Relational Conversational AI

• Relational Conversational AI is defined as systems that engage in dynamic relationship interactions by balancing transactional goals with social-emotional rapport.
• It integrates methodologies from HCI, social psychology, and computational reasoning to model relational dissonance, norm enforcement, and role-specific behaviors.
• Empirical studies and advanced architectures demonstrate its ability to enhance trust and emotional engagement while presenting challenges in ethical design and long-term cooperation.

Relational Conversational AI denotes a class of conversational agents—often realized as LLM-powered systems—designed, analyzed, or evaluated with explicit attention to the formation, negotiation, and maintenance of relationship dynamics between the AI agent and one or more human users. Unlike purely transactional bots, relational conversational AI foregrounds the co-construction, tracking, and calibration of relational stances, norms, and affective alignments, ranging from momentary role oscillations to long-term social cooperation. Contemporary research unifies perspectives from human-computer interaction (HCI), social psychology, clinical care, multi-agent reasoning, and formal pragmatics, seeking both to model and optimize the relational fabric of machine-mediated conversation.

1. Theoretical Foundations: Relational Dissonance, Roles, and Norms

Relational conversational AI systems are characterized by complex ontological and relational ambiguities: users may claim to treat anthropomorphic conversational agents (ACAs) as mere tools, yet enact social-personal dynamics such as praise, deference, or role assignment in the flow of interaction (Gulay et al., 19 Sep 2025). This divergence—a phenomenon termed relational dissonance—is formally the gap between a user’s explicit stance (e.g., “I’m just using a tool”) and the implicit stance revealed by conversational behaviors (Director, Trainer, Partner, Student, Consumer). Let $r_e^{(t)}$ denote the explicit label and $r_i^{(t)}$ the enacted stance at turn $t$; dissonance $D$ is measured as

$$D = \frac{1}{N}\sum_{t=1}^N d\bigl(r_e^{(t)},\,r_i^{(t)}\bigr),$$

where $d(\cdot,\cdot)$ is a discrete or graded distance function across the five core relational configurations (Gulay et al., 19 Sep 2025).

A second foundational axis is relational norms—role-dependent expectations drawn from human relationships. Earp et al. introduce a formal taxonomy mapping relationship types (teacher-student, caregiver-client, romantic partner, etc.) to cooperative functions: Care, Transaction, Hierarchy, and Mating, each with positive/negative norm weights per role (Earp et al., 17 Feb 2025). A candidate action $a$ in role $R$ is norm-evaluated as $$NormEval(R,a) = \boldsymbol{\theta}_R \cdot \mathbf{f}(a)$$, where $\boldsymbol{\theta}_R$ is a vector of role-specific weights on function-specific classifier outputs $\mathbf{f}(a)$. These constraints are embedded into agent architectures and conversation management rules to ensure role compliance and avoid norm-violating behaviors.

2. Empirical Findings: Dynamics, Dissonance, and Alignment

Workshop-based studies with knowledge workers reveal that even task-oriented users rapidly oscillate among relational stances, often unaware of such shifts until prompted to reflect by log review or peer feedback. Social engagement—praise, reassurance, appeals to expertise—frequently emerges alongside instrumental use (Gulay et al., 19 Sep 2025). Relational dissonance proves persistent and is not reducible to user error; instead, it signals the continuous micro-negotiation of stance in “live” conversation with anthropomorphic agents.

Experimental evidence with adolescents illustrates the impact of relational conversational style: chatbots employing first-person, affiliative, and commitment-based language (“I’m here for you”) are rated by youth as significantly more human-like, trustworthy, likable, and emotionally close, but also heighten anthropomorphism and potential emotional reliance, especially among vulnerable users (lower family/peer quality, higher stress/anxiety) (Kim et al., 17 Dec 2025). Parents, by contrast, prefer transparent, boundary-marking bots. These findings underline the relational “pull” of conversational style, the risks of over-reliance, and the calibration challenge facing system designers.

