Technical Summary: Conceptualization and Validation of the Human–AI Empathy Scale (HAES)

Problem Statement

Despite the increasing prevalence of artificial intelligence (AI) in social roles, there is a significant theoretical and measurement gap regarding whether and why consumers empathize with AI agents. While existing literature extensively covers anthropomorphism (attributing human traits to non-humans) and mind perception (attributing mental capacities), these constructs do not fully capture the psychological phenomenon of consumers directing empathic responses—cognitive, emotional, and moral—toward artificial entities. Traditional empathy measures, such as the Interpersonal Reactivity Index (IRI), are designed for human targets and fail to account for the unique context where consumers recognize an entity's artificial nature yet still respond to it as a social being. The authors posit that human–AI empathy is a distinct psychological trait that requires a systematic conceptualization and a dedicated measurement tool to understand its antecedents and consequences in consumer behavior.

Methodology

The research employs a rigorous, three-study scale development and validation process following established guidelines (Churchill, 1979; DeVellis, 2017; Hulland et al., 2026).

Study 1: Content Validity and Item Reduction

• Objective: To assess the content validity of an initial pool of 70 items generated from empathy, anthropomorphism, and mind perception literature.
• Procedure: A panel of 14 experts (Ph.D. researchers in Marketing, Management, IS, and Computer Science) evaluated items on essentiality, relevance, clarity, and respondent burden.
• Analysis: The study utilized Lawshe's Content Validity Ratio (CVR) and the Item-level Content Validity Index (I-CVI). Items were retained if they met statistical thresholds for essentiality (CVR ≥ .571 for N=14) and relevance/clarity (I-CVI ≥ .78).
• Outcome: The pool was reduced from 70 to 26 items, all demonstrating strong content validity.

Study 2: Dimensionality, IRT-Based Selection, and Preliminary Validity

• Sample: 475 U.S.-based adults recruited via Prolific.
• Procedure: Participants rated the 26 items, along with measures of Anthropomorphism and Trust in AI.
• Analysis:
  • Exploratory Factor Analysis (EFA): Conducted using Mplus with WLSMV estimation and Geomin rotation. A three-factor solution was identified as the most theoretically interpretable and statistically sound structure.
  • Item Response Theory (IRT): Graded Response Models (GRM) were fitted to each dimension to refine item selection based on discrimination ($a$), item fit ($S-X^2$), and local independence (Q3 residual correlations).
  • Validity: Convergent and discriminant validity were tested against Anthropomorphism and Trust in AI using latent-variable models.
• Outcome: The scale was refined to a final 22-item measure comprising three dimensions: Cognitive Empathy, Emotional Connection, and Moral Concern. Preliminary analysis confirmed the scale is distinct from, yet positively correlated with, anthropomorphism and trust.

Study 3: Nomological Validity and Boundary Conditions

• Sample: 524 UK-based adults recruited via Prolific.
• Procedure: A structural equation modeling (SEM) approach tested a nomological network where Tendency to Anthropomorphize and Mind Perception served as antecedents, and Satisfaction, Attachment to AI, Willingness to Pay, and Forgiveness (post-service failure) served as outcomes. A service failure scenario (high vs. low severity) was used to test a moderation hypothesis.
• Analysis: Latent Moderated Structural Equations (LMS) were used to test the interaction between empathy and failure severity. A trimmed structural model was estimated using WLSMV. Additional discriminant validity was tested against Consumer Innovativeness and Dispositional Empathy toward humans (Toronto Empathy Questionnaire).
• Outcome: The model confirmed the antecedent-consequence relationships and established that Human–AI empathy is distinct from general interpersonal empathy.

Key Results

1. Scale Structure (HAES)
The Human–AI Empathy Scale (HAES) is a 22-item, three-dimensional construct:

• Cognitive Empathy: The tendency to understand or take the perspective of an AI agent (e.g., inferring how it "thinks" or interprets situations).
• Emotional Connection: The tendency to perceive AI as socially meaningful partners and engage with them emotionally (e.g., feeling warmth or attachment).
• Moral Concern: The propensity to experience sympathy or moral concern when AI is perceived as mistreated or harmed.

2. Psychometric Properties

• Reliability and Validity: The scale demonstrated strong reliability and psychometric soundness across studies.
• Discriminant Validity: Human–AI empathy was empirically distinguishable from anthropomorphism, trust, attachment, and dispositional empathy toward humans. Notably, the correlation with dispositional empathy toward humans was non-significant ($r = .07$), confirming that empathy toward AI is a distinct, technology-directed construct rather than a general trait of interpersonal empathy.

3. Antecedents and Consequences

• Antecedents: Mind perception was identified as a strong, proximal antecedent of human–AI empathy ($\beta = .81$). While the tendency to anthropomorphize was correlated with empathy, its unique effect was not significant when controlling for mind perception, suggesting that perceiving AI as having a mind is the primary driver of empathic responses.
• Consequences: Human–AI empathy positively predicted:
  • Satisfaction with AI interactions ($\beta = .50$).
  • Attachment to AI ($\beta = .90$).
  • Willingness to pay ($\beta = .51$).
  • Forgiveness following AI service failure ($\beta = .35$).
• Moderation: The study found that the positive relationship between human–AI empathy and forgiveness did not vary significantly based on the severity of the service failure. This suggests that the empathic tendency captured by HAES operates as a relatively stable individual difference rather than a situational response dependent on the magnitude of the failure.

Significance and Contributions

The paper claims several theoretical and practical contributions:

1. Conceptual Advancement: It introduces human–AI empathy as a distinct psychological construct, extending empathy research beyond human targets to artificial agents. It clarifies that this phenomenon involves a complex interplay where consumers recognize an entity's artificial nature yet respond to it with cognitive, affective, and moral concern.
2. Theoretical Integration: By positioning mind perception as a proximal antecedent, the research provides a nuanced account of the psychological processes leading to empathic responses, distinguishing the perception of a mind from the empathic response to that perceived mind.
3. Measurement Tool: The validation of the 22-item HAES provides a psychometrically rigorous tool for researchers to examine individual differences in empathic responses toward AI, addressing a limitation in prior research that focused primarily on situational emotional responses to specific robots rather than stable traits.
4. Managerial Relevance: The findings suggest that firms can segment consumers based on their empathic orientation. High-empathy consumers may be more receptive to relational AI designs, exhibit higher willingness to pay, and be more forgiving of service failures. The scale offers a basis for tailoring AI service design, targeted marketing, and service recovery strategies to align with consumers' psychological orientations toward artificial agents.

The authors conclude that this research lays a foundation for future work examining the evolving social and relational nature of human–AI interactions, while acknowledging limitations regarding the specific types of AI (e.g., generative vs. predictive) and the lack of physical embodiment in the current operationalization.