The conductor model of consciousness, our neuromorphic twins, and the human-AI deal

Abstract

Critics of Artificial Intelligence posit that artificial agents cannot achieve consciousness even in principle, because they lack certain necessary conditions for consciousness present in biological agents. Here we highlight arguments from a neuroscientific and neuromorphic engineering perspective as to why such a strict denial of consciousness in artificial agents is not compelling. We argue that the differences between biological and artificial brains are not fundamental and are vanishing with progress in neuromorphic architecture designs mimicking the human blueprint. To characterise this blueprint, we propose the conductor model of consciousness (CMoC) that builds on neuronal implementations of an external and internal world model while gating and labelling information flows. An extended Turing test (eTT) lists criteria on how to separate the information flow for learning an internal world model, both for biological and artificial agents. While the classic Turing test only assesses external observables (i.e., behaviour), the eTT also evaluates internal variables of artificial brains and tests for the presence of neuronal circuitries necessary to act on representations of the self, the internal and the external world, and potentially, some neural correlates of consciousness. Finally, we address ethical issues for the design of such artificial agents, formulated as an alignment dilemma: if artificial agents share aspects of consciousness, while they (partially) overtake human intelligence, how can humans justify their own rights against growing claims of their artificial counterpart? We suggest a tentative human-AI deal according to which artificial agents are designed not to suffer negative affective states but in exchange are not granted equal rights to humans.