XPost: comp.theory   
   From: dnk@OMIT.csse.unimelb.edu.au   
      
   Tero Hakala writes:   
      
   > This may be a well known question or a just result of my misconceptions,   
   > but so far I haven't got any definite answer.. So maybe someone   
   > here could help or clarify things for me.   
      
   > ..   
      
   > I was recently contemplating Turing tests and Turing machines (TM) and   
   > was wondering if the fundamental limitations of TM can be exploited   
   > to discover whether the conversation partner in a Turing test is a   
   > digital computer AI or a real person.   
      
   > As far as I have understood the issue, we have the following points   
      
   > 1) any digital computer+software can in principle be reduced to a   
   > somekind of TM. So the computer can not exceed the computational   
   > capabilities of TM.   
      
   > 2) There are problems that a universal TM can't decide. Eg.   
   > the halting problem:  given TM b and input c, does the machine stop   
   > at some point?   
      
   So far, so good.   
      
   > Now, suppose that we come up with a simple TM with input that does   
   > not stop. Eg. it produces an endless string of aaa..'s.  A human   
   > with sufficient knowledge should be able to see that this   
   > machine never stops.   
      
   > Let's say that we pose this question to our human/AI partner,   
   > ie. we describe our never stopping TM and ask: does it   
   > stop?   
      
   > Now a real human could provide us with a definite answer. However, any   
   > digital computer is subjected to the limitations of TM and   
   > therefore can not say for sure if our machine stops or not.   
      
   Here's where you make an error.  There are indeed TM's that can detect   
   non-terminating TM's, as people can.  2) above doesn't say no TM can   
   detect non-termination of *any* TM's, it says something much more   
   specific, that no TM can correctly determine termination of *every*   
   TM/input pair.  Equally, no person can reliably determine this, since   
   some cases would be more complex than they could possibly understand.   
      
   Indeed, if a person can do it reliably for some case, then so can a   
   specific TM (but that TM will fail on other cases, as the human would).   
   And TM's can solve termination problems that humans never could.   
      
   > So my question is, can we use this kind of scheme to discover whether we   
   > are speaking with an AI implemented on a digital computer or with   
   > a genuine human?   
      
   The short answer is no.  For a long-winded one, try comp.ai.philosophy   
      
   > (of course, an AI mimicing human behaviour would probably say   
   > something like "get a life, smart ass, don't bore me to death", in   
   > which case we couldn't tell :) )   
      
   >  - T.H   
      
   [ comp.ai is moderated ... your article may take a while to appear. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)