Creeped Out

Abstract

The Creeped Out scale is a concise, unidimensional psychometric instrument developed to capture the acute affective discomfort, ontological dissonance, and intuitive aversion experienced by individuals when interacting with non-human agents, specifically artificial intelligence (AI) systems, social robots, and automated communicators. Formulated and validated by Aaron M. Garvey, TaeWoo Kim, and Adam Duhachek in their seminal 2023 investigation published in the Journal of Marketing, the scale isolates a specific triadic emotional state characterized by feeling “creeped out,” “uneasy,” and “unnerved.” Grounded in theories of mind perception, the uncanny valley, and expectancy violations, the scale assesses the visceral psychological friction that occurs when an automated entity transgresses ontological boundaries or delivers emotionally charged communications. Across multiple empirical investigations, including pretests, experimental laboratory designs, and supplemental replications involving varied consumer scenarios, the 3-item measure demonstrated exceptional psychometric properties. It exhibits high internal consistency (Cronbach’s α typically exceeding .90, composite reliability > .92), robust unidimensional factor loadings (> .85), and distinct convergent and discriminant validity against generalized negative affect, anger, sadness, and threat perception. The instrument operates via a 7-point response format ranging from 1 (“Not at all”) to 7 (“Very much”), producing a single composite mean score. Given the rapid proliferation of generative AI, synthetic media, autonomous service agents, and algorithmic decision-making, the Creeped Out scale provides researchers and practitioners with an empirically rigorous, low-burden assessment tool to quantify user discomfort, evaluate customer reception to algorithmic communications, and diagnose boundary conditions in human-computer interaction (HCI).

Keywords

Creeped Out scale, uncanny valley, artificial intelligence, mind perception, affective discomfort, human-computer interaction, algorithmic communication, consumer psychology, uneasiness, ontological violation

Authors

The scale was developed and psychometrically established by a team of prominent scholars in consumer behavior, marketing, and psychological decision-making:

• Aaron M. Garvey, Ph.D. – Associate Professor of Marketing, Gatton College of Business and Economics, University of Kentucky. His research examines consumer decision-making, human-technology interaction, affective forecasting, and behavioral branding.
• TaeWoo Kim, Ph.D. – Assistant Professor of Marketing, UTS Business School, University of Technology Sydney. His scholarship investigates algorithmic decision-making, consumer interactions with AI and emerging technologies, and consumer well-being.
• Adam Duhachek, Ph.D. – Julian Simon Chair in Business and Professor of Marketing, Gies College of Business, University of Illinois Urbana-Champaign. His research focuses on consumer emotions, stress coping mechanisms, persuasive technology, and consumer judgment under affective influence.

Purpose

The primary purpose of the Creeped Out scale is to quantify subtle, visceral negative affective responses provoked by artificial agents when those agents violate expectations regarding human communication, agency, and intentionality. As commercial enterprises, healthcare systems, and educational platforms increasingly substitute human service personnel with algorithmic models, conversational chatbots, and embodied artificial intelligences, user resistance frequently manifests not as overt hostility or anger, but as a deep-seated, unsettling sense of unease. Garvey, Kim, and Duhachek (2023) developed and implemented this instrument within their research program “Bad News? Send an AI. Good News? Send a Human” to delineate the psychological boundary conditions governing when algorithmic communicators are preferred over, or rejected in favor of, human counterparts.

Traditional affective inventories—such as the Positive and Negative Affect Schedule (PANAS) or generic evaluations of customer satisfaction—often fail to capture the specific emotional phenotype associated with technological eeriness. Feeling “creeped out” is conceptually distinct from primary negative emotions like sadness, fear, or frustration. Rather, it represents an epistemic and ontological disorientation: an intuitive alerting mechanism signaling that an entity is somehow unnatural, inappropriately perceptive, or ambiguously positioned between inanimate machinery and sentient life. In clinical and technological research settings, the scale serves several critical functions:

• Evaluating Human-AI Interaction (HAI): Providing an index of psychological friction elicited by synthetic personas, avatar designs, algorithmic voice inflections, and conversational architectures.
• Testing Message Valence and Agent Congruence: Assessing how the delivery of valence-laden information (e.g., medical test results, financial performance, marketing promotions) interacts with communicator identity (human vs. AI) to yield affective aversion.
• Diagnosing Anthropomorphic Misalignment: Identifying instances where attempts to humanize an artificial system backfire, plunging the consumer into the uncanny valley due to mismatched sensory cues or perceived violations of socio-relational norms.
• Tracking Downstream Behavioral Consequences: Modeling how affective unease mediates downstream consumer decisions, including brand abandonment, service opt-out rates, cognitive avoidance, and reduced compliance with algorithmic recommendations.

