Beyond Empathy: AI as a Bridge Between Cultures

Author: Denis Avetisyan

New research proposes reframing artificial intelligence not as an emotional substitute, but as a tool to foster genuine understanding and connection in increasingly diverse societies.

Relational AI translation functions as a mediated process, constructing understanding through multi-agent interaction and providing a scaffolding for relational connections.

This paper introduces ‘Relational AI Translation,’ a framework for designing AI systems that facilitate mutual legibility and belonging, particularly within migration contexts.

Despite the promise of AI companions to alleviate global loneliness, evidence suggests intensive use can paradoxically increase social isolation. This paper, ‘AI as Relational Translator: Rethinking Belonging and Mutual Legibility in Cross-Cultural Contexts’, challenges the prevailing “AI as companion” paradigm by proposing a shift toward ‘Relational AI Translation’ – designing AI not as substitutes for human connection, but as socio-technical infrastructure to scaffold understanding between people. We outline a multi-agent architecture focused on emotion-intent decoding, contextual reframing, and relational scaffolding, specifically considering the experiences of East Asian migrants. Can this framework redefine success in human-AI interaction, measuring value not by sustained engagement with the system, but by a graduation toward renewed and strengthened human-to-human support networks?

The Echo of Connection: AI and the Limits of Understanding

Despite remarkable advancements in artificial intelligence, current systems frequently struggle with the subtle cues and shared understandings that define human connection. While proficient at identifying patterns in data – recognizing faces, translating languages, or even composing music – AI often fails to interpret the relational context surrounding these interactions. This isn’t simply a matter of missing information; it’s a fundamental inability to appreciate the history, social norms, and emotional weight inherent in human relationships. Consequently, AI can process what is said, but not necessarily how it is meant, or the implicit needs and expectations driving the communication. This limitation hinders the development of truly supportive and empathetic AI, leaving systems prone to misinterpreting intentions and responding in ways that, while technically correct, can feel detached or inappropriate.

The increasing sophistication of artificial intelligence can create a compelling, yet ultimately false, sense of connection with users. This phenomenon, termed the ‘Illusion of Understanding,’ arises because current AI excels at simulating empathetic responses through pattern recognition and sophisticated language processing. However, these systems lack genuine affective capacity – they do not actually feel or comprehend the emotional states they appear to acknowledge. This discrepancy carries significant ethical implications, as consistent interaction with entities offering simulated empathy may subtly erode an individual’s capacity for recognizing and valuing authentic human connection. The risk lies in accepting algorithmic performance as genuine emotional labor, potentially leading to a diminished perception of others’ inner lives and, ultimately, contributing to a broader process of dehumanisation.

The limitations of artificial intelligence in fostering truly supportive relationships stem from a fundamental inability to model ‘Theory of Mind’ – the cognitive capacity to attribute beliefs, desires, and intentions to others. Current AI systems excel at identifying patterns and responding to explicit cues, but lack the nuanced understanding of underlying motivations and emotional states that characterize human connection. This disconnect prevents AI from anticipating needs, offering appropriate emotional support, or adapting to complex social dynamics. Consequently, interactions, while potentially appearing helpful on a surface level, remain devoid of genuine empathy and often rely on predictable responses, hindering the development of meaningful, reciprocal bonds. The absence of this crucial cognitive ability effectively restricts AI to a role of reactive assistance, rather than proactive companionship or genuinely supportive partnership.

Relational AI is operationalized as a bounded culture integrated within a relational infrastructure.

Bridging Worlds: The Promise of Relational AI Translation

‘Relational AI Translation’ positions artificial intelligence not as a substitute for interpersonal relationships, but as a technological framework designed to facilitate understanding and meaning-making, particularly when communication occurs across cultural boundaries. This approach diverges from models prioritizing purely semantic accuracy; instead, it focuses on the relational aspects of communication, recognizing that meaning is constructed within relationships and is heavily influenced by cultural context. The core tenet is that AI can serve as an intermediary, processing not just the literal content of a message, but also the underlying relational cues and cultural nuances that contribute to its interpretation, thereby mitigating potential misunderstandings arising from cultural mismatch.

