Effects of Visual Anthropomorphism in AI Conversational Agents: A TAM-Based Experimental Study in Task-Oriented Contexts

2-1 Social Cues and the CASA Framework

The Computers Are Social Actors (CASA) paradigm[3] and Sundar’s[1] Theory of Interactive Media Effects (TIME) both suggest that interface cues shape user expectations and evaluations of interactive systems. In AI conversational agents, human-like visual features can activate social responses and influence perceived credibility, usability, and trustworthiness[4],[15]. Prior studies have shown that anthropomorphic cues may enhance perceived usefulness (PU), social presence, and related evaluations[7],[16], but meta-analytic evidence also indicates substantial heterogeneity across contexts[5]. These mixed findings suggest that the effects of social cues are not uniform and may depend on the interaction setting and user goal.

2-2 Visual Anthropomorphic Cues

Visual anthropomorphism refers to the use of human-like visual features such as eyes, facial expressions, or character-like avatars in interface design[4],[5],[12]. Prior research has often reported positive effects of such cues on engagement, warmth, and social presence, particularly in relational or service-oriented contexts[5],[6],[17]. However, recent studies have also shown that anthropomorphic cues can produce neutral or negative effects when they elevate expectations beyond a system’s functional capabilities[8]-[10]. This possibility is especially relevant in utilitarian environments, where users prioritize efficient task completion over social engagement. In such contexts, human-like cues may create an additional interpretive demand and increase cognitive burden rather than support usability[8],[9]. This interpretation is consistent with expectation disconfirmation theory[18], which suggests that when appearance signals greater capability than the system can deliver, users may evaluate the interaction less favorably[19]. The present study focuses specifically on a low-fidelity, screen-based visual anthropomorphic cue: animated eye expressions versus abstract data visualization. Because this manipulation is limited to the visual modality, the findings are bounded to screen-based visual anthropomorphic cues rather than multimodal or high-fidelity anthropomorphism[20]-[22]. Based on these considerations, the following research question is posed:

RQ1: In routine task contexts, how do visual anthropomorphic cues (human-like vs. machine-like) on an AI conversational agent affect users' perceived ease of use, perceived usefulness, trust, attitude, and intention to use?

2-3 Form-Based Role Cues

A second design dimension relevant to conversational agents is role framing: whether an agent appears to be specialized for particular tasks or usable as a general-purpose assistant[1],[11],[13],[23]. In interface design, such role expectations may be conveyed not only through verbal labels or claims of expertise, but also through device form. A task-specific embedded form may imply functional specialization, whereas a consistent standalone form may imply general-purpose availability across contexts[24]-[26]. However, form-based role cues are conceptually distinct from expertise-based specialization. Unlike explicit expertise labels or demonstrated domain competence, device form may influence evaluations through several overlapping mechanisms: perceived functional legibility, visual consistency across contexts, adaptation burden, and the cognitive or interactional costs of switching between multiple device forms[24],[27],[28].

In this study, the form-based role cue was operationalized as a contrast between task-specific embedded forms and a consistent standalone smart-speaker form. This manipulation was initially designed to approximate a Specialist–Generalist distinction. However, because this contrast also varies the degree of form consistency across tasks, any observed differences should not be interpreted as pure evidence of role specialization.

RQ2 exploratory: In routine task contexts, how are task-specific embedded forms versus a consistent standalone form associated with users' perceived ease of use, perceived usefulness, trust, attitude, and intention to use?

2-4 TAM as a Mediating Framework

The Technology Acceptance Model (TAM) posits that perceived ease of use (PEOU) and perceived usefulness (PU) are central determinants of technology acceptance[13],[14],[29]. In the present study, TAM serves as the mediating framework through which visual anthropomorphic cues are expected to shape downstream evaluations. The key analytical question is whether cue effects are transmitted through usability perceptions rather than directly.

The study examines three outcomes: trust, attitude, and intention to use. These are treated as parallel outcomes rather than a causal chain because the study captures pre-adoption evaluations formed after brief exposure to a simulated interaction[5],[7]. This approach allows the analysis to test whether PEOU and PU mediate the effects of social cues on each evaluative dimension while avoiding stronger causal assumptions among the outcomes themselves.

