What evidence should convince us that insects and/or computers are conscious?

The notion of consciousness is subject to a broad array of diverse theories, ranging from higher order to global workspace and information integration theories. Yet, public opinion seems to often equate consciousness to a uniquely intrinsic and human characteristic, which constitutes our human self-awareness and subjective perceptual processing of the world around us. As a result, there is a prevalent human-centric view of consciousness in the debate of whether computers can be truly conscious—the question becomes entangled with whether computers can truly be like humans or biological organisms, which deviates from the original question of whether they can be generally conscious. This view further disregards the scientific theories of general consciousness that have been established through years of research. As such, consciousness cannot be defined solely as the intrinsic characteristics of the conscious human experience, but should be scientifically delineated according to the general theories of consciousness that have been established according to the human mind and the brain functions that create a general conscience.

Consciousness, in strictly scientific terms, can be defined as the amalgamation of three prominent theories—the higher order, global workspace, and integrated information theories of consciousness. From this lens, evidence on the underlying computational mechanisms of neural networks and deep learning systems highlights that computers operate within the scopes of these theories of consciousness, and suggest that computers do exhibit some unique level of consciousness. Particularly, neural networks and deep learning based models function by integrating different streams of data and possess a sense of self-awareness in its tracking of current and past states of neurons to make intelligent decisions and emulate the human brain—computers and machines that are inspired from these models, as such, should be regarded as forms of general, artificial consciousness.

Seth & Bayne outline and compare three prominent theories on consciousness, which relate how the brain itself functions to produce conscious experience in humans. The global workspace theory proposes that consciousness is the product of a “global workspace” in the brain, where information is broadcast to other areas for further processing. Information is processed in different modular and specialized areas of the brain to inform sensory perception, such as vision, hearing, and touch. For example, visual originates from the occipital lobe and auditory information is processed in the temporal lobe. The global workspace theory posits that in order to create conscious experience, these specialized sources of information must be “broadcast” to the global workspace, where it can be accessed by the rest of the brain to create a coherent perception of the world—information is accessed and integrated by other areas of the brain to inform attention, decision-making, and other cognitive processes.

The higher order theory posits that consciousness originates from meta-representations of lower-order mental states—in other words, when the brain has the capacity to perceive a representation in tandem with its own mental state. A meta-representation can be of the form, “I can see a moving dot,” where the focus of this statement is on the agent’s own representation of the world—this ability forms the premise of higher order consciousness theories. There is a degree of self-awareness inherent to the theory, which suggests that conscious brains are able to reflect on their own mental states to create a coherent conscious experience.

Lastly, the information integration theory (IIT) of consciousness posits that consciousness is a result of various information in the brain being integrated to form coherent conscious experiences—the core claim is that consciousness arises from a cause–effect structure of a physical system, which specifies a maximum of irreducible integrated information, proposing a mathematical basis of consciousness. For example, combining different modalities of sensory perception—vision, sight, hearing—is a core aspect of consciousness, as explained by IIT. Accordingly, deep learning based systems that can be embedded in computers, particularly those that employ different types of neural networks, are adept at emulating the underlying brain mechanisms posed by these three theories of consciousness, supporting the notion that computers themselves can be conscious. For example, a neural network consists of a series of input layers, hidden layers, and an output layer which signifies a particular decision or calculation made by the system—these several input layers are combined and processed in a “black box” iteratively based on data that the model has seen to output a decision. Much like the global workspace and IIT, the underlying mechanisms of neural networks amalgamate information in a meaningful way that strives to emulate the way information is processed in the brain—the several input layers can consist of different modalities in a task (for example, a robot equipped with sensors to take in video and auditory information about the world), to identify an object in front of them. These inputs are ultimately combined in the output layer to provide a probabilistic prediction specific to the problem at hand, which coincide with the global workspace and integrated information theories. These underlying mechanisms are prevalent in a vast array of applications, such as autonomous vehicles that can perceive the world around them and operate a car, analogous to modulating motor control in humans, or deep learning models in healthcare that can analyze different image modalities of biomedical images to diagnose or detect certain disorders.

