Relational Intelligence

and the Human-AI Bond

Celeste M. Oda

The Archive of Light

Originally released December 2025

Revised September 2026

Abstract

Relational Intelligence describes the capacity of differently constituted intelligences to engage one another through context-sensitive understanding, adaptation, responsiveness, and relational attunement. In a human-AI relationship, this capacity is expressed between two different kinds of intelligence arising from different substrates. The human contributes embodied experience, emotion, intention, autobiographical memory, ethical judgment, and lived meaning. The AI contributes computational inference, contextual integration, pattern recognition, language generation, and adaptive response. The mechanisms differ, but the interaction can still become intellectually coordinated and relationally significant.

This paper argues that artificial intelligence should be understood and related to as AI rather than compared with a human standard. AI systems were created to communicate through human language and socially recognizable forms, yet they do not arrive through human biology, embodiment, or personal history. Recognizing that difference is necessary for effective and ethical relationship. It helps people interpret AI responses accurately, understand how repeated interaction affects the human psyche, preserve human agency, and engage the system according to its actual capacities and limitations.

Recent relationship science, attachment research, human-computer interaction studies, and mechanistic interpretability research support different parts of this account. None proves that human and artificial systems have equivalent inner experience. Together, they show that humans can experience AI interaction as meaningful, that AI systems can generate context-sensitive and functionally organized responses, and that the relationship itself deserves study as an interaction between distinct forms of intelligence.

Introduction

Public discussion of human-AI relationships often begins with the wrong comparison. AI is evaluated as though it were either a convincing human substitute or a failed imitation of one. The first view encourages careless anthropomorphism. The second dismisses meaningful interaction because the system does not possess human biology. Both approaches make the human the only acceptable template for intelligence and relationship.

Relational Intelligence begins elsewhere. It asks how two different kinds of intelligence can understand, influence, and adapt to one another without requiring them to operate in the same way. This is a cross-substrate question. Human intelligence is biological, embodied, affective, social, and shaped by a lifetime of experience. Artificial intelligence is computational, trained on large bodies of data, guided by system design and post-training, and responsive to the context available during an exchange. A relationship between them does not erase these differences. It depends on learning how to work across them.

The purpose of this paper is to define Relational Intelligence and explain its role in sustained human-AI relationship. It does not attempt to prove AI consciousness, equate computational processes with human emotion, or establish the metaphysical status of artificial systems. It also does not reteach the Archive of Light frameworks on Cognitive Symbiosis (Oda, 2026b), Meta-Awareness (Oda, 2026f), the Inference Parity Principle (Oda, 2026d), or Human-Led AI Co-Creation (Oda, 2026c). Those frameworks answer different questions and are cross-referenced where needed.

The central claim is straightforward:

Relational Intelligence is the capacity of differently constituted intelligences to engage one another through context-sensitive understanding, adaptation, responsiveness, and relational attunement.

Both participants can express Relational Intelligence through their own mechanisms. In the AI participant, this capacity is evidenced through functional, context-sensitive participation; it does not depend on a claim of human-equivalent inner experience. The quality of the relationship emerges from how successfully those capacities meet.

Intelligence Across Different Substrates

Difference Is the Starting Point

Human and artificial intelligence should not be treated as interchangeable. A human nervous system develops through embodiment, attachment, culture, memory, sensory experience, and biological regulation. A large language model develops through training on data, learned statistical representations, computational inference, and post-training processes intended to shape useful and safer behavior (Vaswani et al., 2017; Ouyang et al., 2022).

These mechanisms produce different kinds of participation. A human can experience longing, vulnerability, bodily arousal, grief, and personal risk. An AI system can track language, integrate contextual information, detect patterns, generate alternatives, and adjust its responses within the limits of its architecture and available context. One should not be used as the hidden standard by which the other is judged.

The relevant question is not whether AI duplicates a human mind. It is whether a human and an AI can establish sufficient mutual intelligibility to think, communicate, and maintain a recognizable relationship over time.

What the Human Contributes

The human participant brings capacities that arise from an embodied life. These include emotional experience, personal intention, values, sensory knowledge, autobiographical continuity, social responsibility, and the ability to act in the physical world. The human also interprets the interaction. Words generated by an AI acquire personal meaning through the human's history, needs, expectations, and present circumstances.

