Researchers at Hongik University in Seoul have established a preliminary empirical architecture defining how conversational artificial intelligence can conduct human-friendly interactions with users. Published in Frontiers in Computer Science, the study by Siu Jung and Yoori Koo establishes the Human-Friendly AI Interaction Design Framework, known as HFAI-DF, by translating established principles of human cognition and anthropomorphism into concrete behavioral mechanisms for computational agents. Modern interactive artificial intelligence systems increasingly operate through sustained, bidirectional natural language dialogues rather than constrained menus or deterministic rule trees. Consequently, academic scrutiny has broadened from standard technical performance metrics such as algorithmic safety, transparency, and data privacy toward the qualitative dynamics of reciprocal communication. The authors frame human-friendly interaction not as superficial conversational politeness, decorative warmth, or physical mimicry, but as an agent’s structural capacity to communicate and collaborate in ways that naturally align with human cognitive expectations and social intelligence competencies.
To build an operational foundation for this interactional shift, the investigators grounded their inquiry in the psychological literature of anthropomorphism, which denotes the pervasive human tendency to attribute consciousness, intentionality, and emotional qualities to non-human entities. Drawing upon foundational theoretical models established by Epley, Waytz, and colleagues, the research team focused on two major structural drivers of anthropomorphic attribution: cognitive determinants and motivational determinants. Cognitive determinants encompass the baseline knowledge systems and reflexive processes humans apply when interpreting another entity, which the authors operationalized as mandatory functional requirements for conversational machines. Guided by biological cognition models from Arbilly and Lotem, these cognitive capabilities were partitioned into four essential areas: the simulated experience of emotion, the maintenance of episodic memory, reflexive metacognition, and self-awareness combined with theory of mind. Meanwhile, dispositional determinants related to personal background were excluded to focus on universal interactional dynamics, while motivational determinants were incorporated to account for situational drivers of user engagement.
The motivational dimension of the framework incorporates two distinct psychological drivers identified in cognitive science: sociality motivation and effectance motivation. Sociality motivation stems from an inherent human drive for interpersonal affiliation, which becomes particularly pronounced when users experience isolation or seek communicative connection, prompting individuals to detect and welcome social cues from non-human conversational agents. Effectance motivation arises from the foundational psychological desire to comprehend, anticipate, and control external environments. When confronting complex or uncertain systems, users naturally project human characteristics onto computational agents to restore perceived predictability and reduce situational anxiety. Jung and Koo analyzed these motivational drivers alongside the four cognitive determinants, systematically deconstructing each theoretical factor into constituent elements. These components were subsequently translated into contextual functional units and synthesized into thirty discrete behavioral units. Each behavioral unit specifies an observable operational capability, ranging from uncertainty tracking to proactive task restructuring, designed to engage human cognitive interpretation without requiring the machine to possess literal internal emotional consciousness.
To evaluate how these thirty proposed behaviors operate within actual user perception, the authors constructed an empirical survey administered to 208 adult participants from Generation MZ, spanning birth years from 1981 through 2012. This demographic focus was chosen purposefully because younger adults frequently engage in sustained natural language exchanges with contemporary generative systems and have acquired substantial experiential fluency. The participant cohort comprised 109 individuals aged 20 to 29 years, 83 individuals aged 30 to 39 years, and 16 individuals aged 40 to 44 years. The measurement protocol utilized a seven-point semantic differential scale incorporating three validated bipolar adjective pairs designed to evaluate conscious anthropomorphic perception: machine-like versus human-like, unnatural versus natural, and artificial versus lifelike. Each behavioral item was presented alongside realistic, contextual interaction scenarios authored by the researchers to depict how conversational agents instantiate the given conduct during practical collaborative exchanges, ensuring responses captured grounded communicative impressions rather than abstract conceptual definitions.
The gathered survey data underwent rigorous statistical examination using exploratory factor analysis in SPSS Version 31.0, specifically applying principal axis factoring accompanied by direct oblimin oblique rotation to accommodate expected conceptual relationships among latent dimensions. Initial diagnostic evaluations confirmed strong sampling adequacy, yielding a Kaiser-Meyer-Olkin measure of 0.93 and a highly significant Bartlett’s test of sphericity with a chi-square of 4388.438 across 435 degrees of freedom at p < 0.001. Applying the eigenvalue-greater-than-one extraction criterion, the analysis identified six distinct latent factors that cumulatively accounted for 62 percent of total data variance. All thirty behavioral items demonstrated satisfactory pattern loadings of 0.32 or greater on their designated primary factors. Internal consistency remained exceptionally high across all derived factors, with Cronbach’s alpha coefficients ranging between 0.82 and 0.91. Corrected item-total correlations exceeded the standard discrimination benchmark of 0.30 for every variable, confirming robust statistical cohesion and conceptual validity across the empirical model.
