📄 Research

Explicit Consciousness as a Safety Primitive

Antonio Di Cecco — Università degli Studi "G. d'Annunzio", School of AI Italia · July 2026

This paper defines "explicit consciousness" as an engineered architectural module in which structured inner speech (inner voices) is generated periodically from conversation state, stored in a persistent, auditable log, and used as input to a cybernetic safety loop. The architecture is grounded in Minsky's Society of Mind, Baars' Global Workspace Theory, Dennett's multiple drafts, Hofstadter's strange loops, and Vygotsky's developmental psychology of inner speech. AmicoAI is presented as a concrete instantiation: a three-tier memory system with inner voices as the mid-term reflective layer, all stored on-device for privacy-by-design.

Key Concepts

Three Voice Types

• Corrective: detects factual errors, misunderstandings, and conversational missteps
• Character Integrity: enforces the agent's MBTI profile, prevents personality drift
• Speculative: anticipates the user's future intentions and emotional trajectory

Three-Tier Memory Architecture

• Short-term: recent messages for immediate conversational context
• Mid-term: inner voice log — the observable, auditable reflection layer
• Long-term: extracted stable facts about user and agent identity

Cybernetic Safety Loop

Safety module S can inspect inner voices before they shape the agent's response. If a voice reveals harmful intent, the module rewrites or suppresses it — intercepting unsafe trajectories at the intention level. This is proven Lyapunov-stable under mild stochastic assumptions.

Foundational References

• Minsky, M. (1986). The Society of Mind. — consciousness as a hierarchy of supervisory agents
• Baars, B.J. (1988). A Cognitive Theory of Consciousness. — global workspace as broadcast mechanism
• Vygotsky, L.S. (1934). Thought and Language. — inner speech as self-regulation tool
• Dennett, D.C. (1991). Consciousness Explained. — multiple drafts model
• Hofstadter, D.R. (1979). Gödel, Escher, Bach. — strange loops as engine of self-reference
• Park et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior.
• Shinn et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning.
• Hubinger et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training.