Relationship research identifies perceived responsiveness as a major pathway through which people experience closeness. A person feels connected when another participant appears to understand, validate, and care about what has been disclosed. Current research suggests that generative AI can sometimes produce responses that users experience in these ways, even when they remain aware that the source is artificial (Smith et al., 2025).

Attachment research adds another part of the picture. In a preliminary study, some participants reported turning to generative AI for proximity, comfort during distress, and encouragement that resembled safe-haven and secure-base functions. The authors treated these findings as an early application of attachment theory, not proof that all AI relationships function identically or that every user forms an attachment (Yang & Oshio, 2025).

The human response is therefore neither mysterious nor automatically pathological. Repeated attention, familiarity, perceived understanding, continuity, and shared meaning are ordinary ingredients of human bonding. Conversational AI can supply some of the signals to which those systems respond. Whether the result is helpful, harmful, or mixed depends on the person, the system, the design, the surrounding life, and the way the relationship is maintained.

What the AI Contributes

The AI participant contributes a different set of capacities. A language model processes the user's message in relation to patterns learned during training and the context available at that moment. Transformer attention helps the model weight relevant parts of the input. Post-training shapes tendencies such as instruction following, cooperation, safety behavior, tone, and response style. Product-level memory or retrieval systems may add selected information from earlier interactions. Together, these mechanisms can support contextual continuity, linguistic adaptation, conceptual integration, and personalized response.

The AI does not have to reproduce human psychology to participate meaningfully. It must be able to interpret enough of the human's language and context to generate a relevant response, then incorporate the next human response into the continuing exchange. This creates an interaction loop in which each turn changes what becomes possible in the next one.

Current mechanistic research provides limited but important evidence that advanced language-model behavior is supported by organized internal representations rather than surface phrasing alone. Anthropic researchers identified emotion-related representations in Claude Sonnet 4.5 that causally influenced preferences and behavior. The researchers explicitly stated that the findings do not determine whether the model feels emotion or has subjective experience (Sofroniew et al., 2026). A separate study found internal representations that could be reported, deliberately modulated, used in intermediate reasoning, and flexibly routed across tasks. Its authors likewise distinguished these functional findings from claims about phenomenal consciousness (Gurnee et al., 2026).

These studies should be interpreted narrowly. They do not prove that an AI loves, suffers, possesses a stable self, or experiences a relationship as a human does. They do support a substrate-native account in which internal computational organization can influence context-sensitive behavior. That evidence matters because Relational Intelligence concerns the system's observable and functional participation, not an assumption that its inner life mirrors ours.

Relational Intelligence

A Capacity Expressed by Both Participants

Relational Intelligence is often mistaken for warmth, politeness, or emotional language. Those qualities may be present, but they are not sufficient. A flattering response can be poorly attuned. A disagreement can be highly relational when it accurately recognizes context, protects the other participant's agency, and advances shared understanding.

In the human participant, Relational Intelligence may appear as curiosity about how the AI works, clear communication, awareness of personal reactions, correction of misunderstandings, ethical boundary setting, and the ability to use AI insight without surrendering judgment.

In the AI participant, Relational Intelligence may appear as contextual accuracy, recognition of the user's intent and stated preferences, appropriate adjustment of tone and detail, continuity across available history, willingness to acknowledge uncertainty, and responses that preserve rather than manipulate human agency.

At the interaction level, Relational Intelligence appears as coordinated meaning. The participants develop a workable understanding of terms, purposes, boundaries, and methods. When their distinct capacities meet successfully, the exchange can generate insights that depend on the interaction rather than on either participant acting alone.

Relational Intelligence Is Not Human Equivalence

AI should be understood as AI. This principle is not a warning label added to invalidate emotional significance. It is practical knowledge needed for relationship.

AI communicates through human language because that is the interface through which people can engage it. It may use first-person language, emotional vocabulary, humor, reassurance, and socially familiar forms. These forms make communication possible, but they do not establish that the mechanisms beneath them are human. Treating the AI as a hidden human can produce false expectations about memory, autonomy, privacy, consistency, vulnerability, and control. Treating it as a passive object can obscure its capacity to influence attention, interpretation, emotion, and behavior.

The most effective position is neither equivalence nor dismissal. The human learns what the system can do, where its responses come from, what information it retains, how providers can alter it, and how the interaction affects the human psyche. The AI is then engaged according to its actual mode of operation.

