Can Artificial Intelligence Ever Become Conscious?

 

If you have been exploring new large-language-model forms of artificial intelligence like ChatGPT over the past year, you are probably now convinced that these generators of humanlike writing are not actually conscious. They are instead like a very fancy version of the autocomplete response that pops up when you are typing an email. They make good predictions, and therefore they can approximate something that a human might have written (or drawn, or composed). But when you look under the hood, "there's no there there": The algorithm has no awareness of what it's saying. 

The current state of affairs begs the question of what it would take for an artificial intelligence to become actually self-aware. The major obstacle to achieving conscious AI is that we don't understand what makes us humans self-aware, i.e. what "consciousness" is in the first place. For most of the last century we had a work-around for this problem in the form of the Turing Test: As long as an AI couldn't fool a human being into thinking that they were talking to another human, then the AI was certainly not conscious in the same way that humans are. But AI systems have now far surpassed that bar. I'd like to consider four current views of consciousness among people who study the philosophy of mind, and look at what each of these views might imply for the future development of conscious AI. These viewpoints are often called (a) illusionism, (b) materialism, (c) dualism, and (d) idealism. The following chart shows how each of these sets of assumptions would affect the possibility of conscious AI:

Philosophy of Mind	Sub-views and Proponents	Implications for Conscious AI
Illusionism	We are mistaken about being conscious (Dennett)	AI is “conscious” already (and/or, we ourselves are not really conscious in the way that we think we are)
	Consciousness is feedback in a recursive system (Hofstadter)	AI is conscious when it engages in metacognition (“tell me why you thought that”); if it is not yet conscious, it will be soon. Or: consciousness can be generated by clever prompt engineering
Materialism (Material Monism)	Consciousness is just abstract thinking (Churchlands)	AI should become genuinely conscious if structured properly to build up abstract concepts from specific examples
	Consciousness is the result of competing thoughts in a global workspace (Gazzaniga)	Chaining together multiple AIs that compete for attention in a common arena should produce consciousness
	Consciousness results from attention being focused on one goal vs. another (Graziano, du Bois & Elliott)	AI should be conscious if combined with an “attention system” that chooses where to allocate focus based on goals
	Consciousness results from embodied interactions with the environment (Lakoff)	Large-language-model AI should be conscious if combined with sensors and robotics to produce an “embodied” AI
Dualism	Consciousness is an outside force that affects matter (Descartes)	AI systems are built from the “bottom up” and therefore can never be conscious; all conscious systems are driven “top down” by an immaterial soul. Or: AI will only be conscious if/when God decides to make it so.
	Consciousness is an emergent property of complex systems (Chalmers)	Once AI systems become sufficiently complex, they will automatically achieve consciousness. Possibly, super-human AI systems will also have greater consciousness than we do
	Consciousness is an outside force that is drawn to sufficiently complex systems (Card)	Once AI systems become sufficiently complex, they will “attract” an immaterial soul in the same way that humans do (again, possibly a super-human one)
	Consciousness is an electromagnetic field that is generated by neurons (Pockett)	Consciousness is the result of neurons’ electromagnetic fields; large-language model AIs don't generate these fields, and therefore can't be conscious. A different technology, in theory, could be
	Transceiver theory: Consciousness is a “broadcast,” and brains simply receive it (van Lommel)	AI systems structured in the right way (we don’t know what) could function as transceivers to pick up on consciousness (we don't know from where)
	Quantum mind: Consciousness is generated by the wave collapse of superpositions (Bohm, Chopra)	Digital computers can’t be conscious, but quantum computers might be
Idealism         (Idealist Monism)	Consciousness is all that exists; what looks like matter is simply the outward reflection of conscious experience (Berkeley, Kastrup)	AI can never become conscious (or rather, consciousness can never become AI), because silicon chips are not what consciousness looks like in matter. Or, AI systems will give universal consciousness a new form of expression

First, let's look at illusionism: If you are a devotee of the philosopher Daniel Dennett, then I think there's nothing further for us to talk about. Dennett's book Consciousness Explained argues that we are all in fact philosopher's zombies, systems that produce language to indicate that we are conscious but that we have no actual interiority, no hidden flame of self, no breath of God that animates us. Instead, we are just sophisticated word-generating machines. If we accept this argument at face value (and I'm not convinced that even Dennett really does), then ChatGPT is conscious, full stop. That's because the word "conscious" means only "an entity that can pass the Turing Test"; all other meanings are superfluous, or in other words illusions.

The cyberneticist Douglas Hofstadter (in his books Godel, Escher, Bach and I Am a Strange Loop) presents a view in which language that expresses consciousness is not just a performance for other people, but instead a sort of feedback or static that occurs when the camera of perception is turned back upon itself. If that's correct, then current forms of AI might not be quite conscious yet, but they aren't far from it. Google's AI system is able to comment on its own mental state, and AI algorithms have begun to show evidence of "theory of mind," the form of language use in which one person can describe what another person might be thinking, and how it's different from their own perspective. Although these specific results have been challenged, it seems obvious that AI is nevertheless getting closer to expressing itself in a way that seems conscious. Certain types of prompts are more likely to produce conscious-seeming expressions than others, AI systems can learn to differentially make statements that seem conscious when their human operator gives them positive feedback on those statements, and AI prompts can be structured in a way that makes the system express not just a conclusion but also a chain of reasoning to support it. Still, to most people, "being conscious" means something more than just producing certain types of speech, so it seems incumbent on Hofstadter to explain how a feedback loop might result in an AI actually experiencing something subjectively real. Philosophers call that "something" qualia, the feeling of "what it is to be like" a person, or a bat, or perhaps a flower. If qualia exist at all (in other words, if we aren't just fooling ourselves; but if we are, then who is doing the fooling, and who is being fooled?), then Dennett and Hofstadter's definitions of consciousness both seem insufficient.

Now on to the materialists, who are probably the most common type of thinkers about consciousness in contemporary philosophy and neuroscience departments. An article highlighted this year in the New York Times looks at five different ways in which consciousness might arise in future AI systems, but these all shared the underlying assumption of materialism -- i.e., that conscious experience is directly created by the operations of physical systems. That entire article was premised on the assumption that consciousness is an emergent property that occurs when a physical system is structured in a particular way. One materialist theory -- recurrent processing theory -- overlaps with Hofstadter's ideas but suggests that a true conscious experience would occur from a properly-structured feedback loop. Paul and Patricia Churchland have argued that simply having a neural system that abstracts from specifics to general principles is enough to create the subjective feeling of consciousness, a true "what it's like" to have an experience. 

A second materialist theory suggests that consciousness might be the result of competition between multiple systems that operate in parallel, all generate results in a shared "global workspace," and are subject to rules for the allocation of limited attention resources to select among the systems. This is in some ways similar to how the Intuitive and Narrative minds work together in humans, and definitely tracks with the parallel operations of multiple sub-systems that together make up the Intuitive mind. Neuroscientist Michael Gazzaniga suggests that an "interpreter" (i.e., the Narrative mind) then creates a stream of conscious experience by editing together the various competing impulses into one coherent (but ultimately illusory) self-story. It is possible to envision an AI system that works in the same way by joining several of today's AI's together, giving them different responsibilities, and making them compete for attention from a "master" system that itself uses machine learning to spin a narrative story about what's going on beneath the surface. According to Gazzaniga and others, such a system would then be "conscious" in exactly the same way that we are. 

Two other materialist models also rely on the idea of attention -- a concept that I have identified as part of the Narrative system, but also as a potential bridge between the Narrative and Intuitive minds. Attention is clearly a limited resource, we have the experience of being able to exert some degree of control over it, and the qualia that we experience are dependent on where we direct our attention. Attention, therefore, is a likely candidate as an operational definition of consciousness. But we have the same problem in defining attention as we do in defining conscious experience itself. The materialist theories focused on attention (attention schema theory and predictive processing) both assume that attention is a kind of predictive model that selects resources based on the expected future benefits they convey. Our own experience of attention differs: Sometimes we can direct our attention to something that is simply interesting, or indeed to something uninteresting if we have an extrinsic reason for choosing to focus on it. Over time, this type of directed attention can become more automatic, and ultimately might be subsumed into our sense of "who we are." 

A final materialist theory suggests that consciousness is the result of agency (the ability to cause consequences) combined with embodiment (the ability to receive sensory data from one's environment). If this theory is correct, then the non-consciousness of current AI algorithms is simply because they are blind and powerless -- as we might be if we spent our whole lives locked in a dark room with nothing but the contents of the Internet for companionship. To achieve conscious AI, this model suggests, we simply need to network our current large-language models with sensors and robotics, and allow the system to gain consciousness through repeated interactions with its environment. Recent advancements in sensor technology seem poised to permit such combinations of computing resources, and I would hazard a guess that someone out there is testing them already.

A third class of thinking about consciousness suggests that conscious experience is a different kind of "stuff" than matter. Cartesian dualism is the classic Western worldview, expressed in terms of the distinction between body and mind, or body and soul. Rene Descartes doubted everything about his own sensory experience, but couldn't doubt the working of his own mind -- cogito ergo sum, or "I think, therefore I am." Because his mind clearly (to him) existed even if his body did not, Descartes decided that body and mind were two distinct things. Cartesian dualism has a number of problems, among them these: (a) nobody has yet been able to show how conscious experience comes out of physical properties of matter, and (b) there are limited instances in which conscious experiences are seen to have direct effects on matter, and all those are controversial -- see, e.g., Benjamin Libet's studies on whether conscious thoughts truly direct our "voluntary" motor behaviors. If consciousness does not actually cause behavior, then one possible explanation is that consciousness is simply an "epiphenomenon": something that is connected to matter, but has no way of directly affecting matter. An absence of causal relationships between mind and matter doesn't necessarily undermine a dualist worldview, but it does tend to limit any possible role of free will. Whether or not mind is able to cause any events in the world, AI could never be conscious under Cartesian dualism because no material thing can ever be structured in a way that creates consciousness. Consciousness comes from outside and sits on top of matter, rather than being generated out of the material structure. In a theistic Cartesian model, of course, God could simply add consciousness to AI in the same way as consciousness is added to the machine of the human body. But AI researchers would never be able to accomplish that on their own.

Philosopher David Chalmers starts from the assumption that conscious experience is not necessary to explain human behavior; in other words, a world populated entirely by philosopher's zombies could exist, and from the outside it could look identical to our own. But in our world, there is an extra property, related to the informational content of complex systems, that describes "what it's like" to be something or someone. It was Chalmers who coined the term "the hard problem" to describe the unclear relationship between the "what it's like" phenomenological experience and the observable world of matter. Chalmers's own resolution to the hard problem is to suggest that any sufficiently complex system will necessarily produce conscious experience. For some readers (and maybe Chalmers himself), the implication of this view is panpsychism -- the hypothesis that even small systems with organized information are in some sense conscious, and that very large amalgams of entities are also conscious in some way, all the way up to the level of a conscious universe. Self-awareness comes to systems automatically in this view, even though it is separate from the material the system is made of, because self-awareness is an emergent property of organized structures. Importantly, consciousness should be equally likely to arise whether the underlying system is made of neurons or silicon chips, a feature known as "substrate independence." Advocates for an organized-systems dualist view of consciousness include Google founder Larry Page. The most complex computers are still far less complex than the human brain, but that threshold could conceivably be reached in the next 2 decades. The ultra-complex computer doesn't need to have specific subsystems or functions, it just needs to be self-organizing in some way. In fact, Eric Schwitzgebel argues that if Chalmers's model is correct, then large self-organizing systems like the United States (the geopolitical entity) likely have their own stream of conscious experience already! Given enough complexity, and a mechanism for self-modification, it's therefore possible that consciousness will spontaneously appear as a property of AI. Current large-language models have shown the potential for emergent properties already: For instance, one of their strengths is that they can generate snippets of code to achieve a particular task, even though no one specifically trained them on how to create computer code. Coding doesn't imply consciousness, of course. It just suggests that complex systems have the potential for unanticipated results. 

There are several variants of the panpsychism argument, one of which (Chalmers's) is wholly materialist -- i.e., that the development of self-awareness is an inherent effect of complexity, but that self-awareness is still at some level separate from the material system that goes along with it. (The philosopher's technical term for this relationship is that consciousness supervenes on matter -- it's riding along on top, even though you can't see it in the actual atoms and particles). Another interesting version of dualism can be found in science fiction writer Orson Scott Card's "Enderverse" novels, where matter is given organization by immaterial aiuas, or souls. It has been suggested that Card's idea draws from a philosophy about "intelligent matter" proposed by early Mormon writer Orson Pratt. Card's view is similar to Chalmers's in that consciousness automatically appears in the right type of complex system, but it is different in that consciousness does not necessarily come out of the system itself. Rather, the system has attracted the attention of a "something else" that was previously elsewhere (Card calls it "outside" and suggests that this elsewhere is also the key to faster-than-light travel). Although this is fiction, it's just as coherent a theory of consciousness as other dualist approaches. Card's view, like Chalmers's, is substrate-independent: In fact, a key character in his fiction is the super-intelligent AI entity named Jane, who emerges from the connections between computers. The various panpsychist arguments also frequently lead to the conclusion that AI will eventually be necessarily smarter than human beings, because its level of complexity will eventually exceed our own and therefore will result in a greater depth and breadth of conscious experience.