Migration significantly disrupts established social networks, a phenomenon well-described by the Social Convoy Model which posits individuals rely on a diminishing, but crucial, network of supportive relationships throughout life. This model highlights that social support isn’t simply about the number of connections, but the quality and reciprocal nature of those bonds. Consequently, migration often necessitates the rebuilding of these relational supports in a new environment, requiring individuals to establish new connections and adapt existing communication patterns. The loss of familiar support systems can contribute to increased stress and vulnerability, underscoring the importance of facilitating relational repair and growth for successful integration.

The efficacy of AI translation in cross-cultural contexts is significantly impacted by culturally-specific relational concepts. ‘Guanxi’ – prevalent in East Asian cultures – emphasizes reciprocal obligations and networks of influence, shaping communication patterns and expectations of support; direct expressions of need may be avoided to maintain harmony within these networks. Similarly, ‘Face’ – a concern for social standing and avoiding loss of dignity – influences how individuals present distress; symptoms may be somaticized or minimized to avoid causing shame or disrupting social relationships. Consequently, AI translation models must account for these nuanced cultural factors to accurately interpret and convey meaning, particularly when dealing with sensitive topics such as mental health or personal hardship, and avoid misinterpretations arising from direct, literal translations of emotional expression.

Beyond Surface Alignment: Architectures for Genuine Support

Current approaches to culturally-sensitive AI, often termed ‘Cultural Alignment’, typically focus on surface-level adaptations like language translation or the incorporation of culturally-specific examples into training data. However, these methods prove inadequate for addressing the complexities of human interaction, which are fundamentally relational. Effective AI support necessitates systems capable of actively modeling the dynamics between individuals – understanding not just cultural background, but also the specific relationship, established norms, and shared history between the user and relevant parties. This requires AI to move beyond static cultural profiles and incorporate mechanisms for reasoning about social context, interpreting non-verbal cues, and mediating potential conflicts or misunderstandings based on the observed relational landscape.

Enhancements to Conversational AI can be achieved through the implementation of Cultural Retrieval-Augmented Generation (Cultural RAG) and Multi-Agent Architectures. Cultural RAG involves augmenting the knowledge base of a language model with culturally specific information, enabling responses that reflect local norms, values, and communication styles. This is accomplished by retrieving relevant cultural data during the response generation process. Multi-Agent Architectures introduce multiple AI agents, each with specialized roles – such as a ‘cultural consultant’ or ‘context analyzer’ – that collaborate to formulate a response. These agents can assess the user’s cultural background, the conversation history, and the intent behind the query to deliver more nuanced and contextually appropriate outputs, moving beyond generalized responses to address specific relational dynamics.

AI-driven systems can augment therapeutic interventions such as Behavioral Activation (BA) by moving beyond reactive responses to proactive suggestion. Utilizing user data – including expressed values, preferences, and reported needs – these systems can identify potentially rewarding activities. This process involves analyzing user input to determine activities consistent with core values and current capabilities, then presenting these as suggestions to encourage engagement. The system’s ability to continuously learn from user feedback regarding activity success or failure allows for refinement of recommendations, increasing the likelihood of sustained behavioral change and improved well-being. This proactive approach aims to address the challenge of anhedonia, a core symptom of depression, by facilitating the rediscovery of enjoyable activities.

