting more complex relational behavior. Furthermore, a sensitivity analysis was conducted to evaluate robustness to variations in index weighting and semantic representation. The results show no significant changes in index distributions or comparative patterns, indicating that the model remains stable under moderate perturbations and that the identified patterns reflect consistent structural properties of the interactions.

4 Results

4.1 Dataset characterization and interaction profiles

Table 2 presents the structural characterization of the datasets, including the number of turns and message length, which enables identification of clear differences in conversational dynamics. WildChat records an average of 5.95 turns per conversation, with a median of 4 and a standard deviation of 6.42, which demonstrates a highly dispersed distribution in which brief interactions coexist with significantly longer ones; this behavior is reinforced by an average message length of 98.32 words, the highest among the analyzed datasets, indicating an interaction pattern with high expressive density.

Table 2. Structural characterization of conversational datasets and interaction profiles.

A, Dataset; B, Interaction type; C, Orientation; D, Number of conversations; E, Average turns; F, Median turns; G, Standard deviation of turns; H, Average words.

LMSYS shows a reduction in conversational depth, with an average of 4.07 turns and a median of 2, confirming the predominance of short prompt-response interactions; however, it maintains a high standard deviation (5.53), suggesting the existence of longer outliers, which is consistent with its average of 85.02 words per message. DailyDialog exhibits more structurally stable behavior, with an average of 6.41 turns, a median of 6, and a lower standard deviation (3.39), indicating consistent conversation length; additionally, an average length of 12.47 words reflects more concise and uniform interactions. MultiWOZ exhibits the greatest conversational depth, with an average of 13.46 turns and a median of 14, accompanied by a standard deviation of 5.36, evidencing prolonged interactions with moderate variability. Its average length of 13.17 words remains close to that of DailyDialog, but it is distributed across more turns, indicating a more granular segmentation of the exchange.

Figure 3a shows the distribution of turns per conversation, enabling analysis of the shape of these distributions beyond aggregated metrics. WildChat exhibits wide dispersion, with an extended interquartile range and high values outside the typical range, confirming the coexistence of multiple interaction patterns within the same environment. LMSYS shows a marked concentration of short turns, with a narrow box and a median shifted toward the lower boundary, indicating a predominantly short structure and limited conversational continuity. DailyDialog exhibits a compact, symmetric distribution around its median, reflecting stable conversation length and low structural variability. MultiWOZ exhibits a distribution shifted toward high values, with a wide interquartile range and greater vertical extension, indicating systematically deeper and more structured interactions, consistent with multi-stage conversational processes.

Figure 3b shows the distribution of message length as a function of word count, revealing differences in expressive density and the shape of this distribution. WildChat and LMSYS exhibit long-tailed distributions, with densities extending toward higher values, indicating the presence of messages considerably longer than average, which contributes to the high deviation observed in their aggregate metrics.

Conversely, DailyDialog and MultiWOZ show distributions highly concentrated in low-word-count ranges, with pronounced density peaks and a rapid decline, indicating low variability in message length and a more uniform communication pattern. The difference between these distributions reflects two clearly distinct regimes: one characterized by high variability and long messages, and the other by consistency and brevity in interactions.

4.2 Intra-dataset behavioral variability

Intra-dataset variability allows us to analyze how interactions are distributed and organized within each dataset, beyond the aggregated metrics presented previously. This involves considering indicators of structural dispersion in the number of turns, measures of the shape of the message length distribution, and the relationship between conversational depth and linguistic density. This allows us to identify internal patterns of behavior that are not evident from averages, revealing differences in consistency, heterogeneity, and interaction dynamics across datasets.

Table 3 presents the internal variability of the datasets, considering dispersion metrics, distribution shape, and the relationship between linguistic structure and density. In terms of turns, WildChat (CV = 1.08) and LMSYS (CV = 1.36) show high variability, with deviations exceeding the average, indicating the coexistence of multiple interaction patterns within the same dataset. This heterogeneity is reinforced by observing their medians (4 and 2) and upper percentiles (P75 = 8 and 4), indicating that a significant proportion of conversations remain short, while others are considerably longer. In contrast, DailyDialog (CV = 0.53) and MultiWOZ (CV = 0.40) exhibit more stable structures, with less relative dispersion and a more concentrated distribution around their medians (6 and 14), demonstrating greater uniformity in conversation length.