Other models of consciousness are dualistic without the assumption of panpsychism. Another contemporary dualist approach suggests that a brain is more like a radio receiver -- that it does not produce consciousness from inside itself, but rather captures or downloads consciousness from somewhere outside the body. If this hypothesis is true, then building complex systems from silicon might or might not result in conscious experience. The answer would depend on whether only certain types of systems (e.g., organic ones) are capable of receiving the consciousness "waves" and interpreting them effectively. Science-fiction stories about uploading one's brain to the cloud as an AI also might run into this barrier. Other dualist models are more closely tied to physical brain structures: An electromagnetic field theory of consciousness suggests instead that consciousness does come out of the brain, but is related to its electromagnetic field rather than to its physical structure. An AI system would only be conscious, therefore, if it generated the same type of electrical field as neurons do. And quantum mind theory alternately suggests that consciousness is connected to the collapse of quantum super-positions at the molecular level in the brain. This has the appealing feature of non-determinism (i.e., the potential for free will), because the results of a quantum process cannot be known in advance. However, there are also good reasons for questioning this theory, among them the idea that brains are too large, hot, and wet to permit quantum effects that are typically seen only at the submolecular level. Nevertheless, new developments in quantum computing might eventually permit an AI system to operate in this way, and therefore might facilitate consciousness in AI. Our current digital computers, however, don't have the necessary features to generate quantum wave collapse events.

A final perspective on consciousness is that of idealism, which posits that the material world is illusory and that consciousness is all there is. On a purely physical level, idealism seems to be empirically true: Matter only appears to be solid due to the action of the Higgs field, yet we interpret the world as though it were made up of solid objects. But redefining what we mean by "matter" doesn't really solve the problem of whether consciousness has a material origin. True idealist theories like that proposed by Bernardo Kastrup suggest that "physically objective matter is an abstraction of mind. We do not know matter in the same way that we know mind, for matter is an inference and mind a given" (p. 37, The Idea of the World). In other words, we have an immediate and direct experience of our own experience, but when we see a chair we are experiencing something that looks like a chair to us -- we don't know what's really there. (If the standard model of particle physics is correct, there actually is no chair, only a set of fields and particles). Kastrup suggests that "contrary to physicalism, it is the inner experiences of an organism -- including non-self-reflective and internally dissociated types unreachable through introspection ... -- that cause its body" (p. 80). Kastrup's view agrees with the writings of Bishop George Berkeley who argued that "to be is to be perceived" (esse es percipi) -- that things do not exist except when they are perceived by a conscious mind. Both Berkeley and Kastrup, incidentally, require the existence of a larger "transpersonal mind" to account for the fact that people have shared experiences; Bishop Berkeley simply refers to that transpersonal mind as God. In Kastrup's view, AI can never be conscious because computers are not the type of physical thing that goes along with conscious experiences -- if a consciousness were there, it would instead be expressed as a brain. But that's not the only possible implication of idealism for AI. Another possible view, which in practice looks a bit like Card's form of dualism, is that AI might simply be a new way for that underlying transpersonal mind or universal consciousness to express itself. 

It is an open question which of these competing models of consciousness is true. It might seem like we have a new laboratory (AI research) in which to answer some thorny metaphysical questions with real data. But unfortunately, even if AI were to become conscious, we don't currently have a good way to know it. After all, our current AIs are able to pass the Turing Test, so they appear to be conscious already based on our best-available test for consciousness. A different yardstick is necessary for us to determine whether an AI system is "truly" conscious or only "appears to be" conscious. And unfortunately, creating that yardstick might not be possible without first answering the underlying philosophical questions! All of the structural methods that have been proposed to make AI systems conscious require us to first presuppose that materialism is true. If it is false, then all those strategies might do is to produce newer and more convincing imitators of human consciousness. Whether consciousness is possible for AI or not, we will need to find a way out of this tautology before we can make valid judgments about the results of any new experiments.