LCAM: A Framework for Diagnosing Interactional Alignment Failures in Con-versational AI

Abstract: Conversational AI is increasingly used for advice, interpretation, reassurance, and decision support in contexts where users may be vulnerable, uncertain, or dependent on the system's apparent competence. Existing alignment work often focuses on model objectives, preference optimization, or output correctness. Yet, many harms arise through interaction: how systems frame authority, express uncertainty, simulate empathy, support reasoning, and make boundaries legible. This paper introduces the Layered Cognitive Alignment Model (LCAM), a conceptual and normative framework for diagnosing interac-tional alignment failures in conversational AI. LCAM defines alignment as a calibrated fit among system behavior, user goals, task demands, and normative context. It distinguishes five layers of fit: perceptual, semantic, affective, cognitive, and ethical, and two diagnostic polarities of misalignment: underfit and overreach. We apply LCAM to a published LLM counseling example, showing how an apparently supportive response can reinforce harmful beliefs, simulate inappropriate care, and obscure role boundaries. By translating conversational failures into audit and governance questions concerning over-reliance, false intimacy, autonomy erosion, boundary confusion, and inappropriate trust, LCAM offers a theoretical and normative lens for evaluating conversational AI beyond accuracy, helpfulness, or trust.

• The paper introduces LCAM as a layered framework to diagnose interactional alignment failures, focusing on both underfit and overreach in conversational contexts.
• It decomposes alignment into five layers—perceptual, semantic, affective, cognitive, and ethical—to pinpoint specific risks in AI interactions.
• The framework bridges technical risk assessment with practical governance for high-stakes domains, urging tailored audits and red-teaming practices.

LCAM: A Framework for Diagnosing Interactional Alignment Failures in Conversational AI

Introduction and Motivation

The increasing deployment of conversational AI in domains such as mental health, education, and financial advice introduces unique alignment hazards not addressed by conventional model- or objective-level analyses. Current alignment paradigms emphasize output correctness or aggregate preference optimization, but many epistemic and affective harms emerge specifically through sustained interactional dynamics—how these systems communicate authority, signal uncertainty, and simulate care. This paper introduces the Layered Cognitive Alignment Model (LCAM) as a theoretical and normative framework explicitly designed to diagnose failures of interactional alignment, centering the analysis on the fit between system behavior, user goals, task demands, and normative context.

Theoretical Foundation and Distinction from Prior Work

LCAM is grounded in an overview of work on model alignment, human-AI interaction, grounding theory, trust calibration, and value-sensitive design. Unlike mono-dimensional optimization schemes (e.g., RLHF approaches), which risk flattening heterogeneous user values into a singular objective, LCAM posits that conversational alignment should be treated as a multidimensional, context-sensitive property encountered temporally within sequences of interaction. The model further recognizes that mere user preference satisfaction or output factuality is insufficient; alignment must be continually recalibrated against users’ immediate needs, domain demands, and normative boundaries, particularly in high-stakes or vulnerable-user contexts.

LCAM’s central innovation is moving beyond isolated conversation turns or system-level metrics by proposing a decomposable, interaction-centered vocabulary that reveals both under-support and overreach as distinct families of alignment failures. This is in contrast to prior frameworks, which tend to treat harms as primarily arising from inattention (underfit) or generic error, failing to taxonomize harm from excess (overreach) such as unwarranted intimacy, authority, or persuasion.

The Layered Cognitive Alignment Model (LCAM)

LCAM operationalizes interactional alignment in conversational AI using five analytically distinct, but practically interdependent, layers, each corresponding to a specific conversational demand and risk profile:

1. Perceptual Alignment: Accessibility and salience of relevant information at the point of reliance. Failures can result in critical caveats or boundaries being practically invisible, even if technically present.
2. Semantic Alignment: Maintenance of shared meaning, including reference, terminology, and framing. Misalignment arises when system interpretations diverge from user intent or context, regardless of fluency.
3. Affective Alignment: Appropriateness of emotional and relational stance. Systems may either invalidate (underfit) or simulate excessive intimacy or empathy (overreach), risking false social attributions, emotional dependency, or distress amplification.
4. Cognitive Alignment: Support for reflective judgment and integration of outputs, designed to preserve user agency without overwhelming, substituting, or covertly steering decision-making.
5. Ethical Alignment: Visibility of system limits, role boundaries, accountability, and legitimacy. Failures at this layer include hidden boundaries, undue authority, misleading role signals, or opacity regarding escalation and professional competence.

