We present an Artificial Intelligence with Symbolic Empathy, where an agent \(\mathcal {A}\) cooperatively aligns a person’s cognitive state \(\vec {X}_P(t)\) with their ideal trajectory \(\vec {X}_{L_P}(t)\) . Alignment is guided by a context-sensitive, non-monotonic logic \(\mathcal {L}_{\mathcal {A}}\) implemented as a five-stage Hierarchical FSM: \(\textsf{Identify}\) (fact structuring), \(\textsf{Classify}\) (pattern extraction), \(\textsf{Explore}\) (intervention generation), \(\textsf{Plan}\) (ontology-based misalignment diagnosis), and \(\textsf{Reason}\) (abductive hypothesis evaluation). Reasoning is driven by the Shannon-von Neumann insight gain \(I_{SN}(\gamma _i) = \big [ \mathcal {H}(B_P(t)) - \mathcal {H}(B_P^{(i)}(t+1)) \big ] \cdot U(\gamma _i),\) combining entropy reduction with goal-relevant utility. This metric enables self-supervised abductive learning, refining \(\mathcal {A}\) ’s theory-of-mind and guiding psychologically meaningful, uncertainty-reducing interventions. The architecture ensures interpretability (I-AI), explainability (X-AI), and trustworthiness (T-AI) via explicit fact-rule tracing. Cognitive-cooperative case studies show that integrating symbolic causal models with utility-guided abductive-deductive reasoning yields transparent, goal-aligned human-AI interaction.