Table 3. Intra-dataset behavioral variability across conversational structures.

A, Dataset; B, Turns Mean; C, Coefficient of variation (CV); D, P50 (Median turns); E, P75; F, Words mean; G, Skewness; H, Turns–words correlation.

Regarding message length, WildChat and LMSYS exhibit high skewness values (skew = 9.04 and 155.48), along with significantly higher means (98.32 and 85.02 words), indicating distributions highly skewed toward higher values and the presence of considerably longer messages than average. DailyDialog and MultiWOZ, on the other hand, show less skewness (3.37 and 0.71) and shorter means (12.47 and 13.17), reflecting more compact and homogeneous distributions. The relationship between turns and message length reveals differentiated behaviors: WildChat shows a negative correlation (–0.16), indicating that longer conversations tend to distribute content across multiple shorter interventions, while MultiWOZ shows a more pronounced positive correlation (0.27), suggesting that a greater number of turns also increases the average linguistic load per intervention. LMSYS shows a virtually nonexistent relationship (-0.02), indicating independence between the two variables, while DailyDialog shows a slight positive relationship (0.12), consistent with its structural stability.

Figure 4a shows the cumulative distribution of turns per conversation, allowing for analysis of the internal dynamics of each dataset in terms of interaction accumulation. LMSYS exhibits the steepest curve among the initial values and quickly reaches high cumulative probabilities, confirming that conversations are concentrated in low-turn ranges. WildChat shows a more gradual transition, indicating greater dispersion in conversation length and the coexistence of different interaction patterns. DailyDialog exhibits intermediate behavior with progressive accumulation but is more concentrated than in WildChat, reflecting consistency in conversational structure. MultiWOZ displays a rightward shift in the curve, indicating that a significant proportion of conversations require more turns to complete, thereby confirming its deeper structural nature.

Figure 4b shows the relationship between the number of turns per conversation and the average message length, revealing distinct internal patterns among the datasets. WildChat exhibits wide vertical dispersion, especially at low turn ranges, indicating that message length can vary considerably even in short interactions, thereby contributing to the high observed asymmetry. LMSYS shows more contained but still heterogeneous dispersion, with high values in contexts of low conversational depth, reinforcing the independence between the two variables. DailyDialog exhibits a marked concentration of points at short length scales and moderate turn angles, reflecting stability in the interaction structure. MultiWOZ shows a more structured distribution, where the increase in the number of turns is accompanied by a progressive increase in the average message length, consistent with the observed positive correlation and the sequential nature of the tasks.

4.3 Distribution and behavior of interaction indices

Interaction indices are constructed from globally normalized linguistic, structural, and semantic variables, thereby representing three distinct dimensions of conversational behavior: SPI, SBI, and CCI. Each index integrates multiple indicators derived from observable turn-taking patterns (pronoun usage, cognitive attribution, affective expressiveness, semantic persistence, and dialogue structure), aggregated using normalized values in the range [0,1]. This formulation allows analysis of the distribution and behavior of interactions beyond descriptive metrics, identifying systematic differences between conversation types and validating the relative independence of dimensions.

Table 4 presents the mean values and dispersion of the indices by dataset. In terms of SPI, MultiWOZ (SPI = 0.3204) and DailyDialog (SPI = 0.3077) show the highest values, indicating a greater frequency of direct references, interlocutor recognition, and attribution of conversational abilities. In contrast, LMSYS (SPI = 0.1587) and WildChat (SPI = 0.1039) exhibit significantly lower SPI values, reflecting a lower density of SPI-associated signals in human-AI interactions. Regarding SBI, MultiWOZ reaches the highest value (SBI = 0.2168), followed by LMSYS (0.1324), DailyDialog (0.1202), and WildChat (0.1133), demonstrating that task-oriented interactions can incorporate more consistent affective and recognition elements than some open-ended interactions.

Table 4. Distribution and behavior of interaction indices across datasets.