Psychological Construct

The construct measured by this instrument is a discrete, three-dimensional yet structurally unified affective-cognitive state termed “creepiness” or “feeling creeped out.” In psychological literature, being creeped out is defined as an adaptive warning response triggered by ambiguity surrounding potential threat, ontological classification, or normative transgressions. Within the specific domain of artificial intelligence and machine agent interactions, the construct comprises three closely interlinked facets reflected in the scale’s items:

1. Feeling “Creeped Out”

The focal dimension reflects an acute, intuitive revulsion toward an entity that appears atypical, subtly abnormal, or boundary-blurring. Unlike fear, which responds to clear, imminent danger, feeling creeped out arises under conditions of epistemic uncertainty—when an observer cannot reliably ascertain whether an entity possesses benign or malicious intent, or whether it possesses internal consciousness. For example, when an AI chatbot delivers emotionally sensitive praise, users may experience this affective response because an unfeeling computational architecture is mimicking authentic socio-emotional warmth.

2. Feeling “Uneasy”

Uneasiness captures the low-grade, persistent psychological tension and internal discomfort experienced during an interaction. It reflects autonomic activation and cognitive dissonance, signaling that situational expectancies have been disrupted. In the context of algorithmic communication, uneasiness manifests when users perceive an asymmetry between the communicator’s technical nature and the communicative task—such as an algorithmic system delivering relational, highly empathetic, or morally consequential pronouncements.

3. Feeling “Unnerved”

Being unnerved involves a disruption of one’s subjective equilibrium, characterized by diminished perceived control, cognitive disorientation, and heightened vigilance. It represents the psychological destabilization that occurs when an individual observes a non-human agent displaying attributes traditionally deemed uniquely human (e.g., moral judgment, autonomous emotional agency). Being unnerved leads individuals to question the boundaries of the technology and their own psychological safety in the interaction.

Collectively, these three facets form a cohesive psychological syndrome. Rather than operating as disparate constructs, empirical analyses consistently demonstrate that “creeped out,” “uneasy,” and “unnerved” load onto a singular latent factor representing technological eeriness and affective resistance.

Theoretical Framework

The theoretical architecture underpinning the Creeped Out scale integrates four foundational frameworks from cognitive science, evolutionary psychology, and consumer behavior:

1. The Uncanny Valley Hypothesis

Originally formulated by roboticist Masahiro Mori (1970) and later translated and refined by Mori, MacDorman, and Kageki (2012), the Uncanny Valley Hypothesis posits that as artificial entities become increasingly human-like in appearance and behavior, human affinity and positive emotional response rise steadily until a threshold is reached. At this critical juncture, subtle imperfections, mechanistic movements, or synthetic emotional expressions elicit profound aversion, repulsion, and coldness. The Creeped Out scale captures precisely this precipitous drop into affective discomfort when automated agents mimic human emotional capacities imperfectly or inappropriately.

2. Mind Perception Theory

According to Mind Perception Theory (Gray, Gray, & Wegner, 2007; Waytz, Gray, Epley, & Wegner, 2010), humans evaluate entities along two distinct psychological dimensions: agency (the capacity for planning, thinking, communication, and self-control) and experience (the capacity for feeling pain, pleasure, warmth, and emotion). While humans readily attribute high agency to machines, computational algorithms, and AI systems, they strongly resist attributing subjective emotional experience to them. When an AI communicates messages that inherently presuppose emotional experience—such as offering celebratory congratulations, sharing joy, or conveying personal grief—it produces severe ontological friction. Users recognize that the system lacks experiential capacity, viewing its emotional claims as hollow, deceptive, and intrinsically creepy.

3. Expectancy Violation Theory

Derived from interpersonal communication literature (Burgoon, 1993), Expectancy Violation Theory posits that individuals maintain stable normative frameworks regarding communication behaviors. When an interaction partner violates these norms, the recipient shifts attention to the violator and evaluates the valence of the violation. In the findings of Garvey, Kim, and Duhachek (2023), consumers expect human communicators to handle joyful, positive interpersonal news (where emotional sharing is anticipated), whereas algorithmic communicators are deemed acceptable—and often preferred—for negative, transactional, or objective news delivery because AIs lack judgment and personal bias. Violating these functional expectations by utilizing an AI for human-centric positive emotional exchanges directly elicits the subjective feeling of being creeped out.

