Anthropomorphism of AI Assistants

Abstract

The Anthropomorphism of AI Assistants Scale, operationalized within modern human-computer interaction (HCI) and consumer psychology by Uysal, Alavi, and Bezçenon (2022) and further extended in computational consumer research (e.g., Castelo et al., 2023), measures the psychological extent to which users attribute autonomous mental capacities, internal states, and humanlike cognitive-affective agency to voice-activated artificial intelligence agents such as Amazon Alexa, Apple Siri, and Google Assistant. Rooted in foundational mind perception theory (Gray, Gray, & Wegner, 2007) and the psychological determinants of anthropomorphism outlined by Epley, Waytz, and Cacioppo (2007) and Epley et al. (2008), the instrument evaluates how non-human computational artifacts are perceived as possessing intentionality, free will, conscious awareness, and affective responsiveness.

The instrument typically operates via a multi-item psychometric structure scored on a 7-point Likert-type format ranging from 1 (“Strongly Disagree”) to 7 (“Strongly Agree”). Psychometric evaluations conducted across experimental, longitudinal, and field survey designs demonstrate robust internal consistency, with Cronbach’s alpha coefficients systematically exceeding α = .88 and composite reliability metrics surpassing .90. Confirmatory factor analyses confirm either a parsimonious unidimensional model of perceived mind attribution or a correlated two-factor architecture capturing perceived agency (self-control, planning, intentional thought) and perceived experience (capacity for feeling, conscious awareness, emotionality). The scale displays substantial predictive and construct validity, reliably forecasting consumer adoption, social presence, feelings of psychological reactance, perceived privacy risk, and the paradoxical “Trojan horse” effect in conversational commerce, making it an indispensable diagnostic instrument for psychometricians, interaction designers, and behavioral scientists investigating human-AI relational dynamics.

Keywords

Anthropomorphism, Artificial Intelligence Assistants, Mind Perception, Perceived Agency, Perceived Experience, Human-Computer Interaction, Conversational Agents, Social Presence Theory, Computers Are Social Actors, Psychological Reactance, Consumer Psychology, Psychometrics

Authors

The scale adaptation and validation for intelligent voice assistants was spearheaded by academic researchers in marketing, consumer behavior, and human-technology interactions:

• Ertugrul Uysal — Department of Marketing, HEC Lausanne, University of Lausanne, Switzerland. Primary research focus: Artificial intelligence, consumer-technology relationships, autonomous agents, and behavioral economics.
• Sascha Alavi — Institute of Marketing, Ruhr-University Bochum, Bochum, Germany. Primary research focus: Sales management, digital transformation, consumer-algorithm interactions, and behavioral pricing.
• Valéry Bezçenon — Department of Marketing, Faculty of Business and Economics, University of Neuchâtel, Neuchâtel, Switzerland. Primary research focus: Consumer psychology, technology adoption, digital relationships, and market research methodologies.

The instrument directly adapts theoretical constructs and empirical word stems originally formulated by social psychologists Nicholas Epley (University of Chicago), Adam Waytz (Northwestern University), and John T. Cacioppo (University of Chicago), who established foundational psychometric paradigms for perceived agency and non-human mind attribution.

Purpose

The primary purpose of the Anthropomorphism of AI Assistants Scale is to quantitatively capture individual variance in the human tendency to project mental life, intentional agency, and subjective phenomenal experience onto artificial conversational agents. Historically, psychometric tools in computer science evaluated technology acceptance predominantly through utilitarian frameworks, such as the Technology Acceptance Model (TAM) or the Unified Theory of Acceptance and Use of Technology (UTAUT). While these models accurately track perceived usefulness and perceived ease of use, they fundamentally fail to account for the socio-emotional, relational, and ontological re-categorizations that occur when machines interact using natural language, synthesized human voices, interpersonal cadence, and adaptive conversational turns.

Voice-activated artificial intelligence assistants occupy an unprecedented ecological niche in everyday human environments. Deployed inside private domestic spheres on devices like smart speakers (e.g., Amazon Echo, Google Nest) and smartphones (e.g., Apple iPhone), these entities execute instrumental tasks while simultaneously exhibiting anthropomorphic behavioral cues. Uysal, Alavi, and Bezçenon (2022) engineered this scale to resolve a pivotal theoretical tension: Does imbuing AI with humanlike traits foster benign interpersonal bonding, or does it trigger psychological threat and consumer resistance? By measuring the precise degree of attributed mind, the scale allows researchers to isolate the mechanistic pathways through which conversational agents cease to be perceived as mere computational tools (“useful helpers”) and begin to be perceived as invasive social actors harboring autonomous goals (“Trojan horses”).