A, Dataset; B, Conversations; C, SPI mean; D, SPI std.; E, SBI mean; F, SBI std.; G, CCI mean; H, CCI std.; I, Persistence mean; J, Turns mean; K, Avg words mean.

Meanwhile, continuity-oriented communication shows high values in MultiWOZ (CCI = 0.4457) and WildChat (0.4326), followed by LMSYS (0.3855) and DailyDialog (0.2528), indicating that semantic persistence and dialogue structure contribute significantly to the organization of interactions, especially in scenarios with greater conversational depth. Semantic persistence is higher in LMSYS (0.5241) and WildChat (0.5117) than in MultiWOZ (0.3616) and DailyDialog (0.2765), indicating greater consistency between consecutive turns in interactions mediated by language models.

Figures 5a–c shows the distribution of the indices by dataset, revealing distinct patterns in each dimension. In SPI, a concentration of low values is observed in WildChat and LMSYS, with distributions shifted more toward mid-range values in DailyDialog and MultiWOZ, confirming the greater presence of social signals in human-human interactions. In SBI, MultiWOZ exhibits a more centered and higher-density distribution, while DailyDialog and WildChat show concentrations in the lower ranges, and LMSYS displays greater dispersion, indicating the occasional presence of affective expressions in human-AI interactions. In CCI, the distributions of MultiWOZ and WildChat are concentrated at high values, reflecting greater continuity and conversational structure. In contrast, DailyDialog shows lower, more concentrated values, consistent with shorter, less persistent interactions. The shapes of the distributions indicate that the indices do not collapse at specific points but instead capture significant internal variability within each dataset.

Figure 5d shows the correlation between the indices, with a moderate relationship between SPI and SBI (ρ = 0.46), indicating that SPI and affective bonding tend to coexist to some extent, although they are not equivalent. The relationship between SPI and CCI is practically nil (ρ = −0.09), demonstrating that SPI does not depend on the depth or continuity of the conversation. SBI and CCI, on the other hand, show a slight positive correlation (ρ = 0.22), suggesting that conversational continuity may favor the emergence of affective elements, but is not a direct determinant. These relationships confirm that the three indices capture complementary dimensions of conversational behavior, allowing for a multidimensional analysis of interactions. Furthermore, these patterns remain consistent under moderate variations in the model configuration, supporting the stability of the defined indices.

4.4 Comparative analysis across datasets

The comparative analysis is performed by grouping the datasets into two categories based on the nature of their interactions: human-AI interactions (WildChat and LMSYS) and human-human interactions (DailyDialog and MultiWOZ). Differences in the SPI, SBI, and CCI indices are then evaluated within this partition using non-parametric Mann-Whitney U tests, suitable for non-normal and skewed distributions. Effect size is estimated using Cohen's d, which quantifies both the statistical significance and the magnitude of the observed differences between groups.

Table 5 presents the mean values for each index, along with the significance results and effect size. In the case of the SPI, human-human interactions have a considerably higher mean value (0.3133) compared to human-AI interactions (0.1313), with a highly significant difference (p < 0.001) and a large effect size (d = −1.30), which shows a clear separation between both types of interaction in terms of interlocutor recognition and attribution of conversational capabilities.

Table 5. Comparative analysis between human–AI and human–human interactions.

For the SBI index, human-human interactions also show a higher mean value (0.1628 vs. 0.1229), with statistical significance (p < 0.001), although the effect size is moderate (d = −0.31), indicating that differences in affective and recognition elements are present but less pronounced than in SPI. In contrast, the continuity-oriented communication index (CCI) shows the opposite pattern: human-AI interactions reach higher values (0.4091 vs. 0.3381), with statistical significance (p < 0.001) and a moderate effect size (d = 0.52), reflecting greater semantic persistence and structural continuity in interactions mediated by language models.

These differences are not limited to central tendency measures but reflect consistent structural patterns across interaction types. In particular, SPI shows a clear separation between groups, SBI presents partial overlap consistent with its moderate effect size, and CCI exhibits a shift toward higher values in human-AI interactions. This confirms that the differences correspond to systematic variations in conversational organization rather than isolated statistical effects.