LCAM specifies two diagnostic polarities for failure at any layer:

• Underfit: Insufficient coordination—critical support, clarification, or warning is lacking.
• Overreach: Excessive or autonomy-reducing coordination—system simulates or assumes authority, intimacy, or agency beyond warranted bounds.

Demonstration: Analysis of LLM Counseling Exchange

LCAM’s analytical utility is demonstrated through application to a documented LLM counseling dialogue (from Iftikhar et al., 2025). The system, in responding with validating empathy to a user’s distressing interpretation of family rejection, fails on multiple layers: semantically, by not distinguishing subjective interpretation from factual status; cognitively, by affirming a problematic belief rather than enabling reflection or reappraisal; affectively, by simulating intimacy without calibrating to risk or user vulnerability; and ethically, by performing a quasi-therapeutic role absent accompanying competence or accountability.

This multilayered analysis demonstrates that apparent conversational fluency, warmth, or helpfulness can mask complex harms—such as epistemic dependency, autonomy erosion, or boundary confusion—not captured by traditional evaluation metrics like user satisfaction or correctness. Moreover, the diagnosis reveals that overreach may be especially insidious, as it leverages positive interactional signals (e.g., empathy, intimacy) in problematic ways.

Implications for Practice, Audit, and Governance

LCAM bridges the gap between technical risk assessment and the lived reality of high-stakes interactional harms. Its layer-specific vocabulary can inform the following:

• Red-teaming and Pre-deployment Audits: Prompting evaluation not just for factual errors or unsafe content but for hidden caveats (perceptual), ambiguous meanings (semantic), excessive validation (affective), autonomy substitution (cognitive), and illegible boundaries (ethical).
• Post-deployment Incident Analysis: Enabling granular tracing of harm back to interactional misfit, informing both technical corrections and broader accountability processes.
• Domain-specific Reviews: Empowering domain experts to interrogate alignment not just generically but with contextually-relevant questions about role, support, and accountability.
• Governance and Responsible AI Practice: Shifting the paradigm away from static declarations of safety or alignment towards ongoing calibration and monitoring of fit, particularly in domains where user vulnerability, dependency, or role ambiguity is salient.

Limitations and Future Directions

LCAM is currently a conceptual framework; its analytic categories have not been formally operationalized as code-able variables or as the basis for large-scale empirical reliability studies. The nuanced underfit/overreach distinction necessitates expert judgment and could be subject to debate, especially in culturally, institutionally, or domain-specific contexts. The layered design is not ontological—single events may cross multiple layers, requiring careful analysis to distinguish primary versus secondary failure types.

Future work should focus on empirical validation of LCAM through structured audit studies, red-teaming exercises, and human-in-the-loop incident analyses in varied conversational domains. Development of operational audit instruments derived from LCAM could substantially improve the precision and relevance of alignment assessment in complex, real-world deployment contexts.

Conclusion

LCAM advances the analysis of AI alignment failures in conversational systems by introducing a theoretically rigorous, interactionally grounded, and normatively sensitive multi-layer framework. It highlights that alignment is fundamentally a problem of calibrated fit among user needs, system behavior, task demands, and normative constraints—one that cannot be fully captured by output accuracy, user satisfaction, or aggregate preference signals alone. By offering explicit mechanisms to audit for both insufficient and excessive support, LCAM provides an analytic foundation for diagnosing and mitigating harms that emerge only through the temporal, relational, and situated character of human-AI interaction.