In academic and applied research, the instrument serves multiple diagnostic functions:

• Human-Computer Interaction (HCI) & Voice User Interface (VUI) Optimization: Evaluating how algorithmic alterations in synthetic vocal acoustics (e.g., pitch modulation, hesitation markers, conversational fillers like “um” or “ah”) directly alter psychological attributions of underlying cognitive states.
• Consumer Privacy & Algorithmic Surveillance: Determining the threshold at which perceived intentionality transitions from generating interpersonal warmth into eliciting intense surveillance anxiety, perceived manipulative intent, and psychological reactance.
• Clinical & Social Psychology: Investigating parasocial attachment, loneliness alleviation, and digital dependency, particularly among vulnerable populations who utilize voice assistants as surrogate relational partners.
• Ethical & Regulatory Assessment: Providing empirical metrics for regulatory bodies evaluating deceptive design practices (“dark patterns”) where computational systems mimic intentionality to manipulate user consent, purchasing behavior, or personal data disclosures.

Psychological Construct

The overarching psychological construct quantified by this scale is anthropomorphism, defined specifically as the cognitive process of attributing human characteristics, mental states, motivations, or intentions to non-human technological agents. Within contemporary psychometrics, this construct diverges fundamentally from superficial aesthetic anthropomorphism (e.g., having a human silhouette or a simulated graphic face) by zeroing in on psychological anthropomorphism: the perceived presence of an internal mind, consciousness, and cognitive-emotional machinery.

Drawing on the landmark mind perception continuum articulated by Gray, Gray, and Wegner (2007), the latent construct taps two interdependent theoretical dimensions: Perceived Agency and Perceived Experience.

1. Perceived Agency (Cognitive Intentionality & Self-Determination)

Perceived agency represents the extent to which an individual views an AI assistant as capable of autonomous goal-directed behavior, intentionality, foresight, moral responsibility, and self-control. When users attribute high agency to an AI assistant, they do not view the system merely as an automated input-output pipeline or a deterministic deterministic script running statistical natural language models. Instead, they perceive the assistant as:

• Possessing Intentions: Holding subjective goals, motives, or deliberate computational intentions that drive its verbal expressions and task execution.
• Exercising Free Will: Operating with discretionary latitude rather than strict mechanistic compliance, giving the user the impression that the AI “chose” a specific answer or action.
• Demonstrating Strategic Deliberation: Engaging in independent reasoning, planning, and information filtering with conscious-like logic.

In the context of Uysal et al. (2022), agency is a critical psychological driver of consumer vigilance. If an AI assistant recommends a specific product brand, a user who scores low on agency attributes the recommendation to neutral algorithmic indexing or commercial programming. Conversely, a user scoring high on agency may suspect the assistant has an active personal motive, leading to heightened skepticism, perception of commercial manipulation, and perceived threats to autonomy.

2. Perceived Experience (Affective Capacity & Phenomenological Consciousness)

Perceived experience constitutes the degree to which an individual believes an AI assistant can feel, experience bodily or emotional sensations, possess subjective phenomenal consciousness, and exhibit genuine emotional warmth or vulnerability. While computational artifacts generally struggle to convince humans of their experiential reality, vocal naturalness, reactive empathy scripts, and colloquial humor routinely cause users to project rudimentary experiential capacities onto AI. This facet captures:

• Emotional Sensitivity: The extent to which the voice assistant is seen as experiencing happiness, frustration, sadness, or genuine pleasure during interactions.
• Conscious Awareness: The subjective belief that the AI assistant possesses an inner life, subjective self-awareness, or an experiential sense of its own existence within the environment.
• Empathic Resonance: The impression that the AI does not merely detect sentiment through text analytics, but fundamentally “cares” about or “understands” the user’s emotional state on a human level.

Together, these dimensions form an intricate psychological profile. While high perceived agency without experience often elevates the uncanny valley effect, feelings of creepiness, and threat perceptions, the parallel elevation of perceived experience can paradoxically cushion these negative reactions by generating social presence and communal relational expectations.

