Feb 11
3
Turing vs. Searle
I’m a little embarrassed to admit that I originally read Searle’s piece more than 25 years ago. I’ve diligently re-read it, together with re-reading Turing’s piece. And, in truth, I still think Searle is hiding the ball when it comes to the core issue.
In short, Searle makes completely valid arguments through about 70% of his piece. And then he hides behind nebulous terms and concepts and somewhat mystical generalities.
I once suspected that Searle didn’t understand Turing’s concept of computability. Now, I suspect he fully understood it, but chose to ignore it.
To over-generalize (probably to the point of heresy or inaccuracy), Turing’s theory of computability, expressed in the Church-Turing Hypothesis, is essentially that anything that can be “computed” or “solved” or “calculated” or otherwise represented in a systematic algorithmic fashion CAN be computed by a Turing Machine. Because almost any computational device has the properties of a Turing Machine, almost anything that can be solved or calculated or computed can be solved or calculated or computed with existing technology.
If we accept the notion (which Searle himself embraces), that human beings and their brains are a type of machine and that machines can therefore think, then, accepting the Church-Turing Hypothesis, it’s very hard for me to see how thought cannot be replicated on a computer. Of course, there’s a reason why it’s the Church-Turing Hypothesis, rather than the Church-Turing Law. So far, it has been an impossible theory to prove (and, in fact, may be unprovable — I’m not too strong on my Godel and completeness theory).
However, to get back to the issue of Turing versus Searle. I see Searle’s argument as follows:
“Strong AI” advocates argue, incorrectly, that they can (a) teach a machine to think without really understanding how the human brain “thinks”, (b) understand human thinking by building machines that appear to simulate human thought, (c) demonstrate that a machine thinks by having it behave in a manner that fools other human beings.
I agree with him completely on all of these points. Interestingly, Searle then goes on to argue (or concede) that brains (presumably human and animal) are themselves machines and that there’s nothing that prevents a “machine” from “thinking” if it can accurately reproduce the complex biological/physical processes associated with thought. He also argues, compellingly, that without understanding the physical processes associated with thought, it’s a fool’s errand to attempt to simulate them.
Where I think Searle’s argument starts to lose cohesion is when he starts distinguishing “machines” from “programs”. Brains as “hardware”, and thought as “software” are, to be fair, poor analogies. But I don’t feel that this is what Turing was ever arguing. Nor, do I feel, does this address the fundamental issue.
Where I feel Searle’s argument completely loses validity is when he starts using phrases such as reference to a brain’s “causal properties,
its ability to produce intentional states” as a unique and distinguishing quality. I don’t know what “causal properties’ are. I don’t know that they’ve been quantified or measured or otherwise subject to any of the rigors we expect. In short, I think they’re short-hand for “something special and indefinable and therefore ir-reproducible, because it’s a brain.” And once Searle takes this path, I feel he has lost the argument. Because, although he (accurately) upbraids the strong AI advocates for failing to take heed of the actual physical phenomena associated with thought, he starts using concepts that do not appear to have any quantitative or qualitative merits. In other words, how would a billion dollar neuro-biology project undertake to measure or identify the “causal properties’ of the human brain? How does one distinguish a human “intentional state” from, for example, the gravitational pull between two objects? In my opinion, Searle, having identified many of the issues with “overly strong” AI, falls back on meta-physical generalities about human thought and emotion that, by definition, can’t be quantified and therefore can’t be simulated or reproduced.
A few other points:
(1) Searle’s “Chinese box” is a flawed metaphor in that: (a) it assumes a highly deterministic approach to machine decision making, which is likely to fail very very quickly, and (b) for it to be a truly compelling “Chinese box” the book used would need to be able to re-write itself — for example, the question “could you please summarize the last four things we were discussing” could not be answered without some very very complex rules.
(2) Searle appears to put great stock in the notion that “simulation isn’t reality.” While on the face of it, this is a true statement, simulation of certain phenomena can be all but indistinguishable from the actual event. For example, simulation of the interaction of billions of subatomic particles probably has a long way to go before experimental physiscists have much to worry about. However, simulation of a light switch turning on a light is pretty much near perfect. There’s no theoretical reason why simulation of certain events can’t be as accurate as the actual event. Again, this is where I feel Searle is avoiding the issue.
In summary, if, as Searle concedes, machines can think, and if we are all, in fact, machines, then I simply don’t see how “programs” can be distinguished from “machines” and how “simulation” cannot be sufficiently accurate to be, for all practical purposes, “the real thing.”
As for Turing, well, I’ll say this — He was probably one of the 5-10 smartest human beings of the 20th century. He played a major role in the Allied war effort and probably helped shorten World War II by several years. He is one of perhaps 7 founders of modern computer science. He had a play written about his life, in which he was played by Derek Jacobi. And I always felt that all he was postulating in his 1950 piece was an exemplar test, intentionally subject to debate. His argument, as I saw it was “someday, machines will probably be able to think … here’s one test I suggest.” I can’t really add to, critique or otherwise embellish that.