Simulated Minds, Artificial Consciousness, and the Theory-Relative Simulation Hypothesis

The simulation hypothesis, which proposes that our world might be a computer-generated environment with simulated minds, is not metaphysically neutral. Standard formulations assume strong substrate-independent computationalism and organizational invariance, thus favoring certain views of the mind. When four leading theories of consciousness—predictive processing, Integrated Information Theory, global workspace theory, and recurrent processing theory—are applied, each yields a distinct theory-relative simulation hypothesis with different substrate demands, architectural requirements, and risks of producing non-conscious "zombie simulations." Attempts to test the hypothesis scientifically, such as through astrophysical signatures or Bayesian modeling, address only narrow implementation choices. The hypothesis is best seen as metaphysics guided by consciousness science, useful for stress-testing theories of consciousness and highlighting ethical and AI-safety stakes.

Study at a glance

Abstract

The simulation hypothesis says that our world might be a computer-generated environment populated by simulated minds. Since Bostrom's [2003] "simulation argument," it has attracted attention in philosophy of mind and, increasingly, in the mind sciences. The simulation hypothesis says that our world might be a computer-generated environment populated by simulated minds, raising questions about artificial consciousness and advanced AI systems. I make three claims. First, standard formulations of the simulation hypothesis quietly assume strong substrate-independent computationalism and something like Chalmers' organizational invariance, and so are not neutral about the metaphysics of mind. Second, plugging four leading theories-predictive processing, Integrated Information Theory, global workspace theory, and recurrent processing theory-into the picture yields a theory-relative simulation hypothesis (TR-SH): there are PP-relative, IIT-relative, GWT-relative, and RPT-relative simulation hypotheses, and they differ sharply in their substrate demands, architectural requirements, and risks of "zombie simulations." Third, current attempts to turn the simulation hypothesis into a standard scientific hypothesis-for example, via astrophysical signatures or Bayesian modelling-only touch narrow implementation choices rather than the hypothesis in full generality. I suggest that the simulation hypothesis is best understood as metaphysics guided and constrained by consciousness science, and that its principal value is to stress-test theories of consciousness and to highlight the ethical and AI-safety stakes of large-scale simulations.