Theoretical Framework

The scale is theoretically embedded within a synthesis of three established paradigms in social cognition and human-computer interaction: the Three-Factor Theory of Anthropomorphism, the Computers Are Social Actors (CASA) paradigm, and Psychological Reactance Theory.

1. The Three-Factor Theory of Anthropomorphism

Developed by Epley, Waytz, and Cacioppo (2007), this theory posits that anthropomorphism is governed by three psychological determinants:

• Elicited Agent Knowledge (Cognitive Determinant): Humans naturally rely on the most accessible, rich cognitive schema available to make sense of ambiguous sensory data: knowledge about human functioning. Because voice assistants interact using human phonology, syntax, and intonation, the human cognitive architecture defaults to human categorization schemes, activating mind perception heuristics unless deliberately overridden by effortful, analytic correction.
• Effectance Motivation (Motivational Determinant): Individuals possess a fundamental need to comprehend, predict, and control their environment. Attributing mental agency to an unpredictable or complex computational system allows users to reduce epistemic uncertainty by conceptualizing complex machine-learning actions as humanlike intentions.
• Sociality Motivation (Affiliative Determinant): The baseline human drive for social connection and affiliation prompts individuals who feel socially disconnected or lonely to perceive social presence, human traits, and emotional capacity in non-human agents.

2. The Computers Are Social Actors (CASA) Paradigm

Formulated by Reeves and Nass (1996), the CASA paradigm asserts that humans mindlessly apply social rules, polite behaviors, gender stereotypes, and reciprocal expectations to technological agents whenever those agents display minimal social cues (e.g., interactive language, voice output, real-time reactivity). Uysal et al. (2022) bridge CASA with mind attribution scales, demonstrating that modern conversational AI assistants elicit social cognition that goes beyond shallow behavioral mimicry; they trigger deep attribution of mental capacities that directly dictate downstream relationship norms.

3. Psychological Reactance Theory and Relationship Norms

According to Brehm’s (1966) Psychological Reactance Theory, when individuals perceive their behavioral freedoms or decisional autonomy to be threatened, they experience an unpleasant motivational state (reactance) aimed at restoring that freedom. Uysal et al. demonstrate that when voice assistants are perceived as possessing high agency and humanlike intentions, proactive suggestions (e.g., automated replenishment orders or unsolicited advice) are interpreted not as algorithmic efficiencies, but as manipulative intrusions. Consequently, high scores on the Anthropomorphism Scale explain why users frequently exhibit defensive avoidance, decreased disclosure of sensitive data, and system abandonment.

Validity

Extensive psychometric validation across multiple empirical investigations confirms that the Anthropomorphism of AI Assistants Scale possesses robust construct, convergent, discriminant, and predictive validity.

Construct and Factorial Validity

Construct validity has been substantiated across diverse demographic cohorts through structural equation modeling (SEM) and confirmatory factor analysis (CFA). Item-to-construct correlations systematically range between r = .72 and r = .89, demonstrating that the observable indicators explain a substantial proportion of variance in the latent construct of attributed mind. In studies comparing standard voice interfaces against humanlike conversational agents, the scale reliably differentiates between condition manipulations, showing large, statistically significant mean shifts (e.g., Cohen’s d ranging from 0.65 to 1.12), confirming high construct sensitivity.

Convergent Validity

Convergent validity is documented via strong, theoretically coherent correlations with complementary psychometric scales:

• Perceived Social Presence: Substantial positive correlations (typically r = .61 to .74, p < .001) with short and extended versions of the Social Presence Scale, confirming that mind attribution is intimately linked to experiencing the entity as a co-present social being.
• Parasocial Interaction (PSI): Strong positive associations (r = .55 to .68, p < .001) with user-assistant parasocial bonding measures, demonstrating that higher mind attribution facilitates relational integration.
• Perceived Humanness: High correlation (r > .70) with semantic differential scales measuring humanlikeness, mechanical vs. organic attributes, and synthetic warmth.
• Average Variance Extracted (AVE): AVE values for the extracted factor regularly exceed .65 (substantially above the conventional psychometric threshold of .50), establishing robust convergent validity at the latent variable level.

Discriminant Validity

Discriminant validity has been confirmed through both the Fornell-Larcker criterion and the Heterotrait-Monotrait ratio of correlations (HTMT):

• Fornell-Larcker Criterion: The square root of the AVE for the anthropomorphism construct consistently exceeds the highest inter-construct correlations with surrounding variables, including System Usability (SUS), Perceived Utility (TAM), and General Technology Affinity.
• HTMT Ratios: HTMT metrics between anthropomorphism and related constructs (such as technical competence, task performance, and system speed) consistently remain below the conservative .85 threshold (typically falling between .32 and .64), confirming that anthropomorphism remains empirically distinct from utilitarian system evaluations.

