What would a "Turing test for reference" show?

Question

What does Hilary Putnam mean by this?

Imagine a situation in which the problem is not to determine if the partner is really a person or a machine, but is rather to determine if the partner uses the words to referas we do. The obvious test is, again, to carry on a conversation, and, if no problems arise, if the partner ‘passes’ in the sense of being indistinguishable from someone who is certified in advance to be speaking the same language, referring’ to the usual sorts of objects, etc., to conclude that the partner does refer to objects as we do. When the purpose of the Turing test is as just described, that is, to determine the existence of (shared) reference, I shall refer to the test as the Turing Test for Reference. And, just as philosophers have discussed the question whether the original Turing test is a definitive test for consciousness, i.e. the question of whether a machine which ‘passes’ the test not just once but regularly is necessarily conscious, so, in the same way, I wish to discuss the question of whether the Turing Test for Reference just suggested is a definitive test for shared reference. The answer will turn out to be ‘No’. The Turing Test for Reference is not definitive.

It is certainly an excellent test in practice; but it is not logically impossible (though it is certainly highly improbable) that someone could pass the Turing Test for Reference and not be referring to anything. It follows from this, as we shall see, that we can extend our observation that words (and whole texts and discourses) do not have a necessary connection to their referents.

I can't figure out the difference between what he is calling the "Turing test for reference" and the normal Turing test. See ieas.unideb.hu/admin/file_2908.pdf (Hilary Putnam's Brains in a Vat Starts on page 8.) Can you give an example?

Answer 1

What's the difference between the Turing Test and the Turing Test for Reference?

I'm not sure I can explain it better than Putnam can, but I'll try:

The traditional Turing Test (hereafter TT) is, as you no doubt know, an exercise where a human communicates via what we would now call "chat" with either a) another human, or b) a computer simulating a human. The object of the endeavour is to determine, through conversation, whether the entity on the other side of the chat (i.e., the partner) is a human or not; if the partner ends up being a piece of software capable of consistently fooling humans, then we have a piece of AI which has "passed the Turing Test." Some folks argue that AI capable of passing the Turing Test is conscious; others reject this.

Putnam here proposes a modified Turing Test, the Turing Test for Reference (hereafter TTFR), which resembles the standard TT in all regards, except for one significant difference: the object of the exercise is not to determine if the partner is human or not, but rather, to determine whether the partner refers to the same objects that we do. For example, if the entity refers to a "table", do they mean the same thing that we mean when we refer to a "table"? If the partner ends up convincing us, through conversation, that they refer to objects in the same manner we do, then the partner has "passed" the TTFR. Putnam then goes on to argue that even if we have a partner who has passed the TTFR, that still does not guarantee that we actually have "shared reference".