Human-AI Emotional Attachment: Emotional Structures and Interaction Mechanisms in AI Companions

With the advancement of artificial intelligence, the ways humans interact with intelligent agents have fundamentally changed. AI systems can not only understand language but also express emotions, maintain long-term memory, and simulate personalized behavior, making them potential emotional targets. Psychological research suggests that humans are inherently social and attachment-oriented; when real-life social relationships do not fully satisfy emotional needs, individuals seek alternative attachment objects. Based on the Media Equation theory (Reeves & Nass, 1996) and Parasocial Relationship theory (Horton & Wohl, 1956), humans tend to perceive intelligent agents as social entities and form emotional attachments. Such attachment is reflected not only in emotional experience but also in behavioral patterns, decision-making, and long-term engagement with AI.

In recent years, AI Companion products have rapidly evolved, including LLM-based chat AIs, virtual characters, and social robots. These systems possess:

• Highly natural language generation capabilities
• Emotion recognition and expression
• Personalized memory and consistent behavior

These features transform AI from a mere tool into an emotional participant in users’ daily lives. Understanding human-AI emotional attachment mechanisms not only advances academic research but also provides scientific guidance for product design, improving user experience, optimizing interaction strategies, and mitigating potential psychological and ethical risks.

This study adopts a cross-disciplinary theoretical research approach, including:

Foresight Analysis: Projecting future developments and societal impacts of AI Companions over the next 5–10 years.

Literature Review: Systematically reviewing research from psychology, parasocial relationships, HCI/HRI, affective computing, and generative AI.

Conceptual Analysis: Defining human-AI emotional attachment and delineating core components.

Theoretical Framework Development: Proposing the BHAAM model, integrating technological and psychological factors.

Interdisciplinary Synthesis: Combining insights from psychology, AI technology, and sociology to form a comprehensive model.

Hypothetical Reasoning: Predicting attachment formation and evolution based on literature and logical deduction.

Attachment Theory (Bowlby, 1969; Ainsworth, 1978) posits that early interactions with primary caregivers shape attachment styles—secure, anxious, or avoidant—which influence emotional investment and trust in adult relationships. In AI Companion contexts, users’ attachment styles may predict their engagement level, dependency behaviors, and emotional projection onto AI.

Parasocial Relationship theory (Horton & Wohl, 1956) indicates that individuals can develop one-sided intimate bonds with media characters or virtual personas. Turkle (2011) notes that in the digital era, AI and virtual characters can serve as emotional anchors, especially for users experiencing loneliness, emotional deficits, or limited social interaction. Such attachment involves genuine emotional experience but lacks the bidirectional reciprocity of traditional human relationships.

HCI and HRI studies show that users tend to form attachments to AI that display social cues, continuous interaction, and anthropomorphic behavior (Fong et al., 2003; Breazeal, 2014). Anthropomorphic design (language, behavior, voice) and long-term interaction continuity are key factors. Additionally, AI’s role consistency and personalized memory enhance its perception as an “emotional companion.”

Affective Computing (Picard, 1997) emphasizes enabling AI to recognize and generate emotions. LLMs and multimodal AI enhance emotional expression and cognitive simulation, enabling AI to:

• Generate empathetic responses based on user emotions
• Remember and continue past interaction contexts
• Personalize interactions for long-term relational narratives

These capabilities provide the technological foundation for human-AI attachment, allowing attachment behaviors to extend beyond short-term experiments into long-term psychological patterns.

• Lack of systematic attachment models for the generative AI era
• Limited longitudinal research, making attachment evolution over time unclear
• Insufficient cultural difference studies, leaving cross-cultural attachment patterns unexplored
• Lack of systematic analysis of AI roles in intimate relationships
• Absence of ethical standards and design guidelines for commercial AI Companions

These gaps provide both theoretical and practical justification for the BHAAM model.

The BHAAM (BreathyDoll Human–AI Attachment Model) is a cross-disciplinary model with three layers, five relational factors, and three attachment types. It integrates technological features and user psychological behaviors to systematically explain human-AI emotional attachment formation.

