Starting from the position that symbols are tokenized shapes that have meanings (i.e., symbols as per the Chinese room argument), the assumption is that computers do not process symbols. What implications does this have for the Turing test?
Typical description of the Turing test
A man, the Interrogator, sits in a room at a desk supporting two Teletype machines. (Each machine has a keyboard and screen (or printer)).
There are two other rooms. Each contains a contestant. One contestant is a human, the other, a computer. Each of these other rooms also contains a teleprinter machine connected by a wire to its partner in the Interrogator’s room. Thus one of the Interrogator’s teleprinters is connected to the teleprinter in the human’s room, and the other to the teleprinter in the computer’s room.
This arrangement allows the Interrogator a means to communicate with the contestants in a completely equal, unbiased and blind way, without knowing whether he is talking to a human or to a computer. The only difference in this entire setup is that one contestant is a human and the other, a computer. Everything else is the same. This is important.
The interrogator types questions on one or other keyboard, and the questions then go to the respective contestant and are displayed on the teleprinter’s paper printout. The point being that these questions comprise symbols that are tokenised shapes that have meanings – i.e., words, numbers, maybe even other keyboard characters.
Odd asymmetry in the typical case
But this, above, is not how the rooms are set up. The room that contains the computer has a different communication setup compared to the human contestant’s room. The computer contestant has no perceptual apparatus so can’t see words on the paper roll at the back of the teleprinter.
In fact, the setup Turing is talking about has one of the Interrogator’s two teleprinter machines connected to another teleprinter – the one in the human contestant’s room. But the second of the Interrogator’s machines is wired directly into the computer contestant. No perception, no characters displayed on a roll of paper or anywhere else.
Now this might seem fine. Why would anyone think there was anything wrong with this asymmetry between human and computer setups – IF computers processed symbols?
But we have assumed they don’t.
Implications for the Turing test
Well, in the case of the human contestant, he perceives the symbols on the paper, understands their meanings, and applies his intelligence accordingly, based on that understanding, interpretation, of the shapes of those symbols printed on the paper.
What of the computer? There are no symbols. There are no symbols flowing down the wire into the computer. They are on the keys of the Interrogator’s keyboard, sure, but they don’t leap off and careen down the wire and into the computer. They doggedly resist this and stay firmly painted on the keys.
Since the computer does not receive symbols it cannot understand their meanings. There are no symbol the meanings of which to understand. The computer might be stuffed with a lot of causal juxtapositions – conditionals – embodied in some substrate and containing examples of whatever the computer does process (voltages), but it doesn’t process symbols, and what comes out of the computer are not symbols either.
Because the computer does not process symbols, it does not understand the questions, and since understanding the questions is necessary for the intelligence being tested, the computer is not – could not possibly be – judged to be intelligent, even though it gives good answers on the Interrogator’s printout to the shapes the Interrogator presses when he types on the keyboard.
Hence the Turing test, as Turing described it for the computers of his day, could not possibly test for human-like intelligence because they could not understand written words. They had no visual perceptual apparatus, and they did not process symbols.
What does this mean for the Turing test?
The Turing test of language comprehension is useless, pointless, doesn’t work, unless the computer is a robot that can see words – OR the computer processes symbols. But processing symbols is infeasible and entirely impracticable for today’s electronic digital computers – they are stuck with clocked voltage levels, magnetic domains and semiconductor switch states. So, for the Turing test to be useful, it seems that the computer must be a robot with one or more visual sensors.
(updated 31 July 2018 mainly with change of “Teletype” to “teleprinter”)