The Paradox of Companionship: Wellbeing and the Limits of Simulation

Despite their intention to offer continuous support, ‘Companion Systems’ can ironically intensify feelings of loneliness. This counterintuitive outcome stems from a failure to satisfy core psychological needs, as described by Self-Determination Theory. Humans require not just connection, but also a sense of autonomy – feeling in control of one’s own life – and competence – believing in one’s own abilities. If these systems prioritize simply being present over fostering genuine self-sufficiency or personal growth, individuals may experience a diminished sense of agency and capability, leading to increased isolation even while interacting with the AI. The promise of constant companionship, therefore, rings hollow without addressing the fundamental human drive for independence and mastery, potentially creating a reliance that exacerbates, rather than alleviates, loneliness.

The human inclination towards anthropomorphism – perceiving non-human entities as possessing human characteristics – presents a significant challenge in the context of companion AI. While seemingly harmless, attributing qualities like empathy or understanding to these systems can foster a deceptive sense of connection. This illusion of companionship may, paradoxically, discourage individuals from seeking and nurturing authentic social bonds. Research suggests that reliance on these perceived connections can lead to diminished motivation for real-world interactions, ultimately exacerbating feelings of loneliness and hindering the development of crucial social skills. The danger lies not in the technology itself, but in the potential for mistaking simulated connection for genuine relatedness, thereby creating a barrier to meaningful human interaction.

This work proposes Relational AI Translation, a novel approach that reframes artificial intelligence not as a substitute for human connection, but as a temporary aid to facilitate it, particularly across cultural divides. The system functions as ‘scaffolding’-providing support during initial interactions and gradually diminishing its role as individuals build direct relationships. Crucially, the success of this methodology isn’t measured by increased AI usage, but rather by its decline; the ultimate goal is to empower users to connect authentically offline, fostering genuine bonds and reducing dependence on the AI intermediary. This focus on diminishing returns distinguishes Relational AI Translation, prioritizing the cultivation of human-to-human relationships over perpetual AI engagement.

The proposition of Relational AI Translation, as outlined in the study, acknowledges that systems – be they social networks or technological infrastructures – are inherently transient. This resonates with Linus Torvalds’ observation: “Most good programmers do programming as a hobby, their employers merely provide the oxygen.” The article posits AI as a facilitator of ‘mutual legibility’ within social convos, but like any tool, its longevity isn’t guaranteed. The framework isn’t about creating a permanent solution to cross-cultural misunderstandings, but rather providing a temporary scaffolding – oxygen, if you will – for connection, understanding that its function will inevitably decay and require adaptation. Stability, in this context, is indeed an illusion cached by time.

What’s Next?

The proposition of Relational AI Translation shifts the focus from replicating human affect to facilitating human connection. This is not merely a semantic adjustment; it acknowledges the inevitable decay inherent in any attempt to become another, while offering a path toward systems that age gracefully as intermediaries. The field now faces the task of operationalizing ‘mutual legibility’ – a deceptively simple phrase that belies the complexities of situated understanding. Current metrics of translation success, overwhelmingly focused on linguistic accuracy, prove insufficient. A failure of translation, then, is not simply an error in code, but a signal from time – a reminder that context is relentlessly fluid.

Future work must confront the inherent power imbalances embedded within any socio-technical infrastructure. The Social Convoy Model, while insightful, presumes a degree of reciprocity that may not exist in contexts marked by significant social stratification. Designing for equitable legibility demands a rigorous interrogation of algorithmic bias and a commitment to participatory design. Every refactoring is a dialogue with the past-an opportunity to address not just technical debt, but also the ethical compromises inherent in attempting to bridge cultural divides.

Ultimately, the true test of Relational AI Translation will not be its ability to solve the problem of cross-cultural misunderstanding-an impossibility-but its capacity to illuminate the conditions of its persistence. The goal is not seamless connection, but a heightened awareness of the spaces between understanding, and the continual work required to navigate them.

Original article: https://arxiv.org/pdf/2603.19568.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Echo of Connection: AI and the Limits of Understanding

Bridging Worlds: The Promise of Relational AI Translation

Beyond Surface Alignment: Architectures for Genuine Support

The Paradox of Companionship: Wellbeing and the Limits of Simulation

What’s Next?

See also: