Abstract:
Much of the recent discourse on AI consciousness has centered around two questions: is artificial consciousness possible in principle and do current systems meet the requirements laid out by prominent theories. For many practical purposes, it matters more whether we are likely to see nontrivial numbers of conscious artificial minds as the technology matures. If artificial consciousness is both possible and easy to produce, it doesn’t immediately follow that an explosion of research, development, and adoption of AI technologies will lead to a corresponding proliferation of conscious artificial minds.
I will argue that we will not see large numbers of conscious artificial minds. To make the strongest case, I’ll concede both computationalism and the thesis that consciousness requires only implementing a relatively straightforward architecture. Instead, I suggest that consciousness involves facets that are inefficient in artificial constructs. Though we may deliberately create conscious minds for experimental purposes, the cheapest routes to the ends that conscious minds could serve won’t involve the right architectures.
My central thesis is that consciousness is likely to be an effective solution to a certain kind of problem faced by biological brains. Future AI systems will be subject to fundamentally different needs and constraints. Given these differences, imposing the solution that works in a biological context is likely to be inefficient.
I will not defend a specific account of the problem that consciousness solves. The claim can be motivated with general considerations on probabilistic grounds. I take it that consciousness is related to such faculties as sensory integration, learning, attention, and self-representation. The problem solved by consciousness in biological minds has something to do with these functions, such that we are unlikely to see consciousness in systems where those functions are completely absent or take a substantially different form.
I focus on two groups of needs and constraints. First, evolutionary and developmental considerations concern the genesis of our brains. Second, organismic and social considerations concern the needs and responsibilities of brains in their natural context. These considerations suggest significant limitations to biological brains that would likely not apply to AI systems.
Evolutionary and developmental constraints relate to the fact that human brains were formed by evolutionary processes and must unfold themselves from a single cell by ordering gene expressions. Cognition is served by wiring patterns that evolved early in our evolutionary history and are resistant to substantial modification. Individuals must face a changing world with little direct instruction and must utilize unsupervised learning techniques to start finding food and evading predators almost immediately. Since our ancestors had brains built for motor control and sensory processing, our minds leverage those faculties, with some modifications, to do all the things we do.
Organismic and social constraints relate to the fact that human brains are housed in, and responsible for sustaining, independent organisms. Humans are separate discrete entities with obligations for their own upkeep. They are born with a fixed set of cognitive machinery that they need to make the most of. They cannot significantly change or scale that machinery. It is hard to justify including computational structures that are highly specialized and rarely used.
Consider the significance of attention in a paradigmatic case. A young mammal spies a predator. It has an instinct to flee, but if it chooses poorly it will not survive. It has never faced this exact problem. It will do best if every available cognitive resource is turned to formulating a response and if it can find a way to smoothly integrate the resulting information into a decision. It must combine what it knows about itself and its own affordances, the intentions of the predator, and the layout of the environment into a decision as quickly as possible.
This is unlike the problems future AI systems will generally face. Most will probably not depend on such extreme time-sensitive choices for their success. They will have the ability to delegate work to scalable specialized systems, from using larger models to spending more time refining an answer. (We already see the start of this with mixture of expert architectures, utilization of python or other models, and speculative decoding.) They will have less need to confront novel problems with sparse sensory evidence and with general-purpose resources. They will probably not be as dependent on systems that were originally built for sensory processing or motor cognition.
Global workspace theory provides a nice example of how different needs might promote different organizations. The theory posits that consciousness relies on a single global repository that receives updates from and broadcasts to a set of modules. These modules include sensory, action planning, and memory systems, among others. The repository serves in information routing and filtering, facilitating the use of the available modules to work on the problems at hand. The global workspace helps with coordinating their activities, of focusing attention, of promoting parallel work, and coming up with flexible solutions to novel problems.
Global workspace architectures may make much more sense in biological brains than future AI systems. A workspace architecture simplifies physical wiring patterns and might naturally develop from loci of sensory integration. Broadcasting allows flexible subsystems to be directed to pertinent tasks instead of sitting idle. That is much more desirable in an organismic setting where the available processors are fixed and contained within a brain. It makes less sense where dynamic aggregations of specialized services can be assembled in the cloud.
Global workspace presents just one example, but the themes will be shared by theories like the attention schema theory, representationalist theories, or higher-order perception theories. There is a common attitude that brains have a particularly sophisticated way of producing intelligence that we should strive to emulate to build better computers. LLMs have demonstrated that we don’t need a human architecture for human-level intelligence. Brains are also subject to stark limitations. Freed of such constraints, we need a good reason to think that any significant design choice, consciousness included, would still be worthwhile from an engineering perspective.