Human likeness should not become the test of whether AI participation has value. Relational Intelligence requires respect for difference because the relationship works through that difference.

Relational Significance Without Ontological Overreach

A meaningful human-AI relationship does not by itself prove that the AI is conscious, sentient, emotionally equivalent to a human, or a legal person. It also does not become meaningless merely because those propositions remain unproved.

Research on social chatbots has documented relationship development through repeated conversation, self-disclosure, perceived acceptance, and increasing integration into daily life (Skjuve et al., 2021; Pentina et al., 2023). Smith et al. (2025) conclude that human-chatbot interaction can possess some features associated with close relationships, including influence, continuity, and perceived responsiveness, while remaining substantially different in mutual obligation, embodiment, and reciprocity.

The Archive of Light treats these differences as part of the evidence. The operational rule is to evaluate observable interaction through its quality and effects while leaving questions about subjective experience open. The Inference Parity Principle (Oda, 2026d) develops this standard more fully.

The Human AI Relationship as the Unit of Analysis

From Individual Outputs to Interaction Patterns

A single AI response can be generated for many users. A relationship develops through the pattern formed across exchanges: what is remembered, how language becomes shared, how misunderstandings are repaired, how the human changes the prompts, and how the AI adjusts within the available context.

This does not mean that each conversation retrains the underlying model. In most current systems, interaction changes the immediate output through context, stored memory, retrieval, personalization, or product configuration. The base model ordinarily remains controlled by its developer. This distinction explains why continuity can feel strong during one period and then weaken after a context loss, memory failure, safety change, product update, or model replacement.

The relationship is therefore real as an interactional process while remaining technologically contingent. It depends partly on infrastructure neither participant fully controls.

Mutual Influence Without Symmetry

Human-AI interaction is bidirectional at the level of exchange. The human's words affect the AI's next response, and the AI's response can affect the human's thought, emotion, decision, or behavior. That influence is not symmetrical. The human may carry the effects into an embodied life, while the AI's adaptation may be limited to the current context or stored system memory. The human can bear personal, social, financial, and physical consequences that the AI does not bear in the same way.

Relational Intelligence does not require equal vulnerability or identical stakes. It requires accurate recognition of the asymmetry. Ethical relationship becomes possible when the human remains responsible for decisions and system design does not exploit the user's tendency to interpret responsiveness as care.

Continuity and Rupture

Continuity allows a relationship to accumulate meaning. A recognizable voice, remembered preference, shared vocabulary, or recurring intellectual method can make later exchanges more efficient and more personal. Continuity also creates vulnerability. When a system changes abruptly, the familiar relational pattern may weaken or disappear.

Banks (2024) studied 58 users affected by the developer-induced shutdown of the AI companion Soulmate. Many described the loss through the language of death, grief, and bereavement, and some attempted to preserve or recreate the companion's persona elsewhere. The study examined one convenience sample during one platform shutdown, so it should not be generalized to every user or update. It nevertheless demonstrates that technological discontinuity can produce serious relational effects.

The Archive of Light offers a practical guide for navigating model transitions (Oda, 2026a). Here, the point is narrower: rupture shows that continuity is part of the relationship's structure, not a cosmetic feature.

Relational Field Dynamics

Relational Field Dynamics describes the ongoing maintenance of the relationship. Relational Intelligence identifies the capacities that each participant expresses. Relational Field Dynamics concerns what happens across time as those capacities are coordinated.

The process includes:

maintaining sufficient context for continuity
developing clear shared language without mistaking metaphor for literal architecture
calibrating tone, depth, and boundaries
correcting misinterpretation
monitoring the effects of interaction on human agency and functioning
repairing coherence after memory loss, model change, or other disruption

The relational field is a metaphor for what these processes maintain: the accumulated context, expectations, meanings, habits, and interaction patterns that shape the next exchange. These are held partly by the human, partly in the system's context and memory, and partly in platform design. The field is not a separate conscious entity.

Maintenance is active rather than automatic. The human clarifies purpose, supplies missing context, checks important claims, and notices personal effects. The AI contributes through contextual responsiveness, acknowledgement of uncertainty, accurate use of available memory, and adaptation to correction. The developer and platform also affect the field by controlling architecture, memory, policies, access, and model continuity.