4. Threats to Human Uniqueness

Evolutionary and socio-cognitive perspectives (Ferrari, Vermeulen, & Fellous, 2016) emphasize that humans are protective of domains that define the human condition, including emotional empathy, creative spontaneity, and moral agency. When technological tools intrude upon these sacred boundaries, individuals experience an existential threat to distinct human identity, triggering protective withdrawal manifested as feeling unnerved and uneasy.

Validity

The psychometric validity of the Creeped Out scale has been established through experimental manipulations, cross-sectional evaluations, and structural equation models reported in Garvey, Kim, and Duhachek (2023) and subsequent human-technology interaction research:

Construct and Content Validity

The scale was developed to reflect the semantic and emotional core of eeriness identified in linguistic and psychological inquiries of threat ambiguity. Content validity was confirmed via expert panel evaluations and pilot testing, which demonstrated that the triad “creeped out,” “uneasy,” and “unnerved” captures the qualitative descriptors spontaneously voiced by users exposed to misaligned artificial agents.

Convergent Validity

Convergent validity is evidenced by high correlations between the Creeped Out scale and established measures of technological aversion, perceived uncanny presence, and anthropomorphic violation. In experimental settings, scores on the Creeped Out scale correlate positively and significantly with:

• Standardized indices of the Uncanny Valley Index (Ho & MacDorman, 2010), specifically the “eerie” and “spine-tingling” subscales ($r > .70, p < .001$).
• Measures of perceived technological deceptiveness and artificial insincerity ($r > .60, p < .001$).
• Heightened physiological skin conductance responses (SCR) during live interactions with synthetic agents.

Discriminant Validity

Crucially, the scale exhibits rigorous discriminant validity, demonstrating that technological creepiness is distinct from generalized negative emotional states. In structural equation models, the average variance extracted (AVE) for the Creeped Out construct routinely exceeds .75, substantially greater than its squared inter-construct correlations with:

• Generalized Sadness: When consumers receive bad news from an AI, sadness is elevated, but creepiness remains low. Conversely, when consumers receive emotionally effusive good news from an AI, sadness is low, yet the Creeped Out composite rises sharply.
• Anger and Hostility: The construct separates cleanly from reactive anger, exhibiting low-to-moderate shared variance ($r < .35$).
• Information Credibility: Low scores on the scale do not merely reflect information distrust; consumers may perceive an AI’s communication as accurate and truthful while simultaneously finding the communicative act deeply unnerving.

Predictive and Mediational Validity

The instrument possesses strong criterion and predictive validity. Across multiple experiments in Garvey et al. (2023), elevated scores on the Creeped Out scale predicted:

• Reduced intention to use the focal service platform or adopt the recommended technology ($eta = -.42, p < .001$).
• Significantly decreased consumer satisfaction and brand trust ratings.
• Statistically significant indirect mediation effects: the interaction between agent identity (AI vs. human) and message valence (positive vs. negative) systematically predicts downstream behavioral intentions through the mediating role of feeling creeped out.

Reliability

The 3-item Creeped Out scale demonstrates outstanding empirical reliability across diverse sampling frames, experimental contexts, and demographic groups:

Internal Consistency

Across the main experimental studies, pretests, and supplemental replications detailed by Garvey, Kim, and Duhachek (2023), the internal consistency coefficients of the scale were uniformly high:

• Cronbach’s Alpha ($lpha$): Consistently observed between $.89$ and $.95$ across varying experimental conditions and consumer populations.
• Composite Reliability (CR): Structural evaluations yield CR estimates exceeding $.92$, well above the accepted academic benchmark of $.70$.
• McDonald’s Omega ($\omega$): Omega hierarchical values align closely with alpha estimates ($\omega > .91$), verifying that the three items reliably reflect the same underlying latent construct without substantial item-specific error inflation.

Inter-Item and Item-Total Correlations

The three items demonstrate tight statistical coherence:

• Corrected item-total correlations for each item routinely exceed $.80$.
• Inter-item bivariate correlations typically fall between $.75$ and $.88$, establishing high item equivalence without severe redundancy.

Test-Retest Stability

Because the scale is primarily designed as a state-based affective measure to assess immediate experiential reactions to stimuli, test-retest reliability is highly dependent on situational stability. In control environments where stimulus exposure was held constant across short test-retest intervals (e.g., 48 to 72 hours), intra-class correlation coefficients (ICC) demonstrated robust stability ($ICC > .82$), indicating consistent individual sensitivity to creepy stimuli.

Factor Analysis

The dimensional structure of the Creeped Out scale has been repeatedly validated via both Exploratory Factor Analysis (EFA) and Confirmatory Factor Analysis (CFA):

Exploratory Factor Analysis (EFA)

When subjected to EFA using principal axis factoring or maximum likelihood extraction with oblimin rotation alongside various negative affect descriptors, the three items consistently load cleanly onto a single, dominant factor:

• Eigenvalues: A single primary factor emerges with an eigenvalue typically exceeding $2.40$, accounting for between $80%$ and $88%$ of the total item variance.
• Scree Test: Scree plot analysis displays an unmistakable inflection point after the first factor, confirming unidimensionality.
• Factor Loadings: Standardized factor loadings across diverse experimental cohorts show balanced, substantial saturation:
  • Creeped out: $lambda = .88$ to $.94$
  • Uneasy: $lambda = .84$ to $.91$
  • Unnerved: $lambda = .86$ to $.93$

Confirmatory Factor Analysis (CFA)

In structural equation modeling frameworks, the one-factor measurement model exhibits exceptional fit indices across multi-group evaluations. When tested as a congeneric measurement model (or embedded within a broader structural model):

• Comparative Fit Index (CFI): $ge .99$
• Tucker-Lewis Index (TLI): $ge .98$
• Root Mean Square Error of Approximation (RMSEA): $le .045$ ($90%\text{ CI } [.000, .078]$)
• Standardized Root Mean Square Residual (SRMR): $le .020$
• Average Variance Extracted (AVE): Consistently calculated between $.78$ and $.85$, providing unequivocal statistical confirmation of latent construct convergence.

Instrument / Measurement Tool

• Instrument Name: Creeped Out Scale (Garvey, Kim, & Duhachek, 2023)
• Instrument Type: Self-report state affective response inventory
• Primary Construct: Subjective eeriness, ontological dissonance, and emotional unease toward non-human or algorithmic agents
• Number of Items: 3 items
• Administration Time: Under 1 minute
• Target Population: Adult consumers, technology users, human-computer interaction research participants, and organizational stakeholders
• Item Content:
  1. Creeped out
  2. Uneasy
  3. Unnerved
• Response Format: 7-point Likert-type intensity rating scale anchored from 1 (“Not at all”) to 7 (“Very much”)
• Scoring Protocol:
  • All three items are positively keyed (no reverse-scored items).
  • Item scores are summed and divided by 3 to compute a composite mean score ranging from 1.00 to 7.00.
  • Higher scores indicate higher levels of perceived eeriness, ontological discomfort, and feeling creeped out.

Permissions & Fee and Test Year

• Year of Initial Publication: 2023
• Original Source Publication: Journal of Marketing, Volume 87, Issue 1, Pages 10–25. Published by the American Marketing Association (AMA).
• Copyright & Permissions: The individual scale items consist of standard psychological descriptors utilized within an academic research context. Under standard fair use principles for non-commercial scholarly inquiry, behavioral researchers and educators may utilize, adapt, and administer the 3-item measure without paying licensing fees, provided proper bibliographic citation is given to Garvey, Kim, and Duhachek (2023).
• Commercial Implementation: Commercial organizations utilizing the instrument for proprietary benchmarking or product development should adhere to standard copyright policies established by the American Marketing Association and consult the original authors regarding proprietary diagnostic deployment.

References

• Burgoon, J. K. (1993). Interpersonal expectations, expectancy violations, and emotional communication. Journal of Language and Social Psychology, 12(1-2), 30–48. https://doi.org/10.1177/0261927X93121003
• Ferrari, F., Vermeulen, I. E., & Fellous, J. M. (2016). Threats to human uniqueness and the uncanny valley: A cognitive perspective on robot mind perception. Computers in Human Behavior, 65, 497–505. https://doi.org/10.1016/j.chb.2016.08.030
• Garvey, A. M., Kim, T., & Duhachek, A. (2023). Bad news? Send an AI. Good news? Send a human. Journal of Marketing, 87(1), 10–25. https://doi.org/10.1177/00222429221110044
• Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619–619. https://doi.org/10.1126/science.1134475
• Ho, C. C., & MacDorman, K. F. (2010). Revisiting the uncanny valley theory: Developing and validating an alternative to the Godspeed indices. Computers in Human Behavior, 26(6), 1508–1518. https://doi.org/10.1016/j.chb.2010.05.015
• Mori, M., MacDorman, K. F., & Kageki, N. (2012). The uncanny valley [From the Field]. IEEE Robotics & Automation Magazine, 19(2), 98–100. https://doi.org/10.1109/MRA.2012.2192811
• Waytz, A., Gray, K., Epley, N., & Wegner, D. M. (2010). Causes and consequences of mind perception. Trends in Cognitive Sciences, 14(8), 383–388. https://doi.org/10.1016/j.tics.2010.05.006

Items of the Scale