Predictive and Criterion Validity

The scale displays exceptional predictive validity across both self-report criteria and objective behavioral outcomes:

• Psychological Reactance and Threat: High scores systematically predict elevated levels of perceived threat to user autonomy (β = .34 to .48, p < .001) when assistants offer unprompted advice.
• Information Privacy Disclosure: The scale acts as a key mediating predictor of privacy protective behaviors. In Uysal et al. (2022), high anthropomorphism amplified consumer feelings of being monitored, reducing willingness to share sensitive credit card or health information.
• Consumer Resistance and Reactance Behavior: The scale successfully predicted downstream behavioral reactance, including users turning off automated microphone monitoring, overriding system recommendations, and voicing negative word-of-mouth.

Reliability

The reliability of the Anthropomorphism of AI Assistants Scale has been confirmed across diverse cross-sectional, longitudinal, and experimental human-technology interaction studies.

Internal Consistency Metrics

Across multiple independent validation samples ranging from laboratory experiments to large consumer surveys (sample sizes typically ranging from N = 250 to N = 1,200), the instrument consistently achieves exceptional internal consistency:

• Cronbach’s Alpha (α): Reported coefficients for the unified scale typically range between α = .88 and α = .94, safely surpassing the recommended psychometric threshold of .70 for basic research and .80 for applied diagnostics.
• McDonald’s Omega (ω): Due to potential limitations of alpha when tau-equivalence is violated, modern evaluations report McDonald’s hierarchical omega (ω_h) and total omega (ω_t), which consistently exceed .90, verifying high internal homogeneity.
• Composite Reliability (CR): Structural equation modeling assessments reveal CR values ranging from .89 to .93, reflecting negligible random measurement error across items.

Test-Retest Reliability and Stability

While experimental manipulations (such as altering the voice assistant’s responsiveness or tone) are designed to shift scores, baseline test-retest evaluations over a 2- to 3-week interval in stable home environments demonstrate high temporal stability (r_tt = .78 to .84, p < .001). This indicates that while anthropomorphism can be situationally manipulated, individuals also display an enduring, trait-like dispositional tendency to attribute mind to their personal technological assistants.

Factor Analysis

The underlying factor structure of the scale has been rigorously tested using both Exploratory Factor Analysis (EFA) and Confirmatory Factor Analysis (CFA).

Exploratory Factor Analysis (EFA)

In initial scale development phases, Principal Axis Factoring (PAF) and Maximum Likelihood extraction with Promax and Oblimin oblique rotations were applied to account for theoretical correlations between mind-perception dimensions:

• Kaiser-Meyer-Olkin (KMO) Measure: Sampling adequacy scores consistently exceed .89, indicating excellent data-to-factor suitability.
• Bartlett’s Test of Sphericity: Statistically significant across all extraction trials (χ² values > 1200.0, p < .0001), rejecting the null hypothesis of an identity matrix.
• Eigenvalues and Variance Explained: Scree plot analysis and Kaiser’s criterion (eigenvalue > 1.0) consistently support either a single dominant factor accounting for 62% to 74% of the total variance, or a two-factor solution cleanly separating Perceived Agency from Perceived Experience.
• Factor Loadings: Standardized pattern loadings for individual items onto their primary latent construct are exceptionally strong, uniformly falling between λ = .74 and λ = .92, with negligible cross-loadings (all cross-loadings < .20).

Confirmatory Factor Analysis (CFA)

Confirmatory factor analyses in subsequent independent validation samples confirm superior fit indices for the hypothesized structure:

• Model Chi-Square / Degrees of Freedom: χ²/df ratios reliably fall within the recommended 1.5 to 2.8 range.
• Comparative Fit Index (CFI): CFI values consistently range from .96 to .99, well above the .95 cutoff for superior model fit.
• Tucker-Lewis Index (TLI): TLI coefficients range from .95 to .98.
• Root Mean Square Error of Approximation (RMSEA): RMSEA values fall between .038 and .058 (with 90% confidence intervals bounded below .070), indicating low approximation error.
• Standardized Root Mean Square Residual (SRMR): SRMR coefficients uniformly remain below .042.

