This concept describes interpretability research into whether neural network language models exhibit an internal distinction analogous to conscious versus unconscious processing, drawing on the global workspace theory of consciousness, which posits that a brain selects a limited set of information into a shared mental workspace that is broadcast for reasoning and control. Researchers identified a set of internal, word-linked activation patterns (termed a "space," derived via the Jacobian, a mathematical tool for measuring how outputs change with respect to inputs) that a model uses for step-by-step reasoning, partial self-directed control of attention, and self-monitoring, distinguishing it from the model's much larger unconscious computational substrate. This belongs to the domain of AI interpretability / mechanistic analysis of neural networks, relating to the parent discipline of cognitive science and consciousness theory through structural analogy, without asserting equivalence to subjective experience.
This concept describes interpretability research into whether neural network language models exhibit an internal distinction analogous to conscious versus unconscious processing, drawing on the globa…