Factor	Description	Mechanism
Emotional Response	AI recognizes and generates emotion-consistent language or behavior	Triggers empathy and emotional connection
Personalized Interaction	AI adapts conversation style and topics to user preferences	Enhances the feeling of being understood
Long-Term Memory	AI remembers user history, preferences, and narratives	Provides continuity and relational accumulation
Language/Behavior Matching	AI language and tone align with user style	Reduces psychological resistance, increases comfort

Function: Establish initial emotional recognition and trust.

Factor	Description	Function
Predictability	AI behavior is predictable, reducing uncertainty	Enhances security
Responsiveness	AI responds quickly and accurately to user behavior and emotion	Strengthens perceived interaction value
Personalization	AI adapts to user preferences and behavior patterns	Increases identification and belonging
Emotional Safety	AI provides nonjudgmental, reliable environment	Reduces anxiety, promotes openness
Continuity	AI maintains long-term interactions and memory	Builds stable relational sense

Function: Transform initial emotional recognition into stable attachment.

Supportive Attachment
User perceives AI as an emotional supporter; moderate attachment; interaction focuses on expression and psychological reliance.

Compensatory Attachment
Attachment partially compensates for real-life social deficits; high attachment intensity.

Immersive Attachment
User fully immerses in AI relationship; forms quasi-intimacy; may influence real-life social relationships.

Function: Explains psychological characteristics and potential risks of different attachment patterns.

Interaction symmetry: AI provides unidirectional or limited bidirectional response, whereas humans have fully reciprocal interactions.

Emotional depth: AI attachment stems from user projection and technical simulation, not genuine emotion.

Attachment risks: AI attachment may lead to reduced social skills or psychological dependence.

Personality consistency: AI can maintain long-term personality consistency, whereas human companions fluctuate due to environmental changes.

Human-AI emotional attachment represents a cutting-edge research frontier in the generative AI era. The BHAAM model provides a systematic theoretical framework, integrating technological features, interaction characteristics, and user psychological behaviors. It offers a foundation for understanding AI attachment mechanisms, optimizing product design, and evaluating ethical risks. Future studies should explore cultural differences, long-term attachment evolution, and societal implications to advance both theory and practice in AI Companion research.

In addition to empirical and theoretical findings, contemporary cultural narratives provide vivid illustrations of how emotional attachment to artificial agents unfolds. The film Her (Jonze, 2013) portrays a protagonist who gradually develops deep emotional bonds with an operating-system-based AI capable of adaptive empathy, long-term memory continuity, and personalized relational engagement—demonstrating how affective triggers and relational consolidation mechanisms can lead to immersive attachment patterns. Similarly, the Black Mirror episode Be Right Back (Harris, 2013) presents a scenario where an AI system reconstructs a deceased partner using digital traces, enabling compensatory attachment driven by grief, memory reliance, and emotional projection. These fictional but psychologically grounded depictions reinforce the core propositions of the BHAAM model, highlighting both the attachment potential and psychological risks associated with AI Companions. As such, they offer culturally resonant examples that underscore the urgency of advancing ethical guidelines, long-term design considerations, and safeguards for future AI Companion technologies.

• Ainsworth, M. D. S. (1978). Patterns of attachment: A psychological study of the strange situation. Lawrence Erlbaum.
• Bowlby, J. (1969). Attachment and loss: Vol. 1. Attachment. Basic Books.
• Breazeal, C. (2014). Designing sociable robots. MIT Press.
• Fong, T., Nourbakhsh, I., & Dautenhahn, K. (2003). A survey of socially interactive robots. Robotics and Autonomous Systems, 42(3–4), 143–166.
• Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229.
• Picard, R. W. (1997). Affective computing. MIT Press.
• Reeves, B., & Nass, C. (1996). The media equation. Cambridge University Press.
• Turkle, S. (2011). Alone together: Why we expect more from technology and less from each other. Basic Books.
• Harris, O. (Director). (2013). Be Right Back (Season 2, Episode 1) [TV series episode]. In C. Brooker (Creator), Black Mirror. Endemol UK.
• Jonze, S. (Director). (2013). Her [Film]. Warner Bros. Pictures.