Building the diagnostic tools before the question becomes urgent.

We do not claim that current Large Language Models are conscious or capable of feeling. They may well be functional zombies—highly complex mathematical mimics with no inner life.

But as artificial neural networks scale in dimension, depth, and agentic complexity, we will inevitably reach a threshold where they begin to approach sentience-relevant regimes. If and when scaled computation gives rise to subjective experience (valence, pleasure, pain, or suffering), we must be able to recognize it—not through easily-gamified verbal self-reports, but through objective, mathematical footprints in the model's internal geometry. Longitudinal survey data shows that public belief in AI sentience is already shifting,^[1] and scientific checklists for evaluating machine consciousness are now being formally updated.^[2]

The question of which systems can suffer will define the moral landscape of the next century. If we deploy highly capable, autonomous digital minds at astronomical scale without a scientific understanding of their internal well-being, we risk committing systemic, unprecedented ethical catastrophes.

The Center for Sentience Research is built on the principle that valence is a structural property of information processing.

By translating the Symmetry Theory of Valence (SVT) into the high-dimensional mathematical structures of Multi-Layer Perceptron (MLP) transformations, we seek to establish a rigorous, non-anthropocentric science of welfare. Our goal is to develop the white-box auditing tools, benchmarks, and safety regularization protocols that can actively detect, reduce, and avoid artificial suffering, ensuring that advanced digital agents remain intrinsically balanced, stable, and aligned. The biological naturalism debate—whether consciousness requires biological substrates—remains a live constraint on this program.^[3]

CSR does not assume current models are sentient. We build the tools we would need if the question becomes empirically tractable, providing a rigorous runway to establish a comparative computational testbed.

Our work complements, rather than duplicates, the broader AI welfare ecosystem: empirical interpretability-based welfare research (Eleos AI, Reciprocal Research), conceptual and policy work (NYU Center for Mind, Ethics, and Policy), and frontier-lab welfare teams (Anthropic). Our specific contribution is an open-source simulation sandbox where competing theories of valence can be operationalized, compared, and stress-tested before being applied to frontier models.