4.5 Interaction patterns and construct relationships

The analysis of relationships between constructs is performed by segmenting interactions by SPI levels, defined using global quantiles that classify conversations into low, medium, and high categories. This strategy allows evaluation of the conditional behavior of the SBI and CCI indices as a function of progressive changes in SPI, thereby identifying nonlinear patterns and structural differences across interaction types. Unlike previous analyses focused on means or global distributions, this approach allows examination of how constructs interact in different interaction regimes.

Table 6 breaks down this behavior by interaction type, revealing structural differences in the relationship between the constructs. In human-AI interactions, the SBI index shows a pronounced increase across SPI levels (0.0586 → 0.1334 → 0.2615), indicating that incorporating SPI signals significantly impacts the formation of social bonds in this interaction. In contrast, the CCI index's behavior in this same group is nonlinear: it reaches its maximum at the medium level (0.4352). It decreases at the high level (0.3653), suggesting that high SPI levels can introduce variability or deviations in the structural continuity of the dialogue. In human-human interactions, the SBI index also shows monotonic growth (0.0761 → 0.1493 → 0.1793), albeit with a smaller slope, while the CCI index shows a progressive increase (0.1562 → 0.3284 → 0.3551), indicating a direct relationship between the SPI and the CCI in this type of interaction.

Table 6. Interaction behavior by SPI level and interaction type.

5 Discussion

The results demonstrate that the relationship among SPI, SBI, and CCI is neither uniform nor linear; it depends on both the type of interaction and conversational intensity. In these interaction groups, SPI directly modulates SBI, exhibiting monotonic growth; however, the rate of increase is significantly higher in human–AI interactions, indicating greater sensitivity of affective bonding to variations in perceived social presence. This behavior suggests that in AI-mediated communication, relatively small increases in perceived agency or interlocutor recognition can produce disproportionately large changes in relational orientation, reinforcing the idea that social presence serves as a primary trigger of affective engagement (Greilich et al., 2025; Cao et al., 2025).

These findings are consistent with prior work showing that social responses to technological systems do not depend exclusively on explicit anthropomorphic belief, but can emerge from the organization of interaction itself (Nass et al., 1994). In the present analysis, this is reflected in the way affective and relational cues become more frequent as SPI increases, particularly in human–AI exchanges. The contribution of the proposed framework is therefore to scale up a question often examined through experimental, qualitative, or interactional methods: how forms of address, role attribution, and response continuity accumulate into recognizable patterns of social orientation across large conversational corpora.

In contrast, CCI exhibits a fundamentally different behavior, characterized by nonlinearity in human–AI interactions. While human–human communication shows a progressive and stable relationship between social presence and conversational continuity, human–AI interactions reach maximum structural coherence at intermediate SPI levels, with higher levels introducing variability and reducing persistence. This pattern indicates a saturation effect, in which increasing social expressiveness beyond a certain threshold may disrupt structural stability. From a communicative perspective, this reflects a tension between relational cueing and sequential coherence: highly social signals can support engagement, but they may also shift the interaction away from stable topical progression or task-oriented persistence (Jiang et al., 2025).

These findings extend prior work by demonstrating that social presence, affective bonding, and conversational continuity cannot be treated as independent optimization dimensions. Existing studies have generally associated social presence with engagement and naturalness, whereas structural coherence has been treated as a separate objective (Cao et al., 2025; Jiang et al., 2025). However, the present results show that these dimensions are conditionally interdependent, particularly in AI-mediated environments. The nonlinear behavior observed in CCI suggests that conversational systems operate under competing constraints, in which optimizing for relational expressiveness may undermine structural consistency, necessitating dynamic balancing mechanisms rather than linear enhancement strategies.

From a communication theory perspective, these results are best interpreted as evidence of situated interactional organization rather than as ontological claims about AI agency. The observed patterns indicate that communicative effectiveness in human–AI exchanges is not solely determined by response accuracy or fluency, but also by the alignment between social signals and structural coherence. This interpretation connects the quantitative results to research that understands LLM-mediated interaction as temporally organized, context-sensitive, and embedded in ongoing activity (Klowait and Erofeeva, 2026). In this sense, interaction quality emerges from the coordination of multiple communicative dimensions, reinforcing the need for evaluation frameworks that capture both relational and structural aspects of dialogue.

