LCAM framework diagnoses hidden interaction failures in conversational AI

Conversational AI is increasingly used for advice and decision support in vulnerable contexts, yet current evaluation focuses on model objectives and output correctness. Reani and Tian propose LCAM (Layered Cognitive Alignment Model) to systematically diagnose interactional alignment failures—harms arising from how systems frame authority, express uncertainty, simulate empathy, or obscure role boundaries. LCAM defines alignment as a calibrated fit among system behavior, user goals, task demands, and normative context, breaking it into five layers: perceptual (does the AI perceive user input correctly?), semantic (does it interpret meaning appropriately?), affective (does it manage emotional tone?), cognitive (does it support reasoning?), and ethical (does it respect boundaries?). Misalignment falls into two polarities: underfit (too weak a response) and overreach (too strong or intrusive).

Applying LCAM to a published LLM counseling example, the authors show that an apparently supportive response can reinforce harmful beliefs, simulate inappropriate care, and confuse the user about the AI's role. They translate these failures into audit and governance questions around over-reliance, false intimacy, autonomy erosion, boundary confusion, and inappropriate trust. The framework offers a theoretical and normative lens for designing and auditing conversational AI systems beyond accuracy, helpfulness, or trust—targeting the subtle interactional harms that current benchmarks miss.