The various senses each have adaptive value to the organism and the brain processes the data from these senses to build a better model of the world it lives in. Evolution grasps for data in any direction and tries to exploit it for adaptive advantage. Our senses develop when evolution finds such useful data sources.
Some sort of intelligence is a natural outcome of such evolutionary elaboration. For this argument I define intelligence as some algorithm to process such data to make a better living for the organism. The brain performs this algorithm. It seems plausible that information about the working of the algorithm proper will eventually come to be a useful data source, not much different from the other senses. It might be thought that the algorithm already had ample access to its own workings but this may not be so; witness the difficulty we have in discovering grammar rules that we evidently employ. Our theory of grammar seems to come from deductions from observed speech patterns, rather than introspection. Most AI schemes in the literature are sublimely non self aware!
The development of sensory nerves that report on the activity of the brain proper, and report it as subject, just as the optic nerve reports images of things out there as subjects to be considered and perhaps put into the model.
The Stretch computer (IBM model 7030) had a scan-out feature designed to help finding hardware failures in the field. When hardware error circuits detected an error, normal execution of instructions would cease and the states of all of the latches (flip-flops) would be copied into a core memory by hardware means specialized just to that task. It, or an unbroken aid, would then proceed to examine this record and try to diagnose the failure. It is noteworthy here that the computer proper could not read its state without these specialized means. It required extra hardware to move the data that was the essence of computing, to a place where it could be computed upon! Many modern processors also have scan-out, such as JTAG, which is now used mainly in later stages of hardware debugging.
Perhaps speech, or some immediate substrate of speech (ala Chomsky), is implicated in this function. Much of what I am directly aware of that transpires in my head, is in the form of speech, or can at least be facilely rendered as speech. If the reports of this wiring merely proceed directly to short term memory, as with the Stretch above, then the observable effects that I can think of are explained. This also fits my subjective experience. After all as I report verbally these self observations, I draw on my memory just as I report other observations verbally.
I imagine that these proposed nerves that report brain function, tap into the brain at several points but leave other parts entirely unreported. The reported points may be at different abstraction levels. We may thus be aware of memories for purposes other than why our memory ability evolved.
The first chapter of Greg Egan’s Diaspora is a stream-of-consciousness report of an AI deducing the existence of itself. It would seem that Egan’s AI lacked what I propose exists naturally in our brain. I agree that a moderately bright AI could indeed deduce its own existence, much as Egan imagines without this specialized wiring. With the extra wiring it would be quicker yet. Without special wiring one might have to deduce one’s thoughts from observed actions, or perhaps from perceived urges to actions. Such deductions indeed seem necessary sometime but more often we know some of the principle precursors to these urges.
John McCarthy thinks that self awareness is and should be wired into people, and that it should be wired into AI’s.
It occurred to me recently that neural nets, such as those in our brains, and that some propose should be emulated digitally to achieve AI, may keep the patterns they know in a form that cannot be decoded by other than themselves. We are not very good at conveying in words how to recognize some particular person, even though that ability would be adaptive; nor are we able to say how to ride a bicycle. Our neural nets do this well.
How this direct introspection relates to consciousness is, I think, more a matter of choice of words than something to be decided in either the laboratory or arm chair.
What does it mean to have first class access to your thought process; how could it be otherwise? The air-conditioning demon alluded to above probably has direct knowledge of the temperature of the rooms under its surveillance, but not direct access to the deductions that it has made about those temperatures. It does not know why it believes what it believes. If it is a classic (1980’s) AI there will be a list of such deductions stored in the memory of the computer and this memory will be integral to the workings of the AI. The AI will most likely have no concept of proposition with which to deal with these deductions. The list of deductions will not include propositions about the deductions. John McCarthy thinks that successful AI’s need this ability. Seeing as how we have no successful general AI’s today he and I may both be right. You don’t need a general AI to control the air-conditioning, but you may do well to manage your deductions about temperatures, much as you manage a collection of temperature readings.
Those who speculate on the evolution of intelligence often claim that its survival value is to understand the thoughts of other humans. If this is so then the adaptive advantage of access to one’s own thoughts is clear. Such an ability may be necessary to create arguments to convince others.
See my notes on “A Brief Tour of Human Consciousness”.