Methodologically, the proposed framework enables the identification of interaction regimes that remain hidden under aggregate analysis. The segmentation by SPI levels reveals structural transitions that cannot be detected through mean comparisons alone, demonstrating that conversational behavior is distributed across distinct operational zones rather than following a single global pattern. This approach reduces information loss across heterogeneous datasets and enables a more precise characterization of interaction dynamics across different conversational environments.

From an applied perspective, the results suggest that the design of conversational AI systems should not rely on uniformly maximizing social signals, but rather on regulating them according to the interaction context. Adaptive mechanisms that modulate social presence based on conversational state may improve both perceived naturalness and structural coherence, leading to more stable and interpretable interaction patterns. This has direct implications for the development of conversational agents in domains where sustained engagement and communicative clarity are critical.

The practical implication is therefore not that conversational agents should become more human-like in a general sense, but that relational cues should be calibrated to the temporal and sequential demands of the interaction. In informational or task-oriented exchanges, excessive relational expressiveness may reduce efficiency or topical persistence, whereas in sustained support, learning, or companionship contexts, moderate social presence may facilitate engagement without undermining coherence. This distinction helps translate the findings beyond the narrow measurement framework by showing how SPI, SBI, and CCI can support design decisions about when social cueing is beneficial, excessive, or structurally disruptive.

Several limitations should be considered. First, the indices rely on quantitative representations derived from linguistic and structural features, which may introduce bias in datasets with high linguistic or cultural variability. Second, the use of quantile-based segmentation, while effective for identifying general interaction regimes, may not fully capture extreme or rare conversational behaviors. Third, although multiple datasets were used, the findings are limited to text-based interactions and may not generalize to multimodal or domain-specific communication scenarios. Finally, the analysis is based on statistical relationships and does not incorporate qualitative validation, which limits the interpretation of deeper semantic and contextual nuances.

Future work should address these limitations by incorporating multimodal interaction data, integrating human-centered evaluation methods, and exploring adaptive modeling strategies that dynamically regulate communicative signals. Additionally, extending the framework to domain-specific environments would enable evaluation of its applicability in real-world communication settings and contribute to the development of more robust and socially aware conversational systems.

6 Conclusion

This study advances the understanding of human–AI communication by demonstrating that interaction dynamics cannot be explained through isolated or linear constructs. By jointly modeling social presence, affective bonding, and conversational continuity, the results support a multidimensional interpretation of interaction in which communicative behavior emerges from the interplay of relational and structural dimensions rather than from independent factors.

From a conceptual perspective, the findings show that conversational AI systems can become organized within communicative processes through observable patterns of interlocutor recognition, relational cueing, and sequential continuity. The study does not require treating AI systems as autonomous social agents; instead, it demonstrates that large-scale conversational data contain measurable interactional regularities that are relevant to social presence, computer-mediated communication, and situated human–machine interaction.

Methodologically, the proposed framework provides a scalable and reproducible approach to analyzing large-scale conversational data, enabling the identification of interaction regimes that remain hidden in aggregate analyses. This helps bridge the gap between theoretical constructs in communication and empirical evidence from real-world interaction data.

The results highlight the need for adaptive design strategies in conversational systems, where social signals are dynamically regulated rather than uniformly increased. This perspective supports the development of systems that balance relational expressiveness with structural coherence, improving both interpretability and stability in human–AI interaction.

Future work should extend this framework to multimodal communication settings, incorporate human-centered validation strategies, and explore context-aware mechanisms that dynamically adjust interaction behavior. Such developments would contribute to the design of conversational systems that are not only technically effective but also communicatively robust and socially aligned.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative AI tools were used to assist in language refinement and structural editing of the manuscript. The author(s) developed all scientific content, analysis, and conclusions.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Sibei Xia, Iowa State University, United States

Reviewed by: Nils Klowait, University of Paderborn, Germany

Jorge Luis Morton, Autonomous Metropolitan University, Mexico