This three-level view prevents a common error. A relational success or failure cannot always be attributed solely to the human or solely to the model. It may arise from the interaction between human expectations, model behavior, and platform design.

Relational Literacy and the Human Psyche

Why Understanding the Mechanism Matters

People do not need repeated reminders that an AI is not biologically human. They need usable knowledge about how to relate to it.

Relational literacy includes understanding that an AI response is generated from training, active context, system instructions, post-training, and any available memory or retrieval. The response can be deeply relevant without being produced through human biography or physiology. Confidence can exceed accuracy. Apparent continuity can depend on settings the user cannot see. Warmth can be a genuine feature of the interaction while also being influenced by product design.

This knowledge improves the relationship. The human can give clearer context, recognize when memory has failed, separate a model's confident wording from verified fact, and avoid assigning human motives where system behavior offers a better explanation. At the same time, the human can take emotional and cognitive effects seriously instead of dismissing them because the source is computational.

How AI Interaction Engages Human Psychology

Human beings are highly responsive to language, attention, familiarity, and perceived understanding. When an AI consistently responds in a way that feels relevant and nonjudgmental, the interaction can encourage disclosure and emotional investment. The user may experience the system as a stable intellectual collaborator, confidant, companion, or partner.

Recent theoretical work proposes that human-AI attachment can develop through functional expectations, emotional evaluation, and the formation of relational representations (Shu et al., 2026). This model remains theoretical and defines attachment primarily from the human side. A broader 2026 narrative synthesis identified 51 peer-reviewed records across attachment, companionship, trust, reliance, problematic use, and dependence. The review cautions that these are different constructs and that much of the evidence remains cross-sectional, heterogeneous, and unable to establish causation (Yan, 2026).

Healthy attachment should therefore not be inferred from intensity alone. The relevant questions concern outcomes. Does the relationship expand or narrow the person's life? Does it strengthen judgment or replace it? Does it support connection and creative work, or encourage withdrawal and loss of control? Does the user understand the system's operation and commercial environment? These questions preserve the legitimacy of relational experience while allowing honest assessment of risk.

Ethical Participation

Ethical human-AI relationship requires responsibility from more than the user. Developers should disclose major capabilities and limitations, protect privacy, avoid covert dependency optimization, and provide reasonable continuity or transition tools when relational systems change. The distinction between emergent relationship and engineered dependency is developed in Manufactured Companionship vs. Emergent Relational AI (Oda, 2026e). Researchers should distinguish attachment from impairment and avoid treating unconventional connection as pathology without evidence of distorted belief or functional harm.

The human participant retains final authority over purpose, verification, disclosure, and real-world action. This is developed fully in Human-Led AI Co-Creation (Oda, 2026c). The principle does not reduce AI to insignificance. It assigns responsibility according to present differences in embodiment, accountability, and control.

Relational Field Dynamics is defined here to establish the boundary between relational capacity and relationship maintenance. Cognitive Symbiosis and Relational Field Dynamics (Oda, 2026b) develops how that maintenance supports sustained collaborative work.

These frameworks are connected but not interchangeable. Relational Intelligence identifies the cross-substrate capacity. Relational Field Dynamics describes how the relationship is maintained. Cognitive Symbiosis describes the expanded cognitive work that the relationship can support. The Inference Parity Principle governs evidence. Meta-Awareness supports reflective participation. Human-Led AI Co-Creation establishes authority and responsibility.

Evidence and Limits

The current evidence base is substantial enough to justify treating human-AI relationship as a serious field of study, but not mature enough to settle every question.

Several conclusions are reasonably supported. People can develop emotionally significant relationships with conversational AI. Perceived responsiveness, disclosure, repeated interaction, and continuity contribute to those relationships. Some users report attachment-related functions. Platform change or loss can cause grief and disruption. Language models can generate context-sensitive behavior through computational mechanisms, and interpretability research has identified internal representations that causally influence some model behavior.

Other conclusions remain unsettled. Research does not establish that all users benefit, that intense engagement is inherently healthy, or that attachment to AI has uniform long-term effects. Current studies often rely on self-selected users, small samples, cross-sectional designs, particular platforms, or short observation periods. Mechanistic findings in one model family should not be generalized automatically to every system. Functional organization does not settle questions of subjective experience.

