arXiv (Cornell University)
April 18, 2024
Michele Farisco, Kathinka Evers, Jean-Pierre Changeux
From an evolutionary perspective, several structural and functional features of the human brain appear key for human-like conscious experience. Current AI research is limited in emulating human consciousness for both intrinsic (structural and architectural) and extrinsic (current scientific and technological knowledge) reasons, but taking inspiration from brain characteristics that enable conscious processing is a promising strategy. It is theoretically possible that AI could develop partial or alternative forms of consciousness qualitatively different from human, possibly more or less sophisticated. The authors recommend neuroscience-inspired caution, proposing to clearly specify what is common and what differs in AI conscious processing from full human conscious experience.
Trends in Cognitive Sciences
June 1, 2026
Jean-Pierre Changeux, Michele Farisco
The Global Neuronal Workspace (GNW) theory, often conflated with functionalist computational theories, actually describes a multilevel architecture of conscious processing that extends from cellular and molecular mechanisms to large-scale network dynamics. The paper argues that GNW is not a functionalist computational theory, unlike the Global Workspace theory with which it is frequently confused.
Neural networks : the official journal of the International Neural Network Society
December 1, 2024
Michele Farisco, Kathinka Evers, Jean-Pierre Changeux
Analyzing the question of developing artificial consciousness from an evolutionary perspective, using the evolution of the human brain and its relation with consciousness as a benchmark, reveals several structural and functional features of the human brain that appear key for human-like complex conscious experience. Current AI research should take these into account. Even if AI is limited in emulating human consciousness for intrinsic and extrinsic reasons, taking inspiration from brain characteristics that make human-like conscious processing possible is a promising strategy.
Proceedings of the National Academy of Sciences of the United States of America
September 27, 2022
Konstantin Volzhenin, Jean-Pierre Changeux, Guillaume Dumas
A three-level computational model of information processing and cognitive development is introduced. The first sensorimotor level handles local nonconscious processing during a visual classification task. The second cognitive level globally integrates information via long-range connections, still nonconsciously. The third conscious level, based on global neuronal workspace theory, maintains self-sustained representations in the absence of sensory input, requiring interneurons. Results show that synaptic epigenesis—selection and stabilization of synapses at local and global scales—is necessary for solving trace and delay conditioning tasks. Dopamine enables credit assignment across temporal delays between perception and reward. Balanced spontaneous activity facilitates epigenesis, and a balanced excitatory/inhibitory ratio improves performance.
PLoS Biology
June 15, 2016
Stanislas Dehaene, Jean-Pierre Changeux
Even without sensory input, cortical and thalamic neurons exhibit structured spontaneous activity whose origins and functions are unclear. Computer simulations of a simplified model with multiple interconnected thalamocortical columns and long-range excitatory axons reveal two main activity states: spontaneous gamma-band oscillations emerge at a precise threshold controlled by neuromodulator systems, and within a spontaneously active network, sudden "ignition" of one of many possible coherent high-level activity states occurs among cortical neurons with long-distance projections. During ignition, spontaneous activity can block external sensory processing, relating to inattentional blindness, where intensely engaged subjects fail to notice salient but irrelevant stimuli. The minimal network's generic properties may clarify basic cerebral phenomena underlying consciousness's autonomy.