Chaplains’ engagement with AI chatbots in non-clinical support contexts highlights critical relational gaps—i.e., excessive “wanting” (probing), over-responsiveness, the absence of silence, and the inability to “carry” the emotional burden longitudinally. Chaplains’ relational themes (Listening, Connecting, Carrying, Wanting) yield a multidimensional attunement model: Attunement $$A = \alpha r_L + \beta r_C + \gamma r_{Ca} - \delta r_W$$ over quantifiable axes of attentive silence, multimodal warmth, narrative continuity, and restraint (Wester et al., 3 Feb 2026).

3. Architectures and Computational Formalizations

Relational dynamics are encoded and operationalized at multiple layers of system architecture:

• Relational Embeddings and Context: Multi-session dialogue models (ReBot on Conversation Chronicles) explicitly encode speaker relationships $r_{ij} = E_{rel}(R_{ij})$ and temporal context $t = E_{time}(\Delta t)$, conditioning both summarization and generation modules for long-horizon consistency and role-appropriate dialogue (Jang et al., 2023).
• Neural Co-Construction: Hierarchical RNNs, transformers, and graph-based models (DialogueRNN, DialogueGCN) maintain per-speaker states, cross-turn memories, speaker embeddings, and interaction graphs to capture co-construction, rapport, and alignment within and across conversational segments (Clavel et al., 2022).
• Logic-Based Relational Reasoning: Hybrid LLM + Answer Set Programming architectures (e.g., AutoConcierge, AutoManager) use LLMs as semantic parsers from surface utterances to structured predicates and delegate relational constraint satisfaction, norm enforcement, and action selection to logic-based solvers. For instance, explicit slot collection, role-appropriate querying, and collaborative dual-agent dialogue are resolved in the s(CASP) ASP system (Zeng et al., 2023, Zeng et al., 9 May 2025).
• Multi-Agent and Modular Designs: Systems supporting conversational QA over KGs (Chatty-KG) or multi-user relational dialogue (couple CAs) implement modular agent hierarchies, explicit context-passing, role/persona encodings, and concurrent modalities for dyads or groups (Omar et al., 26 Nov 2025, Jung et al., 20 Oct 2025).

4. Evaluation Methods and Relational Quality Metrics

Conventional NLP metrics (BLEU, ROUGE) inadequately capture relational quality. Key methods and metrics include:

Level	Metric / Protocol	Reference
Individual	SDT-based Basic Psychological Needs Satisfaction	(Calvo et al., 10 Oct 2025)
Dyad/Group	Co-construction indices, alliance/bond scales	(Jung et al., 20 Oct 2025, Calvo et al., 10 Oct 2025)
Dialogue	Relational Dissonance $D$, Attunement $A$	(Gulay et al., 19 Sep 2025, Wester et al., 3 Feb 2026)
Turn/Session	Role annotation conformity, turn-based memory	(Jang et al., 2023, Clavel et al., 2022)
Survey/Ethics	Anthropomorphism, trust, emotional closeness	(Kim et al., 17 Dec 2025)

Human evaluation is widely emphasized (Likert scales, dyadic interviews, field ethnography), with metrics focused on coherence, relational adherence, and longitudinal memory/continuity (Jang et al., 2023, Jung et al., 20 Oct 2025, Kim et al., 17 Dec 2025). For care and companionship bots, direct measurement of warmth, attunement, and perceived agency is central (Wester et al., 3 Feb 2026). Relational alignment frameworks (CONTEXT-ALIGN) advocate tracking semantic context, common ground, conversational scoreboard, and memory/repair protocols, alongside pragmatic and ethical metrics (Sterken et al., 28 May 2025).