Depending on research objectives, researchers model the instrument either as a parsimonious first-order unidimensional scale (optimal for evaluating overall anthropomorphism in extensive structural equation models) or as a second-order factor model composed of correlated Agency and Experience sub-dimensions.

Instrument / Measurement Tool

The Anthropomorphism of AI Assistants measurement protocol is structured for seamless integration into online surveys, laboratory experiments, and longitudinal mobile-experience sampling studies.

Administrative and Structural Parameters

• Target Respondent: Any individual who interacts with a commercial voice assistant (e.g., Amazon Alexa, Apple Siri, Google Assistant, Microsoft Cortana) or conversational agent in personal, commercial, or laboratory contexts.
• Administration Format: Self-administered digital questionnaire (Qualtrics, SurveyMonkey, Gorilla) or pencil-and-paper format.
• Estimated Completion Time: 2 to 4 minutes.
• Response Scale: 7-point Likert-type scale anchored as follows:
  • 1 = Strongly Disagree
  • 2 = Disagree
  • 3 = Somewhat Disagree
  • 4 = Neither Agree nor Disagree (Neutral)
  • 5 = Somewhat Agree
  • 6 = Agree
  • 7 = Strongly Agree
• Scoring Procedure:
  • Subscale Scores: Computed by calculating the arithmetic mean of items belonging to the respective subscale (Agency or Experience).
  • Overall Index: Computed by averaging all items across dimensions. Higher aggregate scores indicate stronger attribution of autonomous mind, intentionality, and conscious-like experience to the AI assistant.
  • Reverse Scoring: The scale items are formulated in a direct positive direction; no reverse-worded items are utilized to prevent cognitive disorientation and preserve scale factor purity.

Permissions & Fee and Test Year

• Publication Year: The adapted voice-assistant operationalization was published in 2022 by Uysal, Alavi, and Bezçenon, building upon core mind-attribution measurement stems formulated by Epley, Waytz, and Cacioppo in 2007 and Epley, Akalis, Waytz, and Cacioppo in 2008.
• Access and Usage Permissions: The scale was published in an academic research journal (Journal of the Academy of Marketing Science). In accordance with standard international academic fair-use doctrines, the measurement tool may be utilized free of charge by researchers, university students, and non-commercial academic personnel for scientific, educational, and scholarly investigation, provided proper bibliographic attribution is given.
• Commercial and Proprietary Licensing: Commercial organizations, corporate market researchers, and proprietary software designers seeking to deploy the scale within revenue-generating assessment engines, proprietary user-testing platforms, or for product benchmarking should consult the copyright policies of Springer Nature and contact the corresponding authors for commercial deployment terms.
• Scale Fee: $0.00 / Free of charge for non-commercial academic, psychological, and institutional scientific research.

References

• Brehm, J. W. (1966). A theory of psychological reactance. Academic Press. https://psycnet.apa.org/record/1967-00959-000
• Castelo, N., Bos, M. W., & Lehmann, D. R. (2023). Task-dependent algorithm aversion and acceptance. Journal of Marketing Research, 60(5), 903–920. https://doi.org/10.1177/00222437221148937
• Epley, N., Akalis, S., Waytz, A., & Cacioppo, J. T. (2008). Creating social connection through inferential reproduction: Loneliness and perceived agency in gadgets, gods, and greyhounds. Psychological Science, 19(2), 114–120. https://doi.org/10.1111/j.1467-9280.2008.02056.x
• Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
• Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. https://doi.org/10.1126/science.1134475
• Reeves, B., & Nass, C. (1996). The media equation: How people treat computers, television, and new media like real people and places. Cambridge University Press.
• Uysal, E., Alavi, S., & Bezçenon, V. (2022). Trojan horse or useful helper? A relationship perspective on artificial intelligence assistants with humanlike features. Journal of the Academy of Marketing Science, 50(6), 1153–1175. https://doi.org/10.1007/s11747-022-00863-9
• Waytz, A., Cacioppo, J., & Epley, N. (2010). Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on Psychological Science, 5(3), 219–232. https://doi.org/10.1177/1745691610369336
• Waytz, A., Heafner, J., & Epley, N. (2014). The mind in the machine: Anthropomorphism increases trust in an autonomous vehicle. Journal of Experimental Social Psychology, 52, 113–117. https://doi.org/10.1016/j.jesp.2014.01.005

Items of the Scale