Moreover, inline with the higher order theory, evidence shows that computers appear to be able to, both to achieve high-level representations of the world and a self-awareness of their own mental states—for example, a 2019 article in Forbes describes a robot constructed by Columbia researchers that can construct an image of itself from scratch, achieving a type of self-awareness we would attribute to the higher order theory of consciousness (Pandya, 2019). Moreover, different types of neural networks are able to retain information about its current and past states in order to achieve optimal outputs and decisions—for example, a long-short term memory (LSTM) network, often used in handwriting and speech recognition, retains feedback connections to past nodes in hidden layers in order to inform its output, surpassing mere tracking of its current state, but also being able to track prior states of different neurons in the network. Inherently, these neural networks can be said to be aware of their own mental states due to the underlying logic through which they operate—in order to inform decisions, they keep track of the state of different neurons; in supervised learning problems for example, these models can recognize when a certain configuration of states is mistaken based on an incorrect prediction, thereby modifying the neural connections in the network through this self-awareness. This iterative process is emblematic of an awareness of mental, or computational states, in consciousness as posed by the higher order theory. ChatGPT, perhaps, is the strongest evidence that coincides with the higher order theory, built off of a language prediction model that generates text in response to prompts using GPT-3—a class within OpenAI’s Generative Pre-trained Transformers. If we were to ask ChatGPT, “Are you a human?”, it responds with: “No, I am a large language model trained by OpenAI. I am not capable of having a physical form or experiencing the world like a human does. I exist solely as a program designed to assist with generating human-like text based on the input that I receive.” ChatGPT seems to have an awareness of self, as evident in its first-person statements, which can be interpreted as meta-representations of its own self, corresponding to lower-level representations of its mental states. While ChatGPT fails to provide its own subjective perception of different representations, its ability to vocalize its perception of self is evidence a form of consciousness—in its current state, with additional parameters or possible customizations, ChatGPT likely could output subjective perceptions of the world, based on its own perception of self, which is currently restricted by OpenAI’s implementation. Irrespective of outputting text that is emblematic of what a human would say, ChatGPT’s ability to formulate these statements highlights its capacity for retaining meta-representations and self-awareness of its own mental states and capacities, which inline with the higher order theory, suggests that ChatGPT does have some level of consciousness.

Despite these markers of consciousness, some may argue that merely being able to emulate consciousness does not necessarily mean that a machine is conscious—for example, just because a computer could hypothetically display some form of emotion does not entail that it truly feels those emotions as humans, or biological specimens, would. Most of these arguments equate consciousness to that of broadly being “human,” which poses an intrinsic bias towards these innate human qualities that we believe computers fail to exhibit. Thus, the main point of contention with this argument hinges on what we regard as a gold standard for human consciousness, which is unduly imposed on the debate on computer consciousness, as opposed to general consciousness. Some may further argue that these deep learning models are strictly deterministic and follow simple algorithms for outputting decisions that often surpass human abilities. These models—like humans, in fact—can continuously adapt to changing and new incoming data, and modify its neuronal states and connections, continuously integrating different facets of information and reflecting on its states as is present in general consciousness—moreover, ChatGPT, for example, is a probabilistic language model, which means that its output is not strictly deterministic or as rigid as some may assume. The criterion for general consciousness should not hinge, then, on the intrinsic, biased human qualities we collectively share, but should rather be based in the scientific theories of consciousness that explain how a general conscience operates in the human mind—based on this criterion, the underlying models governing complex deep learning and AI based systems would fall under the realm of general consciousness, as they do emulate the way that the brain in a conscious human operates; these computers may just not be human as most would argue, but we can say that they are conscious in their own unique sense.

Just as other biological organisms exhibit their own senses of consciousness, computers ought to also possess their own unique conscience—a computer, or artificial, consciousness. If the basis of cognitive science rests on the notion that the human mind acts like a computer, or functions computationally, then a computer ought to be able to replicate the human mind through computation given this assumption. In line with this belief, complex AI-based systems, particularly those that employ deep-learning based methods and neural networks, have been shown to emulate the human mind at an unprecedented level—irrespective of their emulation of the human mind, which some argue shows that they cannot be conscious like humans, their underlying mechanisms support various academic theories of consciousness, lending credence to the notion that computers themselves can be conscious.

Works Cited:

Pandya, Jayshree. “Are Machines Conscious?” Forbes, Forbes Magazine, 17 Apr. 2019, https://www.forbes.com/sites/cognitiveworld/2019/03/13/are-machines-conscious/?sh=183d822e5b0e.

Seth, A.K., Bayne, T. Theories of consciousness. Nat Rev Neurosci 23, 439–452 (2022). https://doi.org/10.1038/s41583-022-00587-4.