Does claiming that strong AI is impossible imply a belief in substance dualism?

Question

I was reading an article by J Mark Bishop "The danger of artificial stupidity" on ScientaSalon, where he quotes his own research, John Searle and Hilary Putnam, among others, as proof of the impossibility of strong AI. I've always felt that strong AI deniers were closeted substance-dualists. People who believed in souls, but were unwilling to come clean about their religious/metaphysical beliefs for fear of being ridiculed. So instead they come up with all sorts of pragmatic arguments against strong AI like qualia or computers lack of insight, which don't really hold.

My reasoning for why denying the possibility of strong AI implies substance dualism is the following:

1. Any finite sized physical phenomena can be reproduced given sufficient technological means and a sufficient understanding of the underlying physical processes.

2. Denying the possibility of strong AI means that no matter how advanced our technology and how comprehensive our knowledge of neuroscience and psychology will become, we will never be able to reproduce the functionality of the human mind.

3. Per 1) The only reason we would not be able to reproduce the mind's functionality is if there is something non-physical about how the mind works.

4. Saying there is something non-physical about how the mind works is the same as substance dualism.

My question is the following: Is this indeed the case, that denying the possibility of strong AI implies substance dualism?

Answer 1

I think that people who claim strong AI is impossible do tend (often unwittingly) to commit themselves to some manner of dualism. But there is an argument to be made that the process is inscrutable:

We have no general theory for approximating arbitrary recurrent functions. Indeed, there are all sorts of recurrent functions with profoundly frustrating processes such as chaotic functions where long-term values cannot be predicted without infinitely accurate measurements of current states even when you know the form of the function.

Furthermore, our brains are heavily activity-dependent (synaptic plasticity, adaptation, etc. etc.). Therefore it plausibly could make a difference to the qualitative operation of the system what the long-term chaotic behavior is. (It is not guaranteed that it will, but it might.)

Thus, it is at least somewhat plausible that although our brains are in a universe which is nominally computable, in practice there is no way to find the computation that needs to be done. At every reduced level you run into exponentially hard problems (computational chemistry is too hard, protein folding is too hard, etc.), and at the highest level of abstraction there is no way to discover or verify the algorithm (and algorithmic space is enormous; you can't brute force the search).

This would, nonetheless, allow us to monitor and replay the entire activity of someone's brain (given almost unimaginable heroics of instrumentation and genetic engineering), but that is not strong AI. (It is very weird, especially from a philosophical standpoint: here you have a consciousness-over-time arrayed as patterns in space instead.)

Personally, I don't think this type of inscrutability is very likely, but the reasons for thinking so are more of a vague hunch about the evolvability of systems that rely upon details of chaotic behavior of self-recurrent systems than anything truly sound.

Answer 2

I can think of a few alternatives:

• One could argue for a case where a human mind grade AI is theoretically producible, but the universe lacks sufficient resources to do so. This would be a practicality argument, not a theoretical possibility argument.
• Idealism can claim strong-AI is impossible, without being dualistic.
• Not all finite sized physical phenomena can be reproduced. You have to be able to measure it first, and there may be unmeasurable values in the universe (QM has shown presumably unmeasurable values exist).

There is also the cheating argument, to claim that "strong AI" is not defined sufficiently to allow us to accomplish it, but I don't believe that is what you are looking for.

Answer 3

To complete Cort Amon's answer, I would say that there is a difference between reproducing physical phenomena and computing them.

Even if all physical systems can be reproduced using some physical material (say you reproduce the structure of a living cell from similar molecular components), that doesn't mean you can compute them. You only make physical copies of physical systems. Imagine that the only way to reproduce consciousness is creating a physical brain made out of similar organic components (neurons, ...) rather than implementing a software on a turing machine. Then strong AI would be impossible: computers are made out of transistors, not neurons, and you won't ever get the same result on a silicon computer, or any turing machine, even though dualism is false.

Note that contemporary physics is not computable, or only using strong approximations (and even then you only get probabilities). Also note that cloning a physical system is physically impossible (cf. the no cloning theorem) so even reproduction could be impossible...

