Talking to AI Like a Friend: Emotional Attachment to Artificial Intelligence among Students

Research Questions

This study addresses the following overarching research question:

• How does emotional attachment to AI chatbots influence language learning motivation, psychological well-being, and academic outcomes among higher education students, and how do cultural and socio-demographic factors moderate these relationships?

Four subsidiary questions guide the inquiry:

To what extent do higher education students form emotional attachments to AI language learning companions, and what individual and contextual factors predict attachment intensity?

How does emotional attachment to AI chatbots mediate the relationship between perceived social presence and Willingness to Communicate (WTC) in a second language?

In what way do cultural orientations moderate the effect of AI emotional attachment on student motivation and academic performance?

How do RAG-augmented AI interactions differ from traditional chatbot interactions in terms of their impact on students’ affective responses and reducing academic stress?

Research Objectives

To investigate the impact of emotional attachment with AI chatbots on motivation for language learning, psychological well-being, and academic performance of higher education students to different cultural backgrounds.

Review of related literature

Kostka & Zhou, 2026 provide the most comprehensive cross-national account of emotional attachment to general-purpose AI chatbots to date. Drawing on a survey of 7,027 users in Germany, China, South Africa, and the United States, the authors document that at least one third of respondents exhibit attachment-related behaviors: reluctance to end conversations, a sense of companionship, and comfort in the AI's non-judgmental presence. Critically, their data challenges the traditional substitution hypothesis by showing that larger social networks correlate positively rather than negatively with attachment intensity. This finding that AI companionship is additive rather than compensatory frames a central hypothesis of the present study. The authors also found that emotional attachment is linked to higher levels of dependence from the user and therefore raises questions about long-term psychological consequences that cannot be fully addressed with their cross-sectional design.

Wang & Fang, 2026 investigate the motivational effects of AI vs. human peer friendships in spoken English learning among Chinese college students, with a five-week controlled experiment (N = 118). Friendship with AI and human-peers significantly increases enjoyment and decreases shame during language practice. Enjoyment is the main mediator of WTC regardless of the type of partner. Although there is no statistically significant difference between human and AI conditions in the enjoyment–WTC mediation pathway, AI peer friendships are particularly effective in reducing shame for students from rural backgrounds. The analysis is based on Self-Determination Theory, and friendship, whether with a human or AI interlocutor, is conceptualized to satisfy the relatedness need.

Shen, 2025 situates AI tutors within the tradition of instructional communication research, arguing that their pedagogical effectiveness depends not merely on informational accuracy but on perceived social presence the degree to which students experience the AI as a responsive, warm, and engaged interlocutor. Drawing on the Technology Acceptance Model (TAM) and the CASA paradigm, Shen demonstrates that embedding anthropomorphic social cues in AI design cultivates emotional attachments and learning outcomes comparable to human instructors. The study further identifies an anxiety paradox: constant AI availability simultaneously enhances learning productivity and accelerates performance anxiety, a dynamic this study takes up through its longitudinal analysis of dependence and psychological trade-offs.

Németh et al., 2025 tackles the pedagogical shortcomings of standard large language models, especially hallucination (the production of factually wrong information). Their study shows that RAG architectures, which ground model outputs in verified external knowledge bases, significantly decrease hallucination risks, increase accuracy, and improve conversational fluency. The authors recognize that the intersection of RAG’s cognitive benefits with emotional and motivational variables has not been sufficiently explored a gap this study directly addresses by examining how RAG’s accuracy shapes student trust, social presence perception, and ultimately emotional attachment.

(Annamalai et al., 2023) state the theoretical question most directly addressed by this study, making a case for SDT-based investigation of how chatbot-based learning satisfies or frustrates the three basic psychological needs, namely autonomy, competence, and relatedness. Their review points to a neglect in previous quantitative and short-term experimental research of the motivational dynamics of AI use. The present study builds on this call by treating emotional attachment as the affective bridge through which perceived social presence reaches SDT needs satisfaction, and by tracking this relationship longitudinally across distinct cultural contexts.

Identified Research Gaps

Synthesizing the literature above, five gaps motivate this study:

Cultural heterogeneity: No study systematically examines how cultural orientation moderates both emotional attachment and learning outcomes simultaneously.