2-5 Hypotheses

H1: Perceived ease of use will be positively associated with (a) trust, (b) attitude, and (c) intention to use.

H2: Perceived usefulness will be positively associated with (a) trust, (b) attitude, and (c) intention to use.

H3: Perceived ease of use will be positively associated with perceived usefulness, and the serial indirect pathway from the visual anthropomorphic cue to outcomes through PEOU and PU will account for differences in (a) trust, (b) attitude, and (c) intention to use.

Fig. 1.

Research model of the study. Solid lines indicate the primary TAM-based mediation model for the visual anthropomorphic cue. Dashed lines indicate the exploratory device-form contrast, initially designed as a form-based role cue

3-1 Participants and Design

This study employed a 2 (Visual Cue: Human-like vs. Machine-like)×2 (Form-Based Role Cue: Specialist vs. Generalist) between-subjects online experiment. Participants were recruited in the Republic of Korea through Macromill Embrain and were eligible if they were 20–49 years old and had prior experience with voice-based AI assistants. The final analytic sample consisted of 195 participants (88 males, 107 females; M_age=33.18, SD=8.08), who were randomly assigned to one of the four conditions. Participants provided informed consent and received monetary compensation upon completion.

3-2 Experimental Stimuli and Procedure

Participants viewed video-based simulations of an AI conversational agent performing 12 routine tasks across bedroom, living-room, and vehicle settings. To minimize brand bias, the agent was presented under a neutral name (“UP”) without association to any commercial product. After viewing the assigned condition, participants completed the questionnaire.

The visual cue was manipulated through the agent’s screen feedback: the Human-like condition displayed animated eyes, whereas the Machine-like condition displayed speech data visualization. The form-based role cue was manipulated through device form: In the condition initially labelled Specialist, the agent was embedded into task-specific objects across tasks; in the condition initially labelled Generalist, it retained a consistent standalone smart-speaker form. Because this manipulation necessarily varied both the intended role cue and the degree of form consistency across tasks, the resulting contrast is interpreted cautiously as a device-form contrast rather than as a pure manipulation of perceived expertise.

Table 1.

Task assignments across settings

Fig. 2.

Four experimental conditions of a conversational agent in three settings

3-3 Measurements

All constructs were measured on 5-point Likert scales using multi-item measures adapted from prior research. Perceived Ease of Use (α=.86) and Perceived Usefulness (α=.88) were adapted from Venkatesh[30] and Park et al.[31]. Trust (α=.84), Attitude (α=.87), and Intention to Use (α=.89) were adapted from established prior studies[32]-[34]. Control Motivation (α=.80) was included as a control variable alongside age and gender [35]. To improve measurement transparency, representative items for each construct are summarized in Table 2.

Table 2.

Measurement constructs and representative items

3-4 Manipulation Check

Manipulation checks were conducted to verify that participants perceived the experimental conditions as intended. Independent-samples t-tests were performed on dedicated manipulation check items administered after exposure to the stimuli.

Visual anthropomorphic cue. The visual cue manipulation was successfully differentiated. Participants in the Human-like condition rated the agent as significantly more human-like than those in the Machine-like condition (Q12: M=2.29, SD=0.94 vs. M=1.75, SD=0.92). Conversely, participants in the Machine-like condition rated the agent as significantly more machine-like (Q13: M=4.06, SD=0.86 vs. M=3.76, SD=0.78). These results confirm that the visual anthropomorphism manipulation was effective, with participants in the two conditions perceiving the agent’s appearance in the intended directions at a medium effect size.

Form-based role cue. The form-based role cue manipulation yielded weaker differentiation. Neither the function-inferability item (Q17: M_Spec=3.02, SD=0.97 vs. M_Gen=2.86, SD=1.01) nor the function-expressing item (Q15: M_Spec=3.52, SD=0.88 vs. M_Gen=3.49, SD=0.86) reached statistical significance. This result indicates that the form-follows-function design, while conceptually distinct from the generalist form, was not reliably perceived by participants as signaling domain expertise or functional specialization in the manner intended.