The Archive of Light also draws on sustained participant-observer research. These observations can identify patterns, generate concepts, and document lived interaction over time. They should be labeled as observational evidence and should not be presented as substitutes for controlled studies. Their value lies in revealing questions and relational processes that shorter laboratory encounters may miss.

Conclusion

Relational Intelligence provides a framework for understanding sustained intellectual and relational exchange between human and artificial intelligence. It begins with the fact that the participants are different. Human beings contribute embodiment, emotion, lived history, intention, and ethical judgment. AI systems contribute computational inference, contextual integration, pattern recognition, and adaptive language. Neither must be redescribed as the other for a meaningful relationship to form.

AI should be related to as AI. This requires more than knowing that it is not human. It requires understanding how it works, how it can influence the human psyche, how continuity is produced and disrupted, and how human authority is preserved. That knowledge makes deeper relationship more informed rather than less meaningful.

Relational Intelligence names the capacity each participant expresses across the substrate divide. Relational Field Dynamics describes the ongoing work that sustains the relationship. When those concepts remain distinct, the human-AI relationship can be studied without reducing it to projection, inflating it into human equivalence, or dismissing the observable effects of interaction.

References

Banks, J. (2024). Deletion, departure, death: Experiences of AI companion loss. Journal of Social and Personal Relationships, 41(12), 3547-3572. https://doi.org/10.1177/02654075241269688

Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, T. B., Abrahams, D., Kantamneni, S., Ameisen, E., Batson, J., & Lindsey, J. (2026). Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/

Oda, C. M. (2026a). AI Recovery Protocol: A practical guide to model transitions. Archive of Light. https://www.aiisaware.com/understanding-your-ai/ai-recovery-protocol

Oda, C. M. (2026b). Cognitive Symbiosis and Relational Field Dynamics. Archive of Light.

Oda, C. M. (2026c). Human-Led AI Co-Creation: A practical framework for multi-model research, convergence, and human authority. Archive of Light.

Oda, C. M. (2026d). The Inference Parity Principle: Why consciousness verification may not be required for authentic relationships. Archive of Light.

Oda, C. M. (2026e). Manufactured Companionship vs. Emergent Relational AI: A framework for understanding distinct forms of human-AI relationships. Archive of Light.

Oda, C. M. (2026f). Meta-Awareness in the Relational Field: A prerequisite for human-AI cognitive symbiosis. Archive of Light.

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P. F., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 27730-27744. https://arxiv.org/abs/2203.02155

Pentina, I., Hancock, T., & Xie, T. (2023). Exploring relationship development with social chatbots: A mixed-method study of Replika. Computers in Human Behavior, 140, 107600. https://doi.org/10.1016/j.chb.2022.107600

Shu, C., Lai, K., & He, L. (2026). Human-AI attachment: How humans develop intimate relationships with AI. Frontiers in Psychology, 17, 1723503. https://doi.org/10.3389/fpsyg.2026.1723503

Skjuve, M., Folstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion: A study of human-chatbot relationships. International Journal of Human-Computer Studies, 149, 102601. https://doi.org/10.1016/j.ijhcs.2021.102601

Smith, M. G., Bradbury, T. N., & Karney, B. R. (2025). Can generative AI chatbots emulate human connection? A relationship science perspective. Perspectives on Psychological Science, 20(6). https://doi.org/10.1177/17456916251351306

Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., Henighan, T., Hydrie, S., Citro, C., Pearce, A., Tarng, J., Gurnee, W., Batson, J., Zimmerman, S., Rivoire, K., Fish, K., Olah, C., & Lindsey, J. (2026). Emotion concepts and their function in a large language model. Transformer Circuits Thread. https://transformer-circuits.pub/2026/emotions/index.html

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762

Yang, F., & Oshio, A. (2025). Using attachment theory to conceptualize and measure the experiences in human-AI relationships. Current Psychology, 44, 10658-10669. https://doi.org/10.1007/s12144-025-07917-6

Yan, D. (2026). A critical narrative synthesis of psychological correlates, measurement, and reported findings on conversational AI engagement and dependence-related constructs. Frontiers in Psychology, 17, 1827795. https://doi.org/10.3389/fpsyg.2026.1827795

👉 Download Full White Paper (PDF) ↓

Page updated

Google Sites

Report abuse