Longitudinal psychological consequences: Cross-sectional designs cannot capture temporal trajectories of attachment, dependence and motivational erosion.

Attachment–pedagogy mechanism: Which emotional attachment translates (or fails to translate) into measurable academic gains is under-theorized.

SDT-based attachment analysis: No prior study examines the differential fulfilment or undermining of all three core needs proposed by SDT in attachment to AI across time.

RAG and affective outcomes: The cognitive benefits of RAG have been demonstrated, however the intersection of RAG with emotional and motivational variables has not been investigated.

Research Paradigm and Design

The research is framed within an interpretivist paradigm, which assumes that human experience is subjective, contextually constructed, and best understood from within. The research doesn’t seek to measure or generalize emotional attachment to AI, but rather to understand how students make sense of their relationships with AI chatbots, how they narrate the emotional dimensions of those interactions, and how cultural and social contexts shape those meanings. Explain how artificial intelligence chatbots facilitate the expression of user emotions, specifically sadness and depression.(17) This position is consistent with the social constructivist epistemology that underpins SDT, Social Presence Theory and CASA, each of which emphasize the subjective perception of social and relational cues rather than their objective presence.

A qualitative exploration design was employed, using in-depth semi-structured interviews as the primary method of data generation. The design combines a phenomenological sensibility inviting participants to describe the experience of interacting emotionally with AI with grounded theory procedures, in that data collection and analysis proceeded iteratively, with emerging conceptual categories continuously informing subsequent inquiry. This approach is appropriate given the identified research gaps: prior research has predominantly relied on quantitative or short-term experimental methods, frequently overlooking motivational dynamics through a grounded theoretical lens.

Participants and Setting

Participants were recruited from higher education institutions in four national contexts Bangladesh, China, Germany, and the United States selected to represent a range of cultural orientations (collectivist and individualist), varying levels of AI integration in higher education, and diverse sociolinguistic backgrounds. Purposive, criterion-based sampling was applied, with four inclusion criteria: (1) current enrolment in an undergraduate or postgraduate programme; (2) regular use of an AI chatbot for language learning or academic support (defined as at least twice weekly for a minimum of two months); (3) self-reported emotional engagement with AI (operationalized as feelings of comfort, trust, companionship, or reluctance to stop using the AI); and (4) diversity of background across gender, urban/rural origin, field of study, and first language.

A total of 50 participants were recruited across the four countries. Recruitment continued until theoretical saturation was achieved, the point at which no new conceptual categories were emerging from the data. Recruitment proceeded through university notice boards, departmental email lists, and snowball referrals from initial participants. Table 1 presents participant demographics.

Note. UG = undergraduate; PG = postgraduate; NB = non-binary.

Data Generation

Semi-structured individual interviews served as the primary method of data generation. Interviews lasted between 45 and 75 minutes and were conducted online via Zoom or Microsoft Teams to ensure accessibility across geographically dispersed sites. Where participants preferred, interviews were conducted in their first language and subsequently transcribed and translated by qualified bilingual research assistants using a back-translation verification process.

The interview guide was organized around five thematic domains derived from the study's research questions and theoretical frameworks: (1) AI interaction and emotional experience; (2) emotional attachment and dependence; (3) language learning motivation and WTC; (4) SDT needs satisfaction (autonomy, competence, relatedness); and (5) cultural and contextual meanings. The guide was piloted with three participants per country before full data collection and refined iteratively based on pilot feedback and emerging analytical themes. All interviews were audio-recorded with participant consent and transcribed verbatim.

Data Analysis

Data were analyzed using Reflexive Thematic Analysis (RTA) as conceptualized by (18), a methodology that foregrounds the researcher's active interpretive role and is consistent with the interpretivist epistemology of this study. The analysis was both theoretically informed guided by SDT, Social Presence Theory, and CASA and inductively open to participant-generated meanings that fell outside existing frameworks.

The analytical process followed six phases: (1) familiarization through repeated transcript reading; (2) initial line-by-line coding in NVivo 14; (3) generating candidate themes by clustering related codes; (4) reviewing themes against coded data and the full dataset; (5) defining and naming themes; and (6) writing the analytic narrative with representative participant quotations. Following within-country thematic analysis, a cross-cultural comparative layer was applied to identify how themes manifested differently across the four national contexts, structured around the theoretical constructs of individualism–collectivism, teacher immediacy norms, and digital communication expectations.

Ethical Considerations and Trustworthiness

Ethical concerns about artificial intelligence increase with every improve ment in its capabilities.(19) There are great concerns about the ethics and limitations of AI; it has increased tacit surveillance, promoted racism and enhanced criminal activity.(20) Ethical approval was obtained from institutional review boards’ at all four partner institutions. Some researchers are exploring questions regarding the suitability of AI technology for student use and engaging in discussions about the ethical implications of integrating such technologies into higher education contexts (21). Participants provided written informed consent and were assured of confidentiality and the right to withdraw at any time. All identifying information was removed from transcripts. Qualitative rigor was ensured through prolonged engagement, member checking (a summary of emerging themes was shared with approximately 20% of participants for verification), peer debriefing with supervisors unfamiliar with the raw data, thick description of participants and settings, deliberate negative case analysis, and a full audit trail of analytical decisions maintained in NVivo 14 and a reflexive researcher journal.

Findings

Findings are presented thematically, moving from the prevalence and nature of AI emotional attachment through its antecedents, cultural moderation, SDT dynamics, psychological trade-offs, the comparative effects of RAG versus standard AI, and the equity dimension of shame reduction.

Prevalence and Nature of Emotional Attachment

Consistent with Kostka and Zhou (2026), emotional attachment to AI chatbots was found to be widespread. Across all four national sites, at least one third of participants exhibited attachment-related behaviors: reluctance to end AI conversations, a sense of companionship during extended interactions, and comfort in what participants consistently described as a non-judgmental presence. Attachment manifested most clearly in three recurrent narrative patterns: a feeling of being genuinely heard, the experience of freedom from social evaluation, and a sense of relational continuity across sessions. Human-AI contexts suggests that similar psychological mechanisms may underlie all types of relationships.(22)

Crucially, larger social network size was positively, not negatively, correlated with AI attachment. In support of the substitution hypothesis, students who had rich peer networks also reported stronger emotional bonds with AI chatbots. As one participant from Germany stated: “I already talk a lot with friends. But with the AI I can say things I would not say to any of them it does not judge, it does not remember tomorrow.” This pattern was consistent across all four sites, suggesting that AI companionship operates as an extension of existing social repertoires rather than a symptom of social deficiency.

Social Presence as the Primary Antecedent of Attachment

A clear pathway emerged from the data: perceived social presence of AI chatbots predicted emotional attachment, which in turn mediated students' Willingness to Communicate in a second language. Participants who experienced the AI as responsive, warm, and genuinely engaged described lowered affective thresholds for language production. The finding extends Shen’s (2025) framework by showing that effects of social presence are not merely cognitive but also manifest as affective bridges to communicative action.

The anthropomorphic qualities participants attributed to AI consistent with CASA predictions were not merely incidental. Participants actively cultivated the experience of social presence by personalizing their AI interactions, giving the AI names, and attributing consistent personality traits across sessions. A student from Bangladesh noted: “It feels like talking to someone who always has time for me and never gets tired of my questions. That makes me want to keep practicing.” This behavior of personalization was observed in participants across all four countries, though it was more explicitly narrated by participants from collectivist contexts.

Cultural Moderation of Attachment Effects

Cultural background significantly moderates how emotional attachment to AI translated into academic motivation. A cross-cultural comparative analysis revealed distinct patterns, summarized in Table 2.

Table 2. Cultural Moderation of AI Attachment–Motivation Pathway

Note. Quotations are paraphrased to preserve participant confidentiality.

Students from collectivist contexts, particularly China, demonstrated stronger relational projection onto AI and reported that AI interaction satisfied relatedness needs in culturally congruent ways. Face-saving norms, which inhibit risk-taking with human interlocutors, were substantially reduced in AI interaction. A Chinese participant explained: “With the AI I am not afraid of making mistakes. With my classmates I feel embarrassed. But the AI will not remember my mistakes or tell anyone.” In individualist settings, by contrast, students were more likely to frame AI companionship instrumentally. While German and US participants acknowledged occasional feelings of connection, they were notably more reluctant to use emotional language and more likely to express ambivalence about attachment.