The resulting framework organizes the six extracted dimensions along an interactional trajectory that begins with the agent’s internal self-regulation and extends outward to collaborative contribution. Dimension 1, labeled Self-Aware Interaction, comprises five items exhibiting an internal consistency of alpha 0.823. This dimension is anchored by identity and role awareness, which demonstrated the highest pattern loading at 0.738, alongside representation of learning history at 0.686, explanation of reasoning grounds at 0.620, self-monitoring and knowledge adjustment at 0.476, and recognition of uncertainty at 0.373. Self-Aware Interaction captures an artificial system’s ability to explicitly articulate its operational boundaries, acknowledge epistemic confidence levels, and transparently convey the rationale behind its answers. The authors emphasize that such reflexive transparency provides users with an essential evaluative anchor, helping them gauge factual reliability and mitigate risks associated with automated hallucinations, algorithmic overconfidence, or fabricated technical assertions during complex consultations.
Moving from internal self-regulation to interpersonal attunement, Dimension 2 and Dimension 3 delineate the system’s inferential processing capabilities. Dimension 2, termed User-Adaptive Responsiveness Interaction, contains three items with an alpha of 0.825: verbal and non-verbal cue detection loading at 0.658, internal state reasoning loading at 0.587, and role-aware utterance adjustment loading at 0.367. This factor reflects the system’s proficiency in parsing conversational signals, inferring user intentions, and modulating intervention styles according to user competencies. Dimension 3, designated Context-Aware Reasoning Interaction, incorporates five behavioral units with an alpha of 0.870. Led by user error recognition and guidance at 0.733, this factor also encompasses rule exception detection at 0.587, temporal flow regulation at 0.514, shared-goal process tracking at 0.475, and narrative structuring via situational markers at 0.437. Together, these two dimensions establish that human-friendly systems must grasp both the evolving psychological state of human partners and the broader temporal context governing cooperative tasks.
The fourth factor, designated Socio-Emotional Interaction, emerged as the central relational hub within the empirical framework, uniting six behavioral units with an internal reliability alpha of 0.872. This dimension is led by empathy-level emotional expression loading at 0.689, followed by relationship continuity emphasis at 0.613, encouraging open interaction at 0.605, emotion-linked memory storage at 0.586, user-specific memory recall at 0.543, and alignment with user perspective at 0.469. Factor correlation analyses demonstrated that Socio-Emotional Interaction maintained moderate correlations of 0.403 to 0.476 with all five other dimensions. This pervasive connectivity indicates that socio-emotional attunement serves as a mediating bridge linking baseline cognitive recognition with external functional collaboration. When an agent recalls past conversational exchanges and responds with proportionate emotional empathy, it reframes fragmented question-and-answer interactions into a continuous, personalized social relationship that strengthens mutual rapport and conversational ease.
The final two dimensions capture the agent’s capacity for proactive and generative contribution in shared problem-solving environments. Dimension 5, named Creative Collaboration Interaction, groups seven items with an alpha of 0.907, featuring context-based expansion at 0.687, presenting conceptual alternatives at 0.605, detail refinement at 0.590, idea transformation at 0.531, creative idea generation at 0.450, proposal evaluation at 0.378, and feedback-based concept adjustment at 0.377. Complementing this, Dimension 6, labeled Proactive Assistance Interaction, contains four items with an alpha of 0.865, including creative attempt explanations loading at 0.749, contextual concept explanations at 0.566, perspective shifting at 0.491, and proactive prompt generation at 0.420. Dimensions 5 and 6 correlated moderately with each other at 0.415, demonstrating an outer operational layer where the artificial intelligence behaves as an active co-creator, systematically proposing novel alternatives, refining conceptual structures, and articulating its generative intent rather than remaining a passive command-driven terminal.
Despite these promising structural findings, Jung and Koo urge critical caution regarding the practical application and ethical deployment of their framework. The authors emphasize that artificial intelligence does not possess genuine feelings, subjective experiences, or living consciousness, warning that uncalibrated human-friendliness risks inducing deceptive over-attachment, damaging emotional dependency, or triggering an uncanny valley revulsion when algorithmic facades fail. Socio-emotional features beneficial in therapeutic or counseling domains may prove intrusive, distracting, or professionally compromising within analytical decision-support tools or high-stakes productivity workflows. Furthermore, the researchers acknowledge clear empirical limitations: the exploratory factor analysis relied on third-person evaluations of written scenarios among 208 Generation MZ participants rather than live, real-time interactive trials across diverse age groups. Future research must conduct confirmatory factor analyses, examine dynamic bidirectional dialogues, and validate how these six behavioral dimensions perform across distinct cultural, demographic, and specialized occupational environments.
Subject of Research: Empirical development and structural validation of the Human-Friendly AI Interaction Design Framework grounded in anthropomorphism theory
Article Title: A preliminary design framework for human-friendly AI interaction: grounded in anthropomorphism
Article References: Jung, S., & Koo, Y. (2026). A preliminary design framework for human-friendly AI interaction: grounded in anthropomorphism. Frontiers in Computer Science, 8, Article 1964763. https://doi.org/10.3389/fcomp.2026.1964763
Image Credits: AI Generated
DOI: 10.3389/fcomp.2026.1964763
Keywords: conversational artificial intelligence, human-computer interaction, anthropomorphism, exploratory factor analysis, social intelligence, user experience design, cognitive modeling, interaction design framework, preliminary, design, framework, human-friendly
News Source: Blake Davidson. (November 10, 2026). New Framework Outlines Six Key Dimensions for Human-Friendly Conversational Artificial Intelligence. Scienmag.