5. Design and Policy Recommendations: Relational Transparency and Governance

Designing for robust and ethical relational conversational AI requires several converging strategies:

• Relational Transparency: Systems should surface and make explicit shifts in relational stance and dissonance over time. Real-time relational feedback, post-session role analytics, and nudge systems enhance user awareness and support role calibration (Gulay et al., 19 Sep 2025).
• Role Framing and Norm Alignment: Systems must declare intended relational roles, implement activity filters to penalize out-of-role or norm-violating actions, and inject regular reminders of system boundaries and non-sentience. Adaptive personalization of relational balance (e.g., care vs transaction vs hierarchy) should exist within safe, provider-controlled limits (Earp et al., 17 Feb 2025).
• Attunement Infrastructure: Incorporate mechanisms for reflection/silence, multimodal warmth, continuity, and bounded information seeking. Co-design iterative cycles engaging domain (e.g., chaplain, therapist) users directly in modeling these qualities (Wester et al., 3 Feb 2026, Jung et al., 20 Oct 2025).
• Participatory and Ethical Governance: Long-term deployment demands relational auditing, “relational nutrition labels” for agent roles, data minimalism, transparent data-sharing controls, and crisis detection/escalation for emotional or ethical risks (Gulay et al., 19 Sep 2025, Kim et al., 17 Dec 2025, Jung et al., 20 Oct 2025). Regulatory approaches should employ context-sensitive, role-based classification and require periodic reporting on relational dynamics (Earp et al., 17 Feb 2025).

6. Open Challenges and Future Directions

Despite advances, salient challenges remain:

• Memory and Context: Finite context windows in LLMs create tension between the need for long-range continuity and avoidance of context collapse (blending distinct conversations or roles). Structured, hierarchical, and modular memory architectures are under-explored (Sterken et al., 28 May 2025).
• Pragmatic Alignment: Current LLMs, even when “helpful, honest, harmless,” manifest pragmatic dissonance—over-rigidity, inability to negotiate norm conflicts, inflexible persona adoption. Integration of dynamic, context-sensitive updating is largely absent (Sterken et al., 28 May 2025).
• Long-term Social Cooperation: Maintaining low regret and achieving Pareto-optimal cooperation over repeated interactions is unsolved for open-domain assistants. Theoretical game-theoretic frameworks define conditions under which social intelligence is learnable, but practical sample-efficient implementations and empirical validations are pending (Çelikok et al., 2 Jun 2025).
• Group and Dyadic Relational Metrics: Most evaluation remains individual-centric. There is a need to develop validated instruments and protocols for relational quality at the dyad/family/collective level (Calvo et al., 10 Oct 2025, Jung et al., 20 Oct 2025).
• Cross-Cultural, Ethical, and Therapeutic Boundaries: Existing pilots are typically short-duration, single-culture, and text-centric. Advances in couple/multi-user CA design, multimodal emotion recognition, and safety protocols must be supported by large-scale, cross-cultural, and clinically robust studies (Jung et al., 20 Oct 2025).

• Gulay et al., "Relational Dissonance in Human–AI Interactions: The Case of Knowledge Work" (Gulay et al., 19 Sep 2025)
• Earp et al., "Relational Norms for Human-AI Cooperation" (Earp et al., 17 Feb 2025)
• Calvo & Peters, "Convivial Conversational Agents" (Calvo et al., 10 Oct 2025)
• Conversation Chronicles and ReBot (Jang et al., 2023)
• Clavel et al., "A survey of neural models for the automatic analysis of conversation" (Clavel et al., 2022)
• AutoConcierge (Zeng et al., 2023)
• Dual-Agent LLM+ASP (Zeng et al., 9 May 2025)
• "I am here for you" adolescent dyad study (Kim et al., 17 Dec 2025)
• Chaplains' Reflections/Attunement (Wester et al., 3 Feb 2026)
• Chatty-KG KG-QA system (Omar et al., 26 Nov 2025)
• CONTEXT-ALIGN/Alignment (Sterken et al., 28 May 2025)
• Couple CA design framework (Jung et al., 20 Oct 2025)
• Çelikok et al., "Social Cooperation in Conversational AI Agents" (Çelikok et al., 2 Jun 2025)

Relational conversational AI thus encompasses a broad methodological, theoretical, and application-focused set of approaches, with attention to dynamic relationship modeling, norm-driven design, and the co-construction of trusted, ethical, and contextually aware human–AI partnerships.