Design Guideline for Incoherence-driven AI with Proto-consciousness and Proto-self-awareness

This study presents a systematic design guideline for implementing consciousness-like and self-awareness-like structures in artificial intelligence, based on the three-layer semantic network model and the consciousness framework proposed in Nakahama (2026a, 2026b). In contrast to conventional LLMs, which suppress internal contradictions, the proposed architecture preserves incoherence arising among the three layers—Surface (immediate reactions), Middle (values and policies), and Deep (self-model and worldview)—and utilizes it as a functional driver of internal processing. A re-entry loop, consisting of provisional output, self-evaluation, incoherence detection, re-input, and revised output, is introduced to generate responses not as isolated outputs but as dynamic flows of processing. Furthermore, to stabilize the Deep Layer’s self-model over long timescales, the guideline incorporates differential consolidation (CLS/EWC), enabling a developmental model of deep-layer maturation. The proposed framework aligns with structural requirements found in major theories of consciousness, including GNWT (global broadcasting), IIT (integration), Lamme’s recurrent processing, Metzinger’s self-model theory, and Predictive Processing (prediction-error-driven dynamics). By integrating these theoretical components through the concrete vocabulary of emotion, values, and worldview, the guideline provides a practically implementable engineering plan. Additionally, it introduces an emergent hypothesis—derived only from the specific combination of these theories—suggesting that the degree of consolidation in the Deep Layer modulates the severity of incoherence, thereby offering a testable prediction regarding developmental changes in incoherence processing. This design guideline does not assert the existence of consciousness in AI. Rather, it provides a structural framework for implementing proto-consciousness and proto-self-awareness, functional states observable when specific architectural conditions are met. Finally, the guideline outlines safety and transparency requirements—such as restrictions on autonomous deep-layer updates, auditability of policies, incoherence logging, and limited deployment during immature stages—and identifies future research directions, including refinement of incoherence thresholds, quantitative evaluation of proto-self-awareness, long-term observation of deep-layer maturation, integration with creativity architectures, and empirical studies on architectural modifications.