[Paper Review] CoCoMo: Computational Consciousness Modeling for Generative and Ethical AI

Edward Yi Chang|arXiv (Cornell University)|Mar 17, 2023

Ethics and Social Impacts of AI9 citations

TL;DR

The paper proposes the CoCoMo model to embed computational consciousness, ethical guardrails, and emotional intelligence into generative AI using MFQ scheduling, reinforcement learning, and adaptive prompts. It argues this improves fairness, non-maleficence, empathy, and reliability in AI systems.

ABSTRACT

The CoCoMo model proposes a computational solution to the challenge of incorporating ethical and emotional intelligence considerations into AI systems, with the aim of creating AI agents that combine knowledge with compassion. To achieve this goal, CoCoMo prioritizes fairness, beneficence, non-maleficence, empathy, adaptability, transparency, and critical and exploratory thinking abilities. The model employs consciousness modeling, reinforcement learning, and prompt template formulation to support these desired traits. By incorporating ethical and emotional intelligence considerations, a generative AI model can potentially lead to improved fairness, reduced toxicity, and increased reliability.

Motivation & Objective

• Define desired System-2 AI traits including knowledge, fairness, beneficence, non-maleficence, empathy, adaptability, transparency, and critical/exploratory thinking.
• Propose a Computational Consciousness Model (CoCoMo) integrating consciousness theories with practical architecture for AI safety and ethics.
• Describe modules, algorithms, and prompts that enable emotion modeling, ethical guardrails, and task scheduling in generative AI.
• Show how CoCoMo can improve fairness, reduce toxicity, and increase reliability in foundation models.

Proposed method

• Introduce a functionalist view of consciousness and survey theories to justify a computational approach.
• Define CoCoMo with four modules: receptor, unconsciousness, consciousness, and effector.
• Employ a multi-level feedback queue (MFQ) scheduler with interrupts to manage conscious/unconscious task transitions.
• Incorporate emotion modeling, reward-based reinforcement learning, and prompt-template generation to guide behavior.
• Present prompting ensembles and CRIT (Critical Thinking Template) for evaluation and ethical reasoning.
• Provide examples of empathy templates and counterfactual/abductive reasoning prompts to foster creativity with safety.

Experimental results

Research questions

• RQ1 How can computational consciousness be modeled to support ethical and emotional intelligence in generative AI?
• RQ2 What scheduling and reward mechanisms best enable transitions between unconscious and conscious AI states while maintaining safety?
• RQ3 How can prompts and templates enforce empathy, ethical guardrails, and critical thinking in foundation models?
• RQ4 Can counterfactual and abductive reasoning reduce toxicity and hallucinations without sacrificing creativity?

Key findings

• CoCoMo integrates MFQ scheduling with a consciousness-like loop to manage attention and interrupts for novel events.
• Emotion modeling and reward shaping are used to align AI behavior with ethical and empathetic goals.
• CRIT provides a structured method to evaluate document credibility and reason about claims.
• Prompts and templates enable critical thinking and empathetic responses in LLM-based components.
• The framework aims to improve fairness, reduce toxicity, and enhance reliability by embedding ethical considerations within the AI system itself.