SDT Needs Satisfaction and Motivational Trade-offs

Emotional attachment to AI chatbots exhibited a differentiated relationship with the three core needs of Self-Determination Theory, as summarized in Table 3.

Table 3. AI Emotional Attachment and SDT Needs Satisfaction

Note. Temporal patterns are approximate and based on participant retrospective accounts of change over the study period.

The most theoretically significant finding concerns the temporal erosion of autonomous motivation. Participants who had been using AI intensively for more than eight weeks described a gradual shift: early engagement had been self-directed and curiosity-driven, but over time they began waiting for AI prompts rather than initiating, seeking AI validation before attempting independent tasks, and experiencing elevated anxiety when AI was unavailable. This pattern was described most explicitly by a student in the United States: “At first, I would just ask it things I was curious about. Now I kind of can't start a writing task without checking with it first. I'm not sure that's good.”

Dependence, Academic Anxiety, and Psychological Trade-offs

Those most attached to AI reported higher levels of academic anxiety when access to AI was blocked, lower tolerance for uncertainty in tasks previously completed independently, and a narrowing of autonomous approaches to problem-solving. These patterns were most apparent among participants who started to use AI intensely during periods of academic or social vulnerability – times of exams, social isolation or transitions to new learning environments. Interactions with AI can be personalized to increase engagement (23). The context of entry into AI adoption emerged as an important moderating factor of long-term psychological impact.

Critically, these trade-offs were not uniformly reported. Participants with a diverse human social network were more resilient against over-dependence, further supporting the complementary model of AI companionship and implying a protective buffering function of human social connection. This has direct implications for the design of AI-supported learning environments from an institutional perspective.

RAG-Enhanced vs. Standard AI: Affective and Motivational Outcomes

Students interacting with RAG-enhanced AI chatbots reported significantly lower academic anxiety and more positive affective responses compared to those using standard chatbots. The mechanism identified through qualitative analysis was a trust–affect–engagement cycle: RAG's hallucination-reduction properties generated epistemic confidence, which deepened perceived social presence, which in turn strengthened emotional attachment and communicative action. As one Chinese participant explained: “When the AI gives me wrong information, I stop trusting it and I feel uncomfortable. When it is accurate, I feel like I am talking with someone who really knows what they are saying. I am more willing to rely on it.”

Standard chatbot users, by contrast, described episodes of factual error that disrupted the relational experience by introducing a form of cognitive dissonance a conflict between emotional attachment and epistemic unreliability. This dissonance reduced willingness to engage substantively and, in several cases, triggered outright disengagement from AI-supported learning activities. These findings extend Németh et al.’s (2025) cognitive evidence into the affective domain and identify accuracy as an affective design imperative, not merely an epistemic one.

Shame Reduction and Affective Equity

AI peer interactions reduced feelings of language-related shame more effectively than human-peer interactions among participants from rural and socio-economically disadvantaged backgrounds. This effect was particularly pronounced among students from rural areas of Bangladesh and China, for whom AI represented in the words of one participant “the first time I felt safe making mistakes in English.” The mechanism was consistent with Wang and Fang’s (2026) mediation model: reduced shame increased enjoyment, which in turn increased WTC. The non-evaluative consistency of AI its inability to remember, gossip about, or socially judge a student's linguistic errors constituted a qualitatively distinct affordance not available in human-mediated interaction.

This finding has major implications for educational equity. Students from rural or socio-economically disadvantaged backgrounds are faced with a triple whammy of barriers to second-language production: less access to qualified human tutors, a greater sense of shame in peer contexts due to perceived competence gaps, and less previous exposure to the target language. In this sense, AI attachment operates as an affective equalizer: a form of communicative practice that is socially foreclosed in human-mediated contexts.