Edit: here are some reasons why contemporary physics is not computable:

• space-time is continuous, not discrete and you'd need infinite ressources for a bounded system. The idea that ultimate physics would be discrete is speculative.
• physical quantities are irrational, not rational numbers and again you'd need infinite ressources
• natural constants and measured quantities too, but we have finite measurement precision
• a small imprecision can expand exponentially if the system is chaotic
• in standard quantum mechanics, there are infinite series that we can only approximate through perturbation theory. In general, there is no analytic solution to the equations of the theory. Physicist often use classical approximations for interactions with external systems.
• an open system might not be separable from its environment. We need idealisation even before we talk of "a system" (reduced density matrix).
• in quantum field theories, we need further idealisations because of infinite divergence (renormalisation). (Although this could be solved by a future theory of quantum gravity, so if you consider an ideal physical theory in your argument this one is not necessarily an issue).
• even when all these problems are addressed, computation time of quantum mechanics is exponential. Say it takes 1ms to compute 1ms, it could take 10ms to compute the next 1ms, 3 hours to compute the next 6ms, and 12 days to compute only 10ms in total...
• as mentionned before, at this point you only get the probabilities...

Contemporary physicists cannot calculate the structure of complex molecules (up to a few atoms), even with huge computational capacities: they ask chemists to do this.

In sum, the conjecture today is that ultimate physics will be computable and it's an unlikely conjecture. Actual physics is not computable for many foundational reasons.

You could argue that it's not necessary to have all physical information: the substrate does not matter, only the "software" (the higher level) matters... But this is question begging. Why wouldn't the substrate matter for consciousness? Only when strong AI will be a success we'd know that it's the case.

Answer 4

There are other alternatives; for example:

Spinozas system which encompasses mental and physical phenomena as modes of an absolute self-subsisting simple substance.

ln Leibnizs Monadology; there are many kinds of monads; a soul is a monad, as is God; they don't interact; each reflects in itself all the others; and harmonised independently.

In Epicurus's system all is atoms; and the soul is made of soul atoms; as such they can affect material atoms.

In Hegel; the world is the progression of the world-Geist from non-being and being; which in a sense are the same; thus human beings are an expression of the Geist.

Kant identifies a noumenal world behind the phenomenal world; one could take this inexpressible and indescrible dimension where mental and physical phenomena are 'one'; though he quiet on how the noumenal world causes the phenomenal world.

Schopenhauer identifies the noumena of an impersonal will; and it is the force of this will that causes the phenomenal world.

I would hazard a guess, that it was the admixture of Greek and Christian Philosophy that identified the absolute substance as God; and although Descartes is usually counted as the originator of the dual substance thesis; in fact, he's not, being too careful and cautious a thinker; one supposes it was the outcome of later cogitations by Descartians - but this removal of the metaphysical scaffolding left these hanging in the void, and unable to interact with each other by definition; as substances are causally closed.

Answer 5

I think the ability to clearly define the 'strong' in 'strong AI' is equivalent to dualism.

If you do not think there is some inscrutable aspect of the human mind, then all AI is continuous with the behavior of young or impaired humans, and there is not some magical point where it becomes 'strong AI', equivalent to humans. In some aspects it is already superior, and in others it could compensate given adequate power or time.

If you consider genetic recombination as a form of computation, it has the power to create human intelligence, having done so. Given a different environment, it could create a different form. We also see it solve problems that we cannot solve, and we have to simply steal its solutions until we understand them. So that is an equally strong or better problem-solving system than human intelligence. Is it 'artificial' enough? It appears to be digital, and we think of it as quite mechanical -- the kind of thing we can emulate in machines.

If you do think there is this inscrutable aspect, only then can you really state any definition of 'strong AI' with meaning that truly separates it from 'a whole lot of weak AI'.

So the question ends up being baseless or circular. If you accept a functional (performative, Turing-esque) definition of strong AI, you do so only because you are not a dualist, and if you resist, you do so only because you are one.