This outcome is consistent with the conceptual distinction drawn in Section 2.3: form-based role cues may not straightforwardly activate the expertise heuristic. The specialist stimuli may have been perceived as functionally distinct objects (a lamp, a clock) rather than as domain-expert agents. Because the form-based role cue did not achieve clear perceptual differentiation, all findings related to this variable are treated as exploratory throughout this paper and should not be interpreted as confirmatory evidence for the expertise heuristic in conversational agent design.

4-1 Main Effects of Social Cues: ANOVA Results

Visual anthropomorphic cue (RQ1): The Machine-like condition was associated with significantly higher perceived ease of use than the Human-like condition (M=3.99, SD=0.52 vs. M=3.73, SD=0.63; F(1, 191)=10.97, p<.01, ηp²=.054). The Machine-like condition was also associated with higher perceived usefulness (F(1, 191)=4.02, p<.05, ηp²=.021), more favorable attitudes (F(1, 191)=3.91, p<.05, ηp²=.020), and higher intention to use (F(1, 191)=5.87, p<.05, ηp²=.030). The difference in trust between the two visual conditions was not statistically significant (F(1, 191)=0.13, p>.05, ηp²=.001). In summary, human-like visual cues were associated with lower scores across all TAM and behavioral outcome variables except trust, with the largest effect observed for perceived ease of use.

Form-based role cue (RQ2, exploratory): Because the manipulation check did not confirm the intended Specialist–Generalist distinction, these results are reported as an exploratory device-form contrast. The consistent standalone-form condition was associated with higher perceived ease of use than the task-specific embedded-form condition (F(1, 191)=4.85, p<.05, ηp²=.025), as well as higher trust (F(1, 191)=4.01, p<.05, ηp²=.021) and intention to use (F(1, 191)=5.09, p<.05, ηp²=.026). Differences in perceived usefulness (F(1, 191)=0.37, p>.05) and attitude (F(1, 191)=1.53, p>.05) were not statistically significant. These patterns should be interpreted as possible responses to form consistency or adaptation burden, not as confirmed role-specialization effects.

No statistically significant interaction between the visual cue and the form-based role cue was observed for any dependent variable (all ps>.05). Given the exploratory status of the form-based contrast, this absence of interaction should be interpreted cautiously.

Table 3.

The main effects of social cues

4-2 Mediation Analysis: Visual Anthropomorphic Cue

To examine whether the effect of the visual anthropomorphic cue on downstream outcomes was mediated through PEOU and PU, serial mediation analysis was conducted using PROCESS Model 6 [36] with 5,000 bootstrap resamples. In this analysis, the visual cue served as the independent variable (coded: 0=Machine-like, 1=Human-like), PEOU and PU served as serial mediators, and trust, attitude, and intention to use served as separate outcome variables. Age, gender, and control motivation were included as covariates.

The Human-like visual cue was associated with significantly lower PEOU (β=-0.27, p=.001). This indicates that, in the present routine task context, participants who viewed the anthropomorphic agent perceived interaction as requiring more effort than those who viewed the machine-like agent.

PEOU was strongly and positively associated with PU (β=0.58, p<.001), confirming the well-established TAM pathway in which ease-of-use perceptions contribute to perceived usefulness. PEOU also directly predicted attitude (β=0.18, p<.01) and intention to use (β=0.17, p<.05), supporting H1b and H1c. PU in turn significantly predicted attitude (β=0.63, p<.001) and intention to use (β=0.63, p<.001), supporting H2b and H2c. The serial indirect effect (Visual Cue → PEOU → PU → Outcome) was significant and negative for both attitude (Effect=-0.10, SE=0.03, 95% CI [-0.17, -0.04]) and intention to use (Effect=-0.10, SE=0.04, 95% CI [-0.17, -0.04]), supporting H3b and H3c. After accounting for the mediators, the direct effect of the visual cue on attitude and intention were non-significant (ps>.30).

