Full bibliography

Can we make conscious machines?  Some researchers believe we can, with computation:  For example, Dehaene et al., concluding an article about machine consciousness, described their hypothesis as “resolutely computational” (Dehaene et al., 2017); others begin with theoretical computer science, implying that a programmed computer could be conscious (e.g. Blum and Blum, 2022).  But humans are not programmed computers; indeed, Penrose has argued that conscious understanding is non-computable (e.g., Penrose, 2022). Let us imagine a shop selling conscious machines:  Besides standard models, it might offer machines with “bespoke consciousness”, meaning consciousness made to a customer’s specifications.  For example, a customer might request a machine with a specified repertoire of conscious experiences and with a specified relationship between input sensor signals and output motor signals. This work explores ways to make machines with bespoke consciousness.  We begin by avoiding a possible bias favoring computation:  As described here, bespoke consciousness need not employ programmed or algorithmic computation; it might even be analog more than digital.  We also consider and compare general design approaches, with particular attention to “bottom-up” and “top-down” approaches.  We suggest a schematic design for each of these approaches; each schematic design is based on a respective well-known hypothesis about biological consciousness:  Our bottom-up design is based on microtubules, as suggested by Penrose and Hameroff’s orchestrated objective reduction (Orch-OR) hypothesis (Hameroff, 2022); our top-down design starts with machine-scale electrical and/or magnetic (E/M) patterns, as suggested by McFadden’s conscious electromagnetic information (cemi) field hypothesis (McFadden, 2020).  Both designs can share a framework based on the customer’s request:  For example, in either design, a machine can receive sense-like input signals and provide motor-like output signals as requested; between input and output, it has structure that performs non-conscious operations; some of its non-conscious events are involved in providing output signals in accordance with the requested input/output relationship, some correspond surjectively to conscious events in the requested repertoire, and some might do both.  (Mathematically, surjective correspondence would mean that each of the conscious events has at least one non-conscious event corresponding to it. (Beran, 2023)) Looking forward to possible implementation, we find challenges:  For example, an implementation of either schematic design might begin with an appropriate initial structure.  One might add variations of the initial structure to provide additional output signals or to correspond to additional parts of the repertoire.  Or one might add fundamentally different structures for additional output signals or parts of the repertoire.  Such variations or combinations of structures might meet or at least approximate the customer’s request.  But implementations like this depend on identifying or inventing the necessary structures and then combining them—this might take a long time, and success is not guaranteed. Despite this and other challenges, we hope to improve our understanding of both biological and machine consciousness by designing and implementing machines with bespoke consciousness.