I recently had a brief back-and-forth with Bobby Azarian about his new article on Raw Story. Azarian, a neuroscientist at George Mason University, argued that artificial intelligence (AI) could never be conscious. I highly recommend reading Azarian’s article: it’s a great distillation of some key concepts in the philosophy of mind, and he makes an argument that is well worth considering. For the most part, I agree with Azarian’s reasoning regarding current A.I., but I don’t think his argument precludes the possibility of future A.I. being conscious.
First, a summary of Azarian’s key points:
- Computers are Turing Machines, which means they can perform operations on symbols but can’t recognize what those symbols mean (which requires a mind).
- Consciousness is a biological phenomenon, which is produced by processes very different from what happens inside a computer. While brains are in some sense “digital,” since information is carried by a neuron either firing (a “1”) or not firing (a “0”), there are a host of important analogue biological processes in the brain, including, in Azarian’s words, “cellular and molecular processes, biochemical reactions, electrostatic forces, global synchronized neuron firing at specific frequencies, and unique structural and functional connections with countless feedback loops.”
- Simulation is not duplication: even if we simulate a brain, it wouldn’t be conscious, for the same reason that a computer simulation of water isn’t wet.
These are essentially the arguments that philosopher John Searle has made against the possibility of conscious A.I. And these are all good arguments against the possibility of current machines being conscious. But, none of them imply that A.I. couldn’t be conscious.
Machines could be conscious for a very simple reason: machines are physical systems and so are brains. If machines could reproduce the same properties of brains that make brains conscious, then machines would also be conscious.
With this simple point in mind, let’s go through Azarian’s argument again:
- The fact that computers are Turing Machines does not necessarily preclude their ability to consciously understand what they’re doing. Computers happen to be built in such a way that they don’t need sentience to carry out their computations: you feed them some symbols, they manipulate those symbols according to some algorithm, and they spit out their answer. Brains do much the same thing – often without conscious understanding. For example: Your conscious self might be awful at trigonemetry, but you have neurons in your brain that are quite good at it. Whenever you hear a sound, the sound information gets passed on to those neurons, and, following some algorithm that implements a trigonometric calculation, those neurons compute where the sound is coming from. All you consciously perceive is the outcome of that computation, which isn’t different in any important sense from what Turing machines do. The only difference between you and the Turing machine is that the outputs of these unconscious computations get sent to some network of neurons that are doing something that produces you – i.e., your consciousness. Computers simply don’t carry out that extra step.
- Azarian is right to point out that consciousness is a biological phenomenon. But there’s no reason to think that this phenomenon couldn’t be reproduced in a machine: Vitamin D production is a biological process that happens in your skin, but we can reproduce that process with machines. That’s how Vitamin D supplements are made.
- Azarian is right to say that simulation is not duplication: a computer simulation of water is not wet. But that’s fine. If we can’t simulate, then let’s duplicate. If we build a physical system that duplicates the properties of water that make water wet, then that physical system would also be wet. Likewise, if we build a physical system that has all the properties of brains that make brains conscious, then that system would be conscious.
What could such properties be? What makes brains conscious? And are the properties that make brains conscious the sort of properties we might reasonably expect to be able to duplicate in intelligent machines?
While neuroscientists don’t yet know what makes brains conscious, a theory that is gaining increasing traction in the field is the Integrated Information Theory of Consciousness. I’ve written about the theory on this blog before, and my current research at Berkeley is on testing some of its predictions. The theory, in its most basic form, states that consciousness simply is integrated information, i.e. information from the intrinsic perspective of a system. Let’s assume that the theory is correct (which it might not be). What would the theory mean for A.I.?
Integrated Information Theory implies that A.I. could be conscious, but isn’t. All the biological facts about brains – the chemistry of their neurotransmitters, the frequency with which populations of neurons oscillate, the molecular genetics of neurons, etc., are all simply details of implementation. If Integrated Information Theory is correct, then what really matters for consciousness is the presence of integrated information, regardless of the medium in which integrated information is implemented. It could very well be that all these analogue processes are necessary for generating sufficient amounts of integrated information, but that these processes themselves aren’t what matter.
Here’s the rub: current A.I. doesn’t have any integrated information. As far as I understand it, the architecture of current A.I. systems is purely feedforward. Activity in one layer of an artificial neural network gets sent forward to another layer, and that layer’s activity gets sent forward to another layer, until the activity reaches a final output layer, and the activity in the output layer determines the artificial neural network’s actions.
Fundamentally, this is not how information flows in the brain. The brain, and in particular the thalamocortical system (where we think consciousness is produced), is a stupendously integrated system. Everything talks to everything else, everything talks to itself, and information is constantly getting distributed across the entire network. According to Integrated Information Theory, it is out of this integration of information across vast swathes of the thalamocortical system that consciousness emerges.
The remarkable success of A.I. in beating humans at a whole range of tasks only drives home the point, made long ago by John Searle and reiterated in Azarian’s article, that you don’t need consciousness for intelligence. Weak A.I., i.e. non-sentient A.I., can be extremely intelligent. But I doubt if even Google’s DeepMind has an iota of consciousness.
But there’s no reason that the architecture of an A.I. couldn’t reproduce the properties that makes the thalamocortical system conscious. If A.I.s were restructured in a such a way that they reproduced the properties of brains that make brains conscious – be that integrated information or some other property – then A.I. would be conscious. If, on the other hand, the development of A.I. continues along its current course, without adding integrated information (or some other property that makes brains conscious) into the equation, then we shouldn’t expect A.I. to be conscious any time soon.
I discuss some of these points in my new vlog piece: