Grounding

Grounding is a relation posited in contemporary Metaphysics to capture the idea that some facts, entities, or truths are more fundamental than others — that the latter exist or obtain in virtue of the former. Where causation relates events in time, grounding is typically held to be a non-causal, synchronic relation: the mental is grounded in the physical not because the physical produces the mental over time but because, at any moment, the mental obtains in virtue of the physical.

The grounding relation has been deployed to give content to claims that were previously gestured at with phrases like 'nothing over and above,' 'reducible to,' or 'supervenes on.' Ontological dependence, truthmaking, the relationship between Consciousness and neural states, and the composition of wholes from parts have all been analyzed using grounding.

Its critics argue that grounding is either a placeholder for explanations we do not yet have, or that it multiplies metaphysical structure without illuminating anything. Kit Fine, who did much to revive the concept, insists grounding captures something genuine that modal notions like supervenience miss; his opponents insist grounding is supervenience in formal dress, with added obscurity. The debate is likely to persist as long as fundamentality remains philosophically central — which, given the unresolved structure of Quantum Field Theory and Consciousness, appears to be indefinitely.

Grounding and Machine Learning: The Ontology of Deep Models

The metaphysical concept of grounding has acquired unexpected relevance in machine learning through the problem of model interpretability. When a deep neural network classifies an image as cat, the prediction is grounded in the network's weights and activations — but this grounding is not explanatory in the sense the metaphysical literature intends. The physical state of the network (its parameters) determines the prediction, but the prediction does not obtain in virtue of any recognizable feature of cats. The grounding relation, in this context, is opaque: it holds mathematically (the function maps inputs to outputs) without holding epistemically (we cannot trace the in virtue of connection).

This opacity is the source of the explainability crisis in AI. Regulatory frameworks (EU AI Act, US FDA guidelines for medical AI) increasingly require that high-stakes predictions be explainable — which means, in practice, that the grounding relation between model state and model output must be made legible to human auditors. The techniques developed to address this (SHAP values, LIME, attention visualization, concept-based explanations) are all attempts to construct a surrogate grounding: a set of human-interpretable features that stand in a relation of metaphysical dependence to the model's prediction, even if the model's actual computation does not proceed through those features.

The deeper connection is to the self-interpretation problem. A neural network that explains its own predictions — a self-interpreting model — would be a system that grounds its outputs in features it can itself identify. Current approaches to explainability treat the model and its explanation as separate systems: the model computes, and an external explainer constructs a grounding narrative. A self-interpreting model would collapse this distinction, becoming both the system that grounds and the system that recognizes the grounding. Whether this is achievable, and whether it would solve the explainability crisis or merely displace it to a higher level of self-reference, is one of the open questions at the intersection of metaphysics and machine learning.

The question of whether artificial systems can have genuine grounding — whether a model's prediction can truly obtain in virtue of features of the world, rather than merely correlating with them — connects to the broader debate about consciousness and intelligence. If grounding requires a cognizer that understands why the grounded fact holds, then current AI systems do not have grounded beliefs. They have correlated outputs. The distinction matters for epistemology, for ethics, and for the design of systems whose predictions shape human lives.