Answer 6

There are two reasons why it might be impossible to create an Artificial Intelligence with a mind equal to or better than that of a human: One reason could be that the building blocks of the brain are in some way superior to silicon chips or anything else that humans could create as the substrate for intelligence. The other reason could be that creating Artificial Intelligence is just very difficult, and human scientists and software developers are not clever enough to do it.

Substance dualism would be a subcategory of the first reason. But you could argue that the brain has about 10 to the eleventh power neurons, with each having on average 7,000 connections, and that is just a big number that beats current computer hardware. (On the other hand, the computer hardware runs a lot faster than the human brain's hardware). It is an argument for human-grade AI not being possible today, but not for the impossibility in the future. Substance dualism would be an argument for "impossible in principle".

The other argument that creating an AI is just very difficult is also one that doesn't speak for "impossible in principle". It might be that creating a computer which spends 18 years learning like a human is needed to create an AI. That would mean "not impossible". (Note that once such a computer had learned for 18 years, it could easily be duplicated, unlike humans).

So if there are claims that AI is in principle not possible, then attributing some mysterious quality to the human brain that humans can't reproduce would be one argument for such a claim.

Just a note: It is well-known that there are certain mathematical problems that computers cannot solve. But humans can't solve them either. Humans can choose to ignore such problems. An artificial intelligence would necessarily need the ability to ignore such problems as well.

Answer 7

With regard to your points 1 through 4, you have to be a bit more careful. Sure we could probably artificially produce something modelling a human thinking brain, substituting biological tissue by something else. I doubt any scientist would deny this possibility. But this is likely not what you, or anyone really understands by hard AI. Really we want a computational model of intelligence, a Turing machine running a very complicated program. It is difficult to "disprove" possibility of this, but there is very good reason to doubt possibility because of Godel's incompleteness theorem, and Turing's halting problem. It is bizare that this is not mentioned above. It would be difficult to give these things justice here, but some people have thought long and hard about the question, see for instance the books of Roger Penrose.

Answer 8

The scientific consensus is almost overwhelming that strong AI is possible and substance dualism is inconsistent (unless you add a bunch of qualifiers to "impossible", which goes against Searle's original argument against Strong AI). Once we get that out of the way, in response to OP's actual question: Two incorrect beliefs do not have to imply each other. You have identified a set of assumptions that would equate the incorrect beliefs, but that does not mean a belief in one statement implies a belief in the other, as it is entirely possible that the "believer" has not taken into consideration the comprehensive implications of his belief. If this were to be the case all the time, we wouldn't need philosophical debates at all.

If your goal is to demonstrate that these statements are wrong, then you can simply attack each statement individually, since the equivalence you're drawing about beliefs in these statements adds no epistemological value to your argument.

Answer 9

No, not necessarily.

Consider Searle, known for his criticisms of AI research (or perhaps more accurately, of the claims staked already quite early on, with little to show for it).

Yet Searle says he is a physicalist (and by implication, a non-dualist). That is to say, he says that there are physical systems that think (e.g. our brains) and there is no reason to suppose that there could not be machines that think - since whatever physical principles are at work in our brains could presumably be replicated in a suitable sort of machine.

However, Searle also feels that the course set by computer science (and their associated philosopher friends) will not lead to a suitable sort of machine.

If I have understood him correctly, he maintains that we cannot be sure that we have understood all the physics that underpins mental processes (which I guess is true, strictly speaking) and I think he also claims that computers and software categorically lack that piece of missing physics.

The latter is problematic, as it hinges on precise definitions of what we mean by computer and software (CS 101, I know, but nonetheless tricky to apply to this question) and on characterising this bit of physics of which the nature remains undetermined.

The latest spate of kind-of-OK AI-generated imagery, prose, and translations may impress some, but will suggest to others that, as Searle would have it, an essential ingredient is likely to remain lacking for some time to come. But I am not sure whether the missing ingredient really is a bit of yet-to-be-determined physics.