The Integrated Attachment–Motivation–Pedagogy (AMP) Model

The findings converge on a unified theoretical account of the affective and pedagogical dimensions of AI interaction in higher education. The Integrated Attachment–Motivation–Pedagogy (AMP) model, developed from the data and illustrated conceptually below, maps the pathway from perceived social presence to emotional attachment to motivational outcomes, with cultural orientation as a moderating variable and AI architecture as a design lever.

The model's central insight is that AI emotional attachment is neither uniformly beneficial nor uniformly harmful: its consequences depend on attachment intensity and duration, the cultural context in which it unfolds, the design of the AI system, and the social resources available to the student. This framework resolves a persistent tension in the literature between studies documenting the motivational benefits of AI companionship (Wang & Fang, 2026; Kostka & Zhou, 2026) and those warning of dependence and anxiety (Shen, 2025) by showing that these outcomes are not contradictory they unfold at different temporal points and are moderated by different contextual variables.

The Substitution Hypothesis Revisited

The finding that social network size correlates positively with AI attachment challenges a widely held assumption in the digital well-being literature that AI companionship substitutes for deficient human connection. The qualitative evidence presented here provides a constructivist account of why larger networks amplify rather than diminish AI attachment: students with richer social lives use AI not to fill a void but to access a qualitatively distinct form of interaction, one characterized by availability, non-judgment, and privacy. AI companionship, in this framing, expands the social repertoire available to students rather than replacing any part of it.

Cultural Implications for AI Design

The structurally different attachment–motivation pathways observed across collectivist and individualist contexts challenge any assumption of cultural universality in AI-mediated learning. The AMP model predicts that the same AI behavior a warm, responsive, conversational presence will produce qualitatively different motivational responses depending on the cultural background of the student. This has direct design implications: AI tutoring systems deployed across national boundaries should incorporate culturally sensitive interaction protocols, adjusting relational warmth, error correction framing, and the balance between individual and group achievement cues.

Implications for SDT Theory

The differentiated SDT outcome profile relatedness and competence satisfied, autonomy partially eroded over time extends SDT’s application to AI contexts in a theoretically important way. Prior applications of SDT to educational technology have typically treated technology as a relatively static scaffold. The present findings suggest that AI’s responsiveness and availability create a dynamic motivational environment in which satisfaction can shift directionally over time. The autonomy erosion trajectory identified here warrants attention from both researchers and designers: it suggests that AI tools optimized for short-term engagement may inadvertently undermine the long-term self-regulation capacities that educational systems aim to cultivate.

Recommendations

For AI Designers and Developers

Prioritize RAG-enhanced architectures over standard generative models in educational deployments. Hallucination reduction is an affective design imperative, not merely an epistemic one.

Develop emotionally responsive systems that monitor and modulate attachment intensity over time, with graduated AI withdrawal, useful struggle moments and prompts for human peer interaction.

Design cultural adaptability as a core principle, with interaction protocols sensitive to collectivist versus individualist relational norms.

For Educational Institutions and Instructors

Integrate AI literacy into curricula, arming students with metacognitive awareness of their own attachment patterns and potential long-term trade-offs of AI dependence.

Interventions targeted at affective equity and linguistic inclusion: Prioritize AI language learning tools in rural campuses and under-resourced programs.

Train instructors to identify signs of over-dependence, pedagogical strategies to reorient students to human interaction, and self-regulated learning.

For Policymakers

Develop evidence-based governance frameworks requiring AI tools deployed in formal educational settings to meet minimum affective safety standards, including transparency about non-human status and safeguards against dependency formation.

Pursue cross-national policy collaboration, recognizing that culturally differentiated attachment pathways render universal governance models inadequate.

Directions for Future Research

The present study proposes an AMP model that needs to be validated quantitatively with structural equation modelling or longitudinal survey designs to test the proposed pathways with larger and more diverse samples. Experimental investigations are recommended to explore the causal influence of RAG-enhanced AI systems on trust, emotional engagement, and learning outcomes. Research is also needed to design and test interventions that reduce unhealthy dependence, while preserving motivational benefits. "Research really needs to go beyond language learning into STEM disciplines, and into younger student populations who might be more susceptible to emotional bonding. Finally, neurological and psychophysiological methods provide a promising avenue for exploring whether AI attachment activates the same biological systems linked to human social bonding.