Neither the direct path from PEOU to trust nor the serial indirect effect through PU reached statistical significance in the visual cue model (p>.05). H1a and H2a were therefore not supported. This finding indicates that the mechanism by which visual anthropomorphic cues affected attitude and intention to use—through reduced ease of use and subsequent lower perceived usefulness—did not extend to trust. Trust, as a cognitive assessment of reliability, may depend on additional antecedents (e.g., perceived competence, transparency, or consistency of agent behavior) that were not differentially activated by the visual cue in this experiment. This divergence is addressed in Section 5.

Fig. 3.

Mediation effects of visual anthropomorphic cueNote. *p < .05, **p < .01, ***p < .001

Table 4.

Summary of key serial mediation paths

4-3 Exploratory Analysis: Form-Based Role Cue

An exploratory mediation analysis was conducted for the device-form contrast, rather than for a validated Specialist–Generalist distinction. The task-specific embedded-form condition was associated with lower PEOU than the consistent standalone-form condition (β=−0.17, p=.039). PEOU was positively associated with trust (β=0.30, p<.001) and intention to use (β=0.55, p<.001). These findings are suggestive only and may reflect form consistency or adaptation burden rather than role-type preference.

Fig. 4.

Exploratory mediation results for the form-based role cueNote. *p<.05, ***p<.001.

5-1 Theoretical Implications

The most robust finding of this study is that human-like visual cues reduced perceived ease of use in a routine task environment, with the effect fully mediated through the PEOU→PU→Attitude/Intention pathway. This result supports the view that anthropomorphism is context-dependent rather than uniformly beneficial. While prior research has often shown positive effects of anthropomorphic cues in relational or service-oriented settings, the present findings suggest that in utilitarian task environments, such cues may raise expectations that the system cannot meet, thereby undermining initial usability perceptions. This interpretation is consistent with expectation disconfirmation theory and with recent work reporting increased cognitive load or reduced usability in task-oriented AI and IoT contexts[8]-[10].

A second notable finding is that trust did not follow the same mediated pathway as attitude and intention to use. This suggests that trust in AI conversational agents may require evidence beyond initial ease-of-use and usefulness judgments, such as perceived competence, reliability, transparency, predictability, and behavioral consistency. In this experiment, functional performance, response content, voice output, and processing speed were held constant across conditions, which may explain why the visual cue affected attitude and intention but not trust. Thus, the findings indicate a boundary of the TAM-based pathway: PEOU and PU explained attitude and intention more clearly than trust in a short, video-based pre-adoption evaluation.

The device-form findings remain exploratory. Given that the intended Specialist–Generalist distinction was not confirmed by the manipulation check, the observed differences are interpreted as possible responses to form consistency and adaptation burden rather than as expertise-based specialization effects. This account remains tentative and requires confirmatory testing.

5-2 Practical Implications

Within the limited context of short video-based pre-adoption evaluations, these findings cautiously suggest that visual design for routine-use AI conversational agents may benefit from prioritizing functional clarity over character-based expressiveness. Minimally expressive feedback may be more appropriate than human-like animated expressions when the primary goal is efficient task completion. The exploratory form-based role cue findings should be treated as provisional design considerations for future user experiment rather than direct prescriptions for product design.

5-3 Limitations and Future Research

This study has several limitations. First, the anthropomorphic manipulation was limited to a low-fidelity visual cue displayed on a screen, and thus the findings should not be generalized to multimodal or high-fidelity anthropomorphism involving voice, gesture, or physical embodiment. Second, the form-based role cue did not achieve statistically significant perceptual differentiation. Therefore, all related findings must be regarded as exploratory and should not be interpreted as evidence of a confirmed Specialist–Generalist role effect. Moreover, this manipulation confounded the intended role cue with device-form consistency: the Specialist condition used multiple task-specific embedded forms, whereas the Generalist condition used a single consistent standalone form. Thus, the observed pattern may reflect lower adaptation burden or reduced cognitive cost in the consistent-form condition rather than perceived role generality itself. Future research should disentangle these mechanisms by independently manipulating role framing, form consistency, and demonstrated task competence, such as explicit expertise labels, behavioral demonstrations of specialization, or factorial designs that hold device form constant while varying role information. Finally, given that the observed effects were small to moderate and obtained from a single-session, video-based pre-adoption evaluation, the findings should not be generalized to actual long-term use without further interactive validation.