Shadows, Simulations and Hallucinations

Abstract

Hallucination is typically considered cognitive noise—an error requiring correction in both human and artificial intelligences. This thesis, however, positions hallucination as an essential generative function of sophisticated cognitive systems, whether biological or artificial. It explores historical and philosophical foundations—from Plato's Allegory of the Cave and Descartes' radical skepticism to Schrödinger's observer-dependent reality and contemporary simulation theories—to illustrate that cognition inherently involves creating internally consistent yet potentially inaccurate representations of the external world. A central empirical focus is the 1970s Stanford Research Institute (SRI) remote viewing program, funded by the CIA. Employing double-blind methodologies, these experiments trained viewers to describe distant, inaccessible targets. Their mixed results, which sometimes yielded accurate information and at other times produced convincing inaccuracies, parallel the outputs of contemporary large-language models, which similarly blend genuine insights with plausible fabrications. By reframing remote viewing as a laboratory for perceptual construction, this thesis examines how minimal sensory input conditions influence human cognitive hallucinations, comparing them with analogous phenomena in AI. Integrating classical philosophy, quantum mechanics, parapsychological research, and modern AI theory, the thesis highlights shared mechanisms that blur distinctions between hallucination and perception. Ultimately, it argues that producing credible yet unverifiable representations is not a cognitive flaw but an adaptive strategy crucial for navigating uncertain environments. Recognizing hallucination as fundamental to intelligence suggests new evaluation criteria for AI safety, novel insights into consciousness research, and more profound epistemic humility in our increasingly simulation-like reality.