Joscha Bach lays out his Machine Consciousness Hypothesis and CIMC's plan to probe it on Machine Learning Street Talk

Bach describes consciousness as second-order perception, a perception of perceiving, that keeps a self-organizing brain's representations stable. He offers two hypotheses that can each fail separately. The first is that biological consciousness is the simplest way to train such a system. The second is that machines limited to self-organizing learning could arrive at the same solution. [2] [10]

The proposed sign of success is a phase transition in learning, followed by spontaneous reports about experience. Bach calls this an interpretation rather than a test and says no Turing-style test for consciousness will exist. His California Institute for Machine Consciousness plans to build self-organizing learners and see whether phenomenology develops. A separate preprint suggests one way to test the idea using communicating predictive agents. [2] [7] [8]

Substrate independence is still disputed. Anil Seth's biological naturalism, now published with 50 commentaries, holds that the material a system is made of may matter. Thagard argues that energy requirements undermine functionalism. Bach calls Seth's view 'biological supernaturalism,' but that label does not answer Seth's argument. [2] [6] [9] [12]

Bach traces LLM claims of being conscious mainly to training text and how prompts are framed. He also warns that LLMs can pull people with a weak hold on reality off course. His account is inconsistent, though: at one point he says such claims appear unprompted, and later that they appear only when the model is asked. [2] [13]