AI Architectures as Variable Epistemic Conditions: Generating Descriptive Spaces in Which New Scientific Questions Become Formulable

Artificial intelligence is increasingly used in computational neuroscience for prediction, classification, representation learning, parameter estimation, and related analytical tasks. In most applications, however, the scientific question and the target quantities are specified in advance. Different model architectures may be optimized or compared, but they are primarily evaluated by how well they answer the same predefined question rather than by what different scientific descriptions and questions their architectural differences might make possible. We propose a different use of AI: to treat architectural variation itself as an experimental epistemic variable. Different AI architectures embody different inductive biases and therefore make different temporal, relational, topological, geometrical, symbolic, or dynamical structures accessible in the same empirical data. Rather than asking primarily which architecture performs best on a predefined task, we propose deliberately applying structurally different model families to the same data while minimizing prior commitments about which explanatory quantities should matter — for example entropy, dwell time, recovery time, intrinsic neural timescales, or network connectivity. Each architecture can thereby generate a distinct descriptive space. The primary scientific opportunity lies in architecture-specific discovery. A representation may expose a structure and thereby make a scientific question formulable that remains inaccessible within another representation. Cross-architecture convergence provides a second epistemic opportunity: if substantially different descriptive spaces repeatedly reveal translatable versions of the same phenomenon, this recurrence makes that phenomenon a particularly strong candidate for further investigation because it may reflect a more fundamental structure of the empirical system rather than merely the representational preferences of one architecture. Recent work on schizophrenia provides a concrete motivating case. Spatiotemporal analyses have identified temporal imprecision across phase, intrinsic neural timescales, and frequency domains, while complementary state-transition analyses have revealed impaired recovery dynamics despite similar entropy rates between patients and controls. These findings illustrate how changing the mathematical description can shift attention from one plausible explanatory quantity — overall transition entropy — toward structurally different quantities such as spectral gap, mixing time, and recovery dynamics. Schizophrenia serves here as a case study rather than as the intended boundary of the method. The proposed framework is domain-general: wherever substantially different computational architectures can be applied to the same empirical or formal system, architectural variation can in principle be used to explore alternative descriptive spaces. Illustrative model families include recurrent and state-space models, transformers, temporal graph neural networks, latent-dynamics models, and equation-discovery approaches. This list is deliberately non-exhaustive. The framework explicitly allows additional and even initially non-obvious architectures whose inductive biases may open descriptive spaces not represented by these examples. The aim is not primarily to identify the best-performing architecture, but to exploit architectural difference as a method of scientific discovery. We propose shifting AI in science from a tool for answering predefined questions toward a method for systematically generating alternative descriptive spaces in which new scientific questions can become formulable. AI would thereby be used not only to test hypotheses, but to vary some of the conditions under which hypotheses become visible in the first place.