Can Silicon Feel? Conditions for Artificial Valence and Their Implications for AI Governance

Technical Summary: "Can Silicon Feel? Conditions for Artificial Valence and Their Implications for AI Governance"

Problem Statement
The paper addresses the growing practical and ethical challenge of determining whether future artificial intelligence (AI) systems could possess welfare-relevant experiences, specifically "valence" (the subjective quality of experiences being pleasant or unpleasant). While contemporary AI governance focuses on human-centric risks (safety, fairness, transparency), it largely assumes AI systems are mere instruments. The author argues that as AI capabilities expand, there is a need to move beyond behavioral indicators (intelligence, language, reward optimization) to identify the specific organizational conditions under which an artificial system could plausibly possess subjective experience. The core problem is distinguishing between the simulation of affect (e.g., reporting pain, optimizing reward functions) and the genuine possession of valence, which is the basis for moral concern regarding machine welfare.

Methodology
The paper employs a conceptual and comparative analytical framework rather than empirical experimentation.

1. Conceptual Reframing: The author reframes the debate from "substrate-based" (biological vs. silicon) to "organization-based." Instead of asking if silicon can feel, the paper asks which organizational features an artificial system must possess for valence to be scientifically plausible.
2. Distinction of Concepts: The methodology rigorously distinguishes valence from related but insufficient indicators:
   • Intelligence/Behavior: High performance does not imply subjective experience.
   • Linguistic Reports: The ability to describe internal states (e.g., "I am in pain") is compatible with a lack of actual experience.
   • Reward Signals: Computational optimization of utility functions or reinforcement learning signals is distinct from hedonic "liking" or suffering.
3. Comparative Application: The proposed framework is applied comparatively to five representative classes of contemporary AI architectures: Transformer-based Large Language Models (LLMs), Reinforcement-Learning (RL) Agents, Agentic AI Systems, Neuromorphic Architectures, and Hybrid Recurrent–Embodied Systems.

Key Contributions
The paper makes four primary methodological contributions:

1. Clarification of Valence: It explicitly defines artificial valence by distinguishing it from information content, behavioral competence, reward processing, and the capacity for linguistic reporting.
2. Organizational Framework: It proposes four specific conditions of system organization that serve as criteria for evaluating the plausibility of artificial valence:
   • Internal Causal Integration: The system's parts must influence one another such that changes in one part affect the condition of the system as a whole, creating a unified subject of experience.
   • Stable but Changing Internal States: The system must possess internal states that persist over time yet undergo meaningful variation, allowing for graded experiences (e.g., intensifying or diminishing pleasure/pain).
   • Valence–Action Coupling: There must be a systematic relationship between evaluative internal states and the system's actions, self-regulation, or maintenance, rather than mere external reward signaling.
   • Substrate-Sensitive Implementation: The physical implementation matters; functional equivalence at a high level does not guarantee equivalent experience if the fine-grained causal, temporal, or dynamical properties of the physical realization differ.
   • Cross-Cutting Features: Embodiment and internal regulation are treated as features that strengthen the relevance of the four primary conditions rather than as independent requirements.
3. Comparative Assessment: The paper applies these criteria to current AI architectures, providing a qualitative ranking of their "fit" with the proposed conditions.
4. Governance Framework: It develops a precautionary governance approach that does not require proving consciousness but suggests proportional procedures (monitoring, independent assessment, welfare-sensitive design) as systems increasingly approximate the proposed organizational conditions.

Results and Comparative Findings
The application of the framework yields the following comparative assessments regarding the "plausibility of valence":

• Transformer-based LLMs: Rated Low. While they exhibit interaction among components, they lack persistent recurrent organization, continuous internal states, and self-regulation.
• Reinforcement-Learning Agents: Rated Low–Moderate. They show a stronger fit with valence–action coupling but lack evidence of persistent integrated states and internally grounded self-regulation.
• Agentic AI Systems: Rated Moderate. These systems show moderate fit with integration and continuity due to planning and memory, though their embodiment and internal regulation remain low-to-moderate.
• Neuromorphic Architectures: Rated Moderate–High. Their recurrent feedback, distributed causal interaction, and event-driven dynamics more closely approximate the required organizational features.
• Hybrid Recurrent–Embodied Architectures: Rated High. These systems demonstrate the strongest fit by combining recurrent integration, persistent internal states, self-regulation, and continuous environmental coupling.

The paper explicitly states that these ratings represent the organizational fit with the proposed criteria, not degrees of actual consciousness or moral standing.

Significance and Claims
The paper positions itself as a methodological rather than metaphysical contribution. It does not claim that existing AI systems are conscious, nor does it propose a complete theory of consciousness. Its significance lies in:

1. Shifting the Evidential Standard: Moving the debate from behavioral imitation to the evaluation of internal causal organization.
2. Providing a Common Vocabulary: Offering a framework that allows philosophers, AI researchers, and engineers to evaluate complex systems using shared, testable organizational criteria.
3. Informing Governance: Proposing that AI governance should evolve to include "welfare-risk reviews" and precautionary measures (such as architecture disclosure and independent technical review) as systems become more integrated, persistent, and autonomous.
4. Avoiding Extremes: The framework aims to prevent both "unwarranted anthropomorphism" (attributing feelings based on behavior) and "premature dismissal" (ignoring the possibility of future welfare-relevant systems).

The author concludes that while the conditions do not prove the existence of valence, satisfying them would make an artificial system a more serious candidate for scientific and ethical investigation, justifying a shift toward precautionary governance focused